Datasets:

bskang

/

CVPR2023_title_abstract_intro

like 0

Zhang_Prompt_Generate_Then_Cache_Cascade_of_Foundation_Models_Makes_Strong_CVPR_2023

Abstract Visual recognition in low-data regimes requires deep neural networks to learn generalized representations from limited training samples. Recently, CLIP-based methods have shown promising few-shot performance benefited from the contrastive language-image pre-training. We then question, if the more diverse pre-training knowledge can be cascaded to further assist few-shot representation learning. In this paper, we propose CaFo , aCascade of Foundation models that incorporates diverse prior knowledge of various pre- training paradigms for better few-shot learning. Our CaFo incorporates CLIP’s language-contrastive knowledge, DINO’s vision-contrastive knowledge, DALL-E’s vision- generative knowledge, and GPT-3’s language-generative knowledge. Specifically, CaFo works by ‘Prompt, Generate, then Cache’. Firstly, we leverage GPT-3 to produce textual inputs for prompting CLIP with rich downstream linguistic semantics. Then, we generate synthetic images via DALL-E to expand the few-shot training data without any manpower. At last, we introduce a learnable cache model to adaptively blend the predictions from CLIP and DINO. By such col- laboration, CaFo can fully unleash the potential of different pre-training methods and unify them to perform state-of- the-art for few-shot classification. Code is available at https://github.com/ZrrSkywalker/CaFo .

1. Introduction Convolutional neural networks [44] and transform- ers [68] have attained great success on a wide range of vi- sion tasks with abundant datasets [15]. Instead, for some data-deficient and resource-finite scenarios, few-shot learn- ing [63,70] also becomes a research hotspot, where the net- ∗Equal contribution. †Corresponding author aphotoofdog. aaphotoofdog.a CLIPDINODALL-EcontrastcontrastgenerateCaFoLanguage-contrastiveKnowledgeVision-generativeKnowledgeVision-contrastive Knowledge adoglookslike…GPT-3generateWhat a dog looks like?adoglookslike… Language-generativeKnowledgeFigure 1. The Cascade Paradigm of CaFo. We adaptively incor- porate the knowledge from four types of pre-training methods and achieve a strong few-shot learner. works are constrained to learn from limited images with annotations. Many previous works have been proposed in this field to enhance model’s generalization capability by meta learning [20, 71], metric learning [74], and data augmentation [31, 73]. Recently, CLIP [57] pre-trained by large-scale language-image pairs shows favorable zero- shot transfer ability for open-vocabulary visual recogni- tion. The follow-up CoOp [83], CLIP-Adapter [24] and Tip-Adapter [76] further extend it for few-shot classifica- tion and achieve superior performance on various down- stream datasets. This indicates that, even if the few-shot training data is insufficient, the large-scale pre-training has endowed the network with strong representation ability, which highly benefits the few-shot learning on downstream domains. Now that there exist various self-supervisory paradigms besides CLIP, could we adaptively integrate their pre-learned knowledge and collaborate them to be a better This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15211 few-shot learner? To tackle this issue, we propose CaFo , aCascade of Fooundation models blending the knowledge from mul- tiple pre-training paradigms with a ‘Prompt, Generate, then Cache’ pipeline. As shown in Figure 1, we in- tegrate CLIP [57], DINO [7], DALL-E [58], and GPT- 3 [4] to provide four types of prior knowledge for CaFo. Therein, CLIP [57] is pre-trained to produce paired fea- tures in the embedding space for every image and its de- scriptive text. Guided by texts with different categor- ical semantics, CLIP [57] can well classify the images aided by language-contrastive knowledge . DINO fol- lows contrastive self-supervised learning [7] to match the representations between two transformations of one same image, which is expert at distinguishing different images with vision-contrastive knowledge . Similar to CLIP [57], DALL-E [58] is also pre-trained by image-text pairs but learns to predict the encoded image tokens based on the given text tokens. Conditioned on the input text, DALL- E [58] could leverage the vision-generative knowledge to create high-quality synthetic images in a zero-shot manner. Pre-trained by large-scale language corpus, GPT-3 [4] takes a few hand-written templates as input, and autoregressively generates human-like texts, which contain rich language- generative knowledge . Therefore, the four models have distinctive pre-training goals and can provide complemen- tary knowledge to assist the few-shot visual recognition. In detail, we cascade them by three steps.: 1) Prompt . We adopt GPT-3 [4] to produce textual prompts for CLIP based on a few hand-written templates. These prompts with richer language knowledge are fed into CLIP’s textual en- coder. 2) Generate . We adopt DALL-E [58] to gener- ate additional training images for different categories based on the domain-specific texts, which enlarges the few-shot training data, but costs no extra manpower for collection and annotation. 3) Cache . We utilize a cache model to adaptively incorporate the predictions from both CLIP [57] and DINO [7]. Referring to Tip-Adapter [76], we build the cache model with two kinds of keys respectively for the two pre-trained models. Regarding zero-shot CLIP as the distribution baseline, we adaptively ensemble the pre- dictions of two cached keys as the final output. By only fine-tuning the lightweight cache model via expanded train- ing data, CaFo can learn to fuse diverse prior knowledge and leverage their complementary characteristics for better few-shot visual recognition. Our main contributions are summarized as follows: • We propose CaFo to incorporate the prior knowledge learned from various pre-training paradigms for better few-shot learning. • By collaborating CLIP, DINO, GPT-3 and DALL-E, CaFo utilizes more semantic prompts, enriches thelimited few-shot training data, and adaptively ensem- bles diverse predictions via the cache model. • We conduct thorough experiments on 11 datasets for few-shot classification, where CaFo achieves state-of- the-art without using extra annotated data.

Xu_Unsupervised_Domain_Adaption_With_Pixel-Level_Discriminator_for_Image-Aware_Layout_Generation_CVPR_2023

Abstract Layout is essential for graphic design and poster gen- eration. Recently, applying deep learning models to gen- erate layouts has attracted increasing attention. This pa- per focuses on using the GAN-based model conditioned on image contents to generate advertising poster graphic lay- outs, which requires an advertising poster layout dataset with paired product images and graphic layouts. How- ever, the paired images and layouts in the existing dataset are collected by inpainting and annotating posters, respec- tively. There exists a domain gap between inpainted posters (source domain data) and clean product images (target do- main data). Therefore, this paper combines unsupervised domain adaption techniques to design a GAN with a novel pixel-level discriminator (PD), called PDA-GAN, to gen- erate graphic layouts according to image contents. The PD is connected to the shallow level feature map and com- putes the GAN loss for each input-image pixel. Both quan- titative and qualitative evaluations demonstrate that PDA- GAN can achieve state-of-the-art performances and gener- ate high-quality image-aware graphic layouts for advertis- ing posters.

1. Introduction Graphic layout is essential to the design of posters, mag- azines, comics, and webpages. Recently, generative ad- versarial network (GAN) has been applied to synthesize graphic layouts through modeling the geometric relation- ships of different types of 2D elements, for instance, text and logo bounding boxes [10, 19]. Fine-grained controls over the layout generation process can be realized using Conditional GANs, and the conditions might include image contents and the attributes of graphic elements, e.g. cat- egory, area, and aspect ratio [20, 35]. Especially, image *Work done during an internship at Alibaba Group †Corresponding author Logo Text Underlay Figure 1. Examples of image-conditioned advertising posters graphic layouts generation. Our model generates graphic lay- outs (middle) with multiple elements conditioned on product im- ages (left). The designer or even automatic rendering programs can utilize graphic layouts to render advertising posters (right). contents play an important role in generating image-aware graphic layouts of posters and magazines [34, 35]. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10114 This paper focuses on studying the deep-model based image-aware graphic layout method for advertising poster design, where the graphic layout is defined to be a set of el- ements with their classes and bounding boxes as in [35]. As shown in Fig. 1, the graphic layout for advertising poster de- sign in our work refers to arranging four classes of elements, such as logos, texts, underlays, and other elements for em- bellishment, at the appropriate position according to prod- uct images. Therefore, its kernel is to model the relationship between the image contents and layout elements [4,35] such that the neural network can learn how to produce the aes- thetic arrangement of elements around the product image. It can be defined as the direct set prediction problem in [5]. Constructing a high-quality layout dataset for the train- ing of image-ware graphic layout methods is labor inten- sive, since it requires professional stylists to design the ar- rangement of elements to form the paired product image and layout data items. For the purpose of reducing work- load, zhou et.al. [35] propose to collect designed poster im- ages to construct a dataset with required paired data. Hence, the graphic elements imposed on the poster image are re- moved through image inpainting [28], and annotated with their geometric arrangements in the posters, which results in state-of-the-art CGL-Dataset with 54,546 paired data items. While CGL-Dataset is substantially beneficial to the train- ing of image-ware networks, there exists a domain gap be- tween product image and its inpainted version. The CGL- GAN in [35] tries to narrow this domain gap by utilizing Gaussian blur such that the network can take a clean prod- uct image as input for synthesizing a high-quality graphic layout. However, it is possible that the blurred images lose the delicate color and texture details of products, leading to unpleasing occlusion or placement of graphic elements. This paper proposes to leverage unsupervised domain adaption technique to bridge the domain gap between clean product images and inpainted images in CGL-Dataset to significantly improve the quality of generated graphic lay- outs. Treating the inpainted poster images without graphic elements as the source domain, our method aims to seek for the alignment of the feature space of source domain and the feature space of clean product images in the target do- main. To this end, we design a GAN with a pixel-level do- main adaption discriminator, abbreviated as PDA-GAN, to achieve more fine-grained control over feature space align- ment. It is inspired by PatchGAN [13], but non-trivially adapts to pixel-level in our task. First, the pixel-level dis- criminator (PD) designed for domain adaption can avoid the Gaussian blurring step in [35], which is helpful for the net- work to model the details of the product image. Second, the pixel-level discriminator is connected to the shallow level feature map, since the inpainted area is usually small rel- ative to the whole image and will be difficult to discern at deep levels with large receptive field. Finally, the PD is con-structed by three convolutional layers only, and its number of network parameters is less than 2%of the discriminator parameters in CGL-GAN. This design reduces the memory and computational cost of the PD. We collect 120,000 target domain images during the training of PDA-GAN. Experimental results show that PDA-GAN achieves state-of-the-art (SOTA) performance according to composition-relevant metrics. It outperforms CGL-GAN on CGL-dataset and achieves relative improve- ment over background complexity, occlusion subject de- gree, and occlusion product degree metrics by 6.21%, 17.5%, and 14.5%relatively, leading to significantly im- proved visual quality of synthesized graphic layouts in many cases. In summary, this paper comprises the follow- ing contributions: • We design a GAN with a novel pixel-level discrimina- tor working on shallow level features to bridge the do- main gap that exists between training images in CGL- Dataset and clean product images. • Both quantitative and qualitative evaluations demon- strate that PDA-GAN can achieve SOTA performance and is able to generate high-quality image-aware graphic layouts for advertising posters.

Zhang_Dense_Distinct_Query_for_End-to-End_Object_Detection_CVPR_2023

Abstract One-to-one label assignment in object detection has suc- cessfully obviated the need for non-maximum suppression (NMS) as postprocessing and makes the pipeline end-to- end. However, it triggers a new dilemma as the widely used sparse queries cannot guarantee a high recall, while dense queries inevitably bring more similar queries and en- counter optimization difficulties. As both sparse and dense queries are problematic, then what are the expected queries in end-to-end object detection? This paper shows that the solution should be Dense Distinct Queries (DDQ). Con- cretely, we first lay dense queries like traditional detectors and then select distinct ones for one-to-one assignments. DDQ blends the advantages of traditional and recent end- to-end detectors and significantly improves the performance of various detectors including FCN, R-CNN, and DETRs. Most impressively, DDQ-DETR achieves 52.1 AP on MS- COCO dataset within 12 epochs using a ResNet-50 back- bone, outperforming all existing detectors in the same set- ting. DDQ also shares the benefit of end-to-end detec- tors in crowded scenes and achieves 93.8 AP on Crowd- Human. We hope DDQ can inspire researchers to con- sider the complementarity between traditional methods and end-to-end detectors. The source code can be found at https://github.com/jshilong/DDQ . * Equal contribution.

1. Introduction Object detection is one of the most fundamental tasks in computer vision, which aims at answering what objects are in an image and where they are. To achieve the objective, the detector is expected to detect all objects and mark each object with only one bounding box. Due to the complex spatial distribution and the vast shape variance of objects, detecting all objects is quite chal- lenging. To solve the problem, traditional detectors [17,21, 27] first lay predefined dense grid queries1to achieve a high recall. Convolutions with shared weights are then applied to quickly process dense queries in a sliding-window manner. At last, one ground truth bounding box is assigned to multi- ple similar candidate queries for optimization. However, the one-to-many assignment results in redundant predic- tions and thus requires extra duplicate-removal operations (e.g., non-maximum suppression) during inference, which causes misaligned inference with training and hinders the pipeline from being end-to-end (as shown in Fig. 1.(a)). This paradigm is broken by DETR [2], which assigns only one positive query to each ground truth bounding box (one-to-one assignment) to achieve end-to-end. This scheme requires heavy computation to refine queries and adopts self-attention to model interactions between queries 1 Anchors [17, 21] or anchor points [27] in conventional detectors play the same role as sparse object queries in [2]. Hence, we collec- tively refer to densely distributed anchor boxes and anchor points as dense queries. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7329 to facilitate the optimization of one-to-one assignment, which unfortunately limits the number of queries. For ex- ample, DETR only initializes hundreds of learnable object queries. Therefore, compared to the densely distributed queries in conventional detectors, the sparse queries fall short in recall rate, as shown in Fig. 1.(b). Some recent works have also tried to integrate dense queries into one-to-one assignment [24, 28, 32]. How- ever, dense queries in end-to-end detectors face unique challenges. For example, our analysis shows that this paradigm would inevitably introduce many similar queries (potentially representing the same instance) and that it suffers difficult and inefficient optimization as similar queries are assigned opposite labels under one-to-one as- signment.(Fig. 1.(c)). Now that both sparse queries (low recall) and dense queries (optimization difficulty) under one-to-one assign- ment are sub-optimal, what are the expected queries in end- to-end object detection? In this study, we demonstrate that the solution should be dense distinct queries (DDQ) , meaning that the queries for object detection should be both densely distributed to de- tect all objects and also distinct from each other to facilitate the optimization of one-to-one label assignment. Guided by such a principle, we consistently improve the perfor- mance of various detector architectures, including FCN, R-CNN, and DETRs. For one-stage detectors composed of fully convolutional networks (FCN), we first propose a pyramid shuffle operation to replace heavy self-attentions to model the interaction between dense queries. Then, a dis- tinct queries selection scheme ensures that the one-to-one assignment is only imposed on the selected distinct queries, preventing contradictory labels from being assigned to simi- lar queries. Such an end-to-end one-stage detector is named DDQ FCN and achieves state-of-the-art performance in one-stage detectors. DDQ also naturally extends to DETR and R-CNN structures [7, 19, 26, 38] by first laying dense queries as in [38] and then selecting distinct queries for later refining stages, which are respectively dubbed DDQ R-CNN andDDQ DETR . We have conducted experiments on two datasets—MS- COCO [18] and CrowdHuman [23]. DDQ FCN and DDQ R-CNN obtain 41.5/44.6 AP, respectively, on the MS- COCO detection dataset [18] with the standard 1x sched- ule [17, 21]. Compared to recent DETRs, DDQ DETR achieved 52.1 AP in just 12 epochs with the DETR-style augmentation [2]. The strong performance demonstrates that DDQ overcomes the optimization difficulty in end-to- end detectors and converges as fast as traditional detectors with higher performance. Object detection in crowded scenes such as CrowdHu- man [23] is another arena to testify to the effectiveness of DDQ. It is extremely cumbersome to tune the post-processing NMS in traditional detectors, as a low IoU threshold leads to missing overlapping objects, while a high threshold brings too many false positives. Recent end-to- end structures also struggle to distinguish between dupli- cated predictions and overlapping objects due to their diffi- cult optimization. In this study, DDQ FCN/R-CNN/DETR achieve 92.7/93.5/93.8 AP and 98.2/98.6/98.7 recall on CrowdHuman [23], surpassing both traditional and end-to- end detectors by a large margin.

Zala_Hierarchical_Video-Moment_Retrieval_and_Step-Captioning_CVPR_2023

Abstract There is growing interest in searching for information from large video corpora. Prior works have studied rele- vant tasks, such as text-based video retrieval, moment re- trieval, video summarization, and video captioning in iso- lation, without an end-to-end setup that can jointly search from video corpora and generate summaries. Such an end- to-end setup would allow for many interesting applications, e.g., a text-based search that finds a relevant video from a video corpus, extracts the most relevant moment from that video, and segments the moment into important steps with captions. To address this, we present the HIREST (HIerarchical REtrieval and STep-captioning) dataset and propose a new benchmark that covers hierarchical infor- mation retrieval and visual/textual stepwise summarization from an instructional video corpus. HIREST consists of 3.4K text-video pairs from an instructional video dataset, where 1.1K videos have annotations of moment spans rel- evant to text query and breakdown of each moment into key instruction steps with caption and timestamps (totaling 8.6K step captions). Our hierarchical benchmark consists of video retrieval, moment retrieval, and two novel moment segmentation and step captioning tasks. In moment segmen- tation, models break down a video moment into instruction steps and identify start-end boundaries. In step caption- ing, models generate a textual summary for each step. We also present starting point task-specific and end-to-end joint baseline models for our new benchmark. While the baseline models show some promising results, there still exists large room for future improvement by the community.1

1. Introduction Encouraged by the easy access to smartphones, record- ing software, and video hosting platforms, people are in- creasingly accumulating videos of all kinds. To fuel the *equal contribution 1code and data: https://github.com/j-min/HiRESTsubsequent growing interest in using machine learning sys- tems to extract and summarize important information from these large video corpora based on text queries, progress has been made in video retrieval [2, 17, 18, 41, 42], moment re- trieval [10, 16, 17], video summarization [9, 24, 33, 34], and video captioning [13, 20, 41, 42]. Previous works have gen- erally focused on solving these tasks independently; how- ever, all these tasks share the common goal of retrieving in- formation from a video corpus, at different levels of scales and via different modalities. Hence, in this work, we intro- duce a new hierarchical benchmark that combines all four tasks to enable novel and useful real-world applications. For example, a text-based search service that finds a rele- vant video from a large video corpus, extracts the most rel- evant moment from that video, segments the moment into important steps, and captions them for easy indexing and retrieval. To support this, we introduce H IREST, a hierar- chical instructional video dataset for a holistic benchmark of information retrieval from a video corpus (see Sec. 3). HIREST consists of four annotations: 1) 3.4Kpairs of text query about open-domain instructions ( e.g.,‘how to make glow in the dark slime’ ) and videos, 2) relevant moment timestamps inside the 1.1Kvideos, where only a part of the video ( <75%) is relevant to the text query, 3) moment breakdown in several instructional steps with timestamps (7.6steps per video, total 8.6Ksteps), and, 4) an manually curated English caption for each step ( e.g.‘pour shampoo in container’ ). We collect fine-grained step-wise annota- tions of H IREST in a two-step annotation process with on- line crowdworkers on instructional text-video pairs from the HowTo100M [23] dataset (see Sec. 3.1). The instructional videos often come with clear step-by-step instructions, al- lowing fine-grained segmentation of the videos into short steps. While there are existing video datasets with step an- notations, they are based on a small number of predefined task names [36, 46] (thus step captions are not diverse), or are limited to a single topic ( e.g. cooking [45]). H IREST covers various domains and provides diverse step captions with timestamps written by human annotators (see Table 1), presenting new challenging and realistic benchmarks for hi- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23056 1) Video Retrieval 3) Moment Segmentation 4) Step Captioning Pour shampoo in container... 48s... 63s ... Pour glow stick in shampoo90s 104s 0s 44s 119s 162s Retrieved Moment Segmented MomentQuery: “How to Make Glow in the Dark Slime” 2) Moment Retrieval ... Break glow stick67s 77s Figure 1. Overview of four hierarchical tasks of our H IREST dataset (Sec. 3). 1) Video retrieval: find a video that is most relevant to a given text query. 2) Moment retrieval: choose the relevant span of the video, by trimming the parts irrelevant to the text query. 3) Moment segmentation: break down the span into several steps and identify the start-end boundaries of each step. 4) Step captioning: generate step-by-step textual summaries of the moment. erarchical video information retrieval. Using the H IREST dataset, we benchmark four tasks: 1) video retrieval, 2) moment retrieval, 3) moment segmen- tation, and 4) step captioning (see Fig. 1 and Sec. 3.3). In the video retrieval task, models have to identify a video that is most relevant to a given text query. In the moment re- trieval task, models have to select the relevant span of the video, by trimming the parts irrelevant to the text query (blue boundary in Fig. 1). In the moment segmentation task, models have to break down the relevant portion into sev- eral instructional steps and identify the start-end boundaries of each step (green boundaries in Fig. 1). Finally, in the step captioning task, models have to generate step captions (e.g.‘spray the warm water on carpet’ ) of the instructional steps. To provide good starting points to the community for our new task hierarchy, we show the performance of re- cent baseline models on H IREST. For baselines, we use strong models including CLIP [27], EV A-CLIP [8], Frozen- in-Time [2], BMT [13], and SwinBERT [20]. On all four tasks, we find that finetuning models on H IREST improve performance; however, there exists a large room to improve performance. We summarize our contributions in this paper: 1) We present H IREST dataset and propose a new benchmark that covers hierarchy in information retrieval and visual/textual summarization from an instructional video corpus. 2) Un- like existing video datasets with step captions based on predefined task names or limited to a single topic, our HIREST provides diverse, high-quality step captions with timestamps written by human annotators. 3) We provide a joint baseline model that can perform moment retrieval, moment segmentation, and step captioning with a single ar- chitecture. 4) We provide comprehensive dataset analysesand show experiments with baseline models for each task, where there is a large room to improve model performance. We hope that H IREST can foster future work on end-to-end systems for holistic information retrieval and summariza- tion on large video corpus. In addition, our manually an- notated step captions can also be a good source for training and testing the step-by-step reasoning of large multimodal language models [40, 44].

Xu_Meta_Compositional_Referring_Expression_Segmentation_CVPR_2023

Abstract Referring expression segmentation aims to segment an object described by a language expression from an image. Despite the recent progress on this task, existing models tackling this task may not be able to fully capture semantics and visual representations of individual concepts, which limits their generalization capability, especially when han- dling novel compositions of learned concepts . In this work, through the lens of meta learning, we propose a Meta Com- positional Referring Expression Segmentation ( MCRES ) framework to enhance model compositional generalization performance. Specifically, to handle various levels of novel compositions, our framework first uses training data to con- struct a virtual training set and multiple virtual testing sets, where data samples in each virtual testing set contain a level of novel compositions w.r.t. the virtual training set. Then, following a novel meta optimization scheme to opti- mize the model to obtain good testing performance on the virtual testing sets after training on the virtual training set, our framework can effectively drive the model to better cap- ture semantics and visual representations of individual con- cepts, and thus obtain robust generalization performance even when handling novel compositions. Extensive experi- ments on three benchmark datasets demonstrate the effec- tiveness of our framework.

1. Introduction Referring expression segmentation (RES) [12, 38, 40] aims to segment a visual entity in an image given a lin- guistic expression. This task has been receiving increas- ing attention in recent years [5, 18, 34, 37], as it can play an important role in various applications, such as language- based human-robot interaction and interactive image edit- *Corresponding Author T r aining Data T esting Data No v el Composition "dark horse" "t he coff ee mug""mug wit h dark coff ee" (a) w or d-w or d w or d-phr ase phr ase-phr ase bir d whit ewhit e bir d standing in wat er standing in wat er wat er in“whit e bir d standing in wat er ” (b) Figure 1. Illustration of novel compositions and various levels of compositions. (a) An example of the testing sample containing the novel composition of “dark coffee” in RefCOCO dataset [41]. Such a novel composition itself does not exist in the training data, but its individual components (i.e., “dark” and “coffee”) exist in the training data. (b) An example of various levels of compositions in an expression. We perform constituency parsing of the expres- sion using AllenNLP [9]. Based on the obtained parsing tree as shown above, we can then obtain various levels of compositions (e.g., word-word level, word-phrase level) in this expression. ing. However, despite the recent progress on tackling this task [5, 34, 42], existing methods may struggle with han- dling the testing samples, of which the expressions contain novel compositions of learned concepts. Here a novel com- position means that the composition itself does not exist in the training data, but its individual components (e.g., words, phrases) exist in the training data, as shown in Fig. 1a. We observe that testing samples containing such novel compositions of learned concepts widely exist in RES datasets [26, 28, 41]. However, existing RES models may not be able to well handle novel compositions during test- ing. Here we test the generalization capability of multi- ple state-of-the-art models [25, 37, 42] in terms of handling This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19478 novel compositions, as shown in Table 2. Specifically, we first split each testing set of the RefCOCO dataset. In each testing set, one split subset includes the data samples, in which all the contained compositions are seen in the Ref- COCO training set. While another subset includes the data samples containing novel compositions, of which the indi- vidual components (e.g., words, phrases) exist in the Re- fCOCO training set but the composition itself is unseen in the training set, i.e., containing novel compositions of learned concepts. Then we evaluate the models [25, 37, 42] on the two subsets in each testing set, and find that for each model, its testing performance on the subset contain- ing novel compositions drops obviously compared to the performance on the other subset. For these models, the per- formance gap between the two subsets can reach 14%−17% measured by the metric of overall IoU. Such a clear perfor- mance gap indicates that existing models struggle with gen- eralizing to novel compositions of learned concepts. This might due to that the model does not effectively capture the semantics and visual representations of individual concepts (e.g., “dark”, “coffee” in Fig. 1a) during training. Then the trained model may fail to recognize a novel composi- tion (e.g., “dark coffee”) at testing time, which though is composed of learned concepts. Thus to handle this issue, we aim to train the model to effectively capture the semantics and visual representations of individual concepts during training. Despite the concep- tual simplicity, how to guide the model’s learning behavior towards this goal is a challenging problem. Here from the perspective of meta learning, we propose a Meta Composi- tional Referring Expression Segmentation ( MCRES )frame- work, to effectively handle such a challenging problem by only changing the model training scheme. Meta learning proposes to perform virtual testing during model training for better performance [7, 29]. Inspired by this, to improve the generalization capability of RES mod- els, our MCRES framework incorporates a meta optimiza- tion scheme that consists of three steps: virtual training , virtual testing andmeta update . Specifically, we first split the training set to construct a virtual training set for virtual training, and a virtual testing set for virtual testing. The data samples in the virtual testing set contain novel com- positions w.r.t. the virtual training set. For example, if the expressions of data samples in the virtual training set con- tain both words “dark” and “coffee” but do not contain their composition (i.e., “dark coffee”), the virtual testing set can include this novel composition correspondingly. Based on the constructed virtual training set and virtual testing set, we first train the model using the virtual train- ing set, and then evaluate the trained model on the virtual testing set. During virtual training, the model may learn the compositions of individual concepts as a whole with- out truly understanding the semantics and visual representa-tions of individual concepts, which though can still improve model training performance. For example, if there are many training samples containing the composition of “yellow ba- nana” in the virtual training set, the model can superficially correlate “banana” with “yellow” and learn this composi- tion as a whole, since using such spurious correlations can facilitate the model learning [1, 10, 39]. However, learning the compositions as a whole over the virtual training set may not improve model performance much on the virtual testing set in virtual testing, since the virtual testing set contains novel compositions w.r.t. the virtual training set. Thus to achieve good testing performance on such a virtual testing set, the model needs to effectively capture semantics and visual representations of individual concepts during virtual training. In this way, the model testing performance on the virtual testing set serves as a generalization feedback to the model virtual training process. Thus after the virtual training and virtual testing, we can further update the model to obtain better testing per- formance on the virtual testing set (i.e., meta update), so as to drive the model training on the virtual training set to- wards the direction of learning to capture semantics and vi- sual representations of individual concepts, i.e., learning to learn. In this manner, our framework is able to optimize the model for robust generalization performance, even tackling the challenging testing samples with novel compositions. Moreover, given that expressions can often be hierarchi- cally decomposed, there can exist various levels of novel compositions. Specifically, to identify meaningful compo- sitions in an expression, we can parse an expression into a tree structure based on the constituency parsing tool [9] as shown in Fig.1b. In such a parsing tree, under the same par- ent node, each pair of child nodes (e.g., “white” and “bird”) are closely semantically related, and thus can form a mean- ingful composition. Since each child node can be a word or a phrase as in Fig. 1b, there can naturally exist the following three levels of novel compositions: word-word level (e.g., “white” and “bird”), word-phrase level (e.g., “standing” and “in water”) and phrase-phrase level (e.g., “white bird” and “standing in water”), which correspond to different levels of comprehension complexity. To better handle such a range of novel compositions, we construct multiple virtual test- ing sets in our framework, where each virtual testing set is constructed to handle one level of novel compositions. Our framework only changes the model training scheme without the need to change the model structure. Thus our framework is general, and can be conveniently applied on various RES models. We test our framework on multiple models, and obtain consistent performance improvement. The contributions of our work are threefold: 1) We propose a novel framework (MCRES) to effectively im- prove generalization performance of RES models, espe- cially when handing novel compositions of learned con- 19479 cepts. 2) Via constructing a virtual training set and mul- tiple virtual testing sets w.r.t. various levels of novel com- positions, our framework can train the model to well han- dle various levels of novel compositions. 3) When applied on various models on three RES benchmarks [28, 41], our framework achieves consistent performance improvement.

Zhang_Frame-Event_Alignment_and_Fusion_Network_for_High_Frame_Rate_Tracking_CVPR_2023

Abstract Most existing RGB-based trackers target low frame rate benchmarks of around 30 frames per second. This setting restricts the tracker’s functionality in the real world, espe- cially for fast motion. Event-based cameras as bioinspired sensors provide considerable potential for high frame rate tracking due to their high temporal resolution. However, event-based cameras cannot offer fine-grained texture in- formation like conventional cameras. This unique comple- mentarity motivates us to combine conventional frames and events for high frame rate object tracking under various challenging conditions. In this paper, we propose an end-to- end network consisting of multi-modality alignment and fu- sion modules to effectively combine meaningful information from both modalities at different measurement rates. The alignment module is responsible for cross-style and cross- frame-rate alignment between frame and event modalities under the guidance of the moving cues furnished by events. While the fusion module is accountable for emphasizing valuable features and suppressing noise information by the mutual complement between the two modalities. Exten- sive experiments show that the proposed approach outper- forms state-of-the-art trackers by a significant margin in high frame rate tracking. With the FE240hz dataset, our approach achieves high frame rate tracking up to 240Hz.

1. Introduction Visual object tracking is a fundamental task in computer vision, and deep learning-based methods [7, 9,10,15,35,56] have dominated this field. Limited by the conventional sen- sor, most existing approaches are designed and evaluated on benchmarks [13, 24,38,53] with a low frame rate of approx- imately 30 frames per second (FPS). However, the value of a higher frame rate tracking in the real world has been proved [16, 21–23]. For example, the shuttlecock can reach speeds of up to 493km/h , and analyzing its position is es- sential for athletes to learn how to improve their skills [46]. Utilizing professional high-speed cameras is one strategy ⋆Xin Yang ([email protected]) is the corresponding author. ToMP-MF DeT HMFT FENet AFNet GTFigure 1. A comparison of our AFNet with SOTA trackers. All competing trackers locate the target at time t+ ∆t with conven- tional frames at time tand aggregated events at time t+ ∆t as inputs. Our method achieves high frame rate tracking up to 240Hz on the FE240hz dataset. The two examples also show the comple- mentary benefits of both modalities. (a) The event modality does not suffer from HDR, but the frame does; (b) The frame modality provides rich texture information, while the events are sparse. for high frame rate tracking, but these cameras are inac- cessible to casual users. Consumer devices with cameras, such as smartphones, have made attempts to integrate sen- sors with similar functionalities into their systems. How- ever, these sensors still suffer from large memory require- ments and high power consumption [49]. As bio-inspired sensors, event-based cameras measure light intensity changes and output asynchronous events to represent visual information. Compared with conventional frame-based sensors, event-based cameras offer a high mea- surement rate (up to 1MHz), high dynamic range (140 dB vs. 60 dB), low power consumption, and high pixel band- width (on the order of kHz) [14]. These unique proper- ties offer great potential for higher frame rate tracking in challenging conditions. Nevertheless, event-based cameras cannot measure fine-grained texture information like con- ventional cameras, thus inhibiting tracking performance. Therefore, in this paper, we exploit to integrate the valuable information from event-based modality with that of frame- based modality for high frame rate single object tracking under various challenging conditions. To attain our objective, two challenges require to be ad- dressed: (i) The measurement rate of event-based cameras is much higher than that of conventional cameras. Hence This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9781 for high frame rate tracking, low-frequency frames must be aligned with high-frequency events to disambiguate target locations. Although recent works [34, 45,48,50] have pro- posed various alignment strategies across multiple frames for video-related tasks, they are specifically designed for conventional frames of the same modality at different mo- ments. Thus, applying these approaches directly to our cross-modality alignment does not offer an effective solu- tion. (ii) Effectively fusing complementary information be- tween modalities and preventing interference from noise is another challenge. Recently, Zhang et al. [61] proposed a cross-domain attention scheme to fuse visual cues from frame and event modalities for improving the single object tracking performance under different degraded conditions. However, the tracking frequency is bounded by the conven- tional frame rate since they ignore the rich temporal infor- mation recorded in the event modality. To tackle the above challenges, we propose a novel end- to-end framework to effectively combine complementary information from two modalities at different measurement rates for high frame rate tracking, dubbed AFNet, which consists of two key components for alignment and fusion, respectively. Specifically, (i) we first propose an event- guided cross-modality alignment (ECA) module to simulta- neously accomplish cross-style alignment and cross-frame- rate alignment. Cross-style alignment is enforced by match- ing feature statistics between conventional frame modality and events augmented by a well-designed attention scheme; Cross-frame-rate alignment is based on deformable convo- lution [8] to facilitate alignment without explicit motion es- timation or image warping operation by implicitly focusing on motion cues. (ii) A cross-correlation fusion (CF) mod- ule is further presented to combine complementary infor- mation by learning a dynamic filter from one modality that contributes to the feature expression of another modality, thereby emphasizing valuable information and suppress- ing interference. Extensive experiments on different event- based tracking datasets validate the effectiveness of the pro- posed approach (see Figure 1as an example). In summary, we make the following contributions: •Our AFNet is, to our knowledge, the first to combine the rich textural clues of frames with the high temporal reso- lution offered by events for high frame rate object tracking. •We design a novel event-guided alignment framework that performs cross-modality and cross-frame-rate align- ment simultaneously, as well as a cross-correlation fusion architecture that complements the two modalities. •Through extensive experiments, we show that the pro- posed approach outperforms state-of-the-art trackers in var- ious challenging conditions.

Zhang_Efficient_Map_Sparsification_Based_on_2D_and_3D_Discretized_Grids_CVPR_2023

Abstract Localization in a pre-built map is a basic technique for robot autonomous navigation. Existing mapping and local- ization methods commonly work well in small-scale envi- ronments. As a map grows larger, however, more mem- ory is required and localization becomes inefficient. To solve these problems, map sparsification becomes a prac- tical necessity to acquire a subset of the original map for localization. Previous map sparsification methods add a quadratic term in mixed-integer programming to enforce a uniform distribution of selected landmarks, which requires high memory capacity and heavy computation. In this pa- per, we formulate map sparsification in an efficient linear form and select uniformly distributed landmarks based on 2D discretized grids. Furthermore, to reduce the influence of different spatial distributions between the mapping and query sequences, which is not considered in previous meth- ods, we also introduce a space constraint term based on 3D discretized grids. The exhaustive experiments in different datasets demonstrate the superiority of the proposed meth- ods in both efficiency and localization performance. The relevant codes will be released at https://github. com/fishmarch/SLAM_Map_Compression .

1. Introduction To realize autonomous navigation for robots, localiza- tion in a pre-built map is a basic technique. Lots of algo- rithms have been proposed for mapping and localization us- ing different sensors, including camera [20] and lidar [27]. These algorithms commonly work well in some small-scale environments now. When applied in large-scale environ- ments or in long-term, however, new challenges appear and need to be settled for practical applications. Some of these problems are time and memory consuming. When using cameras, visual simultaneous localization and mapping (SLAM) is a commonly used method to build maps. In visual SLAM, redundant features are extracted from images and then constructed as landmarks in a map, Figure 1. Localization results in a compact map. The original map is constructed from sequence 00 of the KITTI dataset [14]. The original map consists of 141K landmarks, indicated by gray points; while the compact map consists of only 5K landmarks, indicated by blue points. 96.15% of the query images are localized successfully in the compact map, indicted by green poses. such that camera poses can be tracked robustly and accu- rately. These redundant landmarks promise good localiza- tion results in small-scale environments. As more and more images are received when working in large-scale environ- ments or in long-term, however, memory consumption is increasing unboundedly. Localizing in such large maps will also be more time-consuming. These problems are espe- cially severe for some low-cost robots. Actually, not all landmarks are necessary for robots to localize in pre-built maps. In theory, even only 4 matched landmarks can determine a camera pose using EPnP [16] (more landmarks are commonly used for robust and ac- curate estimation). This reveals that maps can be com- pressed and still retain comparable performance for local- ization. Map compression can be classified into two types: descriptor compression [5, 18, 22] and landmark sparsifica- tion [17, 23]. The research of this paper falls in the latter This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12470 one, which is to find a subset of an original map while main- taining comparable localization performance. A subset map is called compact map in this paper. For example, as indi- cated in Fig. 1, only 3.91% of the original landmarks are selected as the compact map, in which more than 96% of the query images can still be localized successfully. To select an optimal subset for localization, map sparsi- fication is related to a K-cover problem [17], which means the number of landmarks in a compact map is minimized while keeping the number of associated landmarks in each image larger than a threshold (i.e., K). To solve the K- cover problem, it can be formulated as mixed-integer lin- ear programming, through which an optimal subset is ob- tained. The original formulation only considers the number of landmarks for localization, while their distribution also affects localization performance. Therefore, some works design and add quadratic terms, formulating mixed-integer quadratic programming to enforce a more uniform distri- bution of selected landmarks [11, 23]. However, these quadratic terms slow down the optimization speed heav- ily. The required high memory capacity and heavy com- putation become severe limitations of these map sparsifica- tion methods. For example, in our experiments, the mixed- integer quadratic programming methods cannot be used for the maps containing more than 55K landmarks, because all computer memory has been consumed. To select uniformly distributed landmarks and at the same time maintain the computation efficiency, we keep map sparsification formulation in a linear form in this pa- per. We firstly discretize images into 2D fix-sized grids. Then for all observed landmarks, we can find which cells they fall in. Therefore, more occupied cells reflect a more uniform distribution of landmarks. This can be formulated in a linear form easily. In this way, uniformly distributed landmarks are selected efficiently, and thus the localization performance in compact maps will be better. Another severe limitation is that all of past works assume the spatial distribution of the query sequence is close to that of the mapping sequence. Then landmarks are selected only based on their association with the images of the mapping sequence. However, this assumption cannot be guaranteed in real robotic applications. The perspective difference be- tween query and mapping sequences may cause the local- ization in compact maps to fail unexpectedly. To ensure more query images from whole 3D space can be localized successfully in compact maps, we propose to select landmarks based on not only their association with mapping images but also their visibility in 3D space. The visibility region of a landmark is defined based on its view- ing angle and distance. The 3D space is also discretized into a 3D discretized grid. For each 3D cell, all visible land- marks are collected, and a constraint on the number of visi- ble landmarks is added into map sparsification formulation.In this way, landmarks are selected to maintain localization performance for query images from the whole space. In summary, the contributions of this paper are as fol- lows: 1) We propose an efficient map sparsification method formulating uniform landmark distribution in a linear form to keep the computation efficiency. 2) We propose to per- form map sparsification involving the visibility of land- marks to achieve better localization results for the query images from the whole space. 3) We conduct exhaustive ex- periments in different datasets and compare with other state- of-the-art methods, showing the effectiveness and superior- ity of our methods for different kinds of query sequences.

Zhang_Skinned_Motion_Retargeting_With_Residual_Perception_of_Motion_Semantics__CVPR_2023

Abstract A good motion retargeting cannot be reached without reasonable consideration of source-target differences on both the skeleton and shape geometry levels. In this work, we propose a novel Residual RET argeting network (R2ET) structure, which relies on two neural modification modules, to adjust the source motions to fit the target skeletons and shapes progressively. In particular, a skeleton-aware mod- ule is introduced to preserve the source motion semantics. A shape-aware module is designed to perceive the geometries of target characters to reduce interpenetration and contact- missing. Driven by our explored distance-based losses that explicitly model the motion semantics and geometry, these two modules can learn residual motion modifications on the source motion to generate plausible retargeted motion in a single inference without post-processing. To balance these two modifications, we further present a balancing gate to conduct linear interpolation between them. Ex- tensive experiments on the public dataset Mixamo demon- strate that our R2ET achieves the state-of-the-art perfor- mance, and provides a good balance between the preserva- tion of motion semantics as well as the attenuation of in- terpenetration and contact-missing. Code is available at https://github.com/Kebii/R2ET .

1. Introduction As a process of mapping the motion of a source charac- ter to a target character without losing plausibility, motion retargeting is a long-standing problem in the community of computer vision and computer graphics. It has a wide spec- trum of applications in game and animation industry, and is a cornerstone of the digital avatar and metaverse tech- nologies [25]. In recent years, learning-based retargeting methods started sparkling in the community. Among them, the neural motion retargeting [1, 16, 24, 25], which has ad- vantages in intelligent perception and stable inference, be- *Most of this work was done during Jiaxu’s internship at Tencent. †Corresponding author: [email protected] 5678910 0 0.5 1 1.5 2 2.5Skeleton -aware Source Copy Ours MSEInterpenetration (%) OursNKN PMnetSAN Motion Retargeting Ours Copy SourceShape -aware Semantics Geometry PMnet*Figure 1. Our R2ET fully considers the source-target differences on both the skeleton and shape geometry levels. The retargeted motion of R2ET preserves motion semantics, eliminates interpen- etration, and keeps self-contact without post-optimizations. comes a new research trend. The previous learning-based methods utilize a full-motion mapping structure, which de- codes joint rotations of the target skeleton as outputs, with joint positions [16, 24, 25] or joint rotations [1] as inputs. However, due to the gap between the Cartesian coordinate space and the rotation space, the full joint position encoding unavoidably introduces motion distortion. Meanwhile, the full joint rotation encoding always leads to discontinuity in the rotation space [1, 26]. In animation, we observe that artists usually copy the motion of the source character, and then manually modify it to preserve motion semantics and avoid translation artifacts, e.g., interpenetration, during motion reuse in new charac- ters. Inspired by this observation, we design a Residual RET argeting network (R2ET) with a residual structure for motion retargeting. This structure takes the source motion as initialization and involves neural networks to imitate the modifications from animators, as illustrated in Figure 1. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13864 With this design, the coherence of the source motion is well maintained, and the search space for retargeting solutions during training is effectively reduced. The key to achieving physically plausible single-body motion retargeting is to understand two main differences be- tween the source and target characters, 1) the differences in bone length ratio; 2) the differences in body shape geome- try. To reach this goal, we explore two modification mod- ules, i.e., the skeleton-aware module and the shape-aware module, to perceive the two differences. On the skeleton level, the skeleton-aware module takes the skeleton configurations as input to assist the transfer of the source motion semantics, such as arm folding and hand clapping, to the target character. To overcome the lack of paired and semantics-correct ground truth, we directly take the supervision signal from the input source motion. The motion semantics is explicitly modeled as a normalized Dis- tance Matrix (DM) of skeleton joints. Accordingly, the se- mantics preservation is achieved by aligning the DM be- tween the source and target motions (Figure 1- Semantics ). On the shape geometry level, the shape-aware module senses the compatibility between the target character mesh and the skeleton adjusted after motion semantics preserva- tion to avoid interpenetration and contact-missing. To train the module end-to-end, we introduce two voxelized Dis- tance Fields, i.e., the Repulsive Distance Field (RDF) and the Attractive Distance Field (ADF) (Figure 1- Geometry ), as the measurement tools for interpenetration and contact. We sample the distance of the query vertices on the target character mesh to the body surface in these two fields to es- timate the degree of interpenetration and contact. With this manner, the whole process is differentiable during training. In practice, we find there always exists a contradic- tion between the preservation of motion semantics and the avoidance of interpenetration. We, therefore, propose a bal- ancing gate to make a trade-off between the skeleton-level and geometry-level modifications by learning an adjusting weight. By leaving the weight to the user, our R2ET also accepts interactive fine control from users. With the above main designs, our R2ET preserves the motion semantics of the source character and avoids in- terpenetration and contact-missing issues in a single-pass without post-processing. We evaluate our method on var- ious complex motion sequences and a wide range of char- acter geometries from skinny to bulky. The qualitative and quantitative results show that our R2ET outperforms the ex- isting learning-based methods by a large margin. The con- tributions of this work are summarized in three-fold: • A novel residual network structure is proposed for neural motion retargeting, which involves a skeleton- aware modification module, a shape-aware modifica- tion module, and a balancing gate. • A normalized joint Distance Matrix is presented toguide the training of the skeleton-aware module for explicit motion semantics modeling, and two Distance Fields are introduced to achieve differentiable pose ad- justment learning. • Extensive experiments on the Mixamo [2] dataset demonstrate that our R2ET achieves the state-of-the- art performance qualitatively and quantitatively.

Yang_Good_Is_Bad_Causality_Inspired_Cloth-Debiasing_for_Cloth-Changing_Person_Re-Identification_CVPR_2023

Abstract Entangled representation of clothing and identity (ID)- intrinsic clues are potentially concomitant in conventional person Re-IDentification (ReID). Nevertheless, eliminating the negative impact of clothing on ID remains challenging due to the lack of theory and the difficulty of isolating theexact implications. In this paper , a causality-based Auto- Intervention Model, referred to as AIM1, is first proposed to mitigate clothing bias for robust cloth-changing person ReID (CC-ReID). Specifically, we analyze the effect of clothing on the model inference and adopt a dual-branch model to simulate causal intervention. Progressively, clothing biasis eliminated automatically with model training. AIM is encouraged to learn more discriminative ID clues that are free from clothing bias. Extensive experiments on two stan- dard CC-ReID datasets demonstrate the superiority of the proposed AIM over other state-of-the-art methods.

1. Introduction Short-term person Re-IDentification (ReID) aims to match a person within limited time and space conditions, under the assumption that each individual maintains clothing consistency. Of both traditional and deep learning methods, the best way to deceive current ReID models is by having pedestrians alter their clothing. This highlights the inade-quacy of existing short-term ReID methods [ 4,42,45]. To solve this issue, Cloth-Changing person ReID (CC-ReID) [ 2] has been recently explored, which is increasingly critical in public security systems for tracking down disguised criminal suspects. For example, witnesses typically provide descrip- tive details ( e.g., clothing, color, and stature) when describ- ing suspects, but it is unlikely that criminals will wear the same clothes upon their reappearance. It follows that cloth- ing information will disrupt the existing ReID system [ 40], *Corresponding author: [email protected] will publicly available at https://github.com/BoomShakaY/AIM-CCReID .ID classification ID B (b) Clothing affect person recognition ID A ID C ProbeGallery match 0.20.40.60.8 ID C(a) Entangled representation of clothing and ID ID Non-ID (Clothing, et al.) Entangled Camera #1 ID Non-ID (Clothing, et al.) Latent space Person #1Entangled match Person #1Camera #2 Figure 1. Illustration of the entangled representation of clothing and ID and how clothing bias affects model prediction. which leads to a growing need felt by the research commu- nity to study CC-ReID task. As one of the accompanying objects of people, clothing is an essential factor in social life. There are two possible responses when people identify others: confusing the per-ception of identity (ID) or clearly distinguishing differentIDs through immutable visual appearance (faces or soft- biometrics). The former manifests as the mix-up of IDs due to the similarity in flexible visual appearance ( e.g., cloth- ing) of people, while the latter is caused by the high-level semantic ( e.g., ID-clues) perceived by humans, transcending the similarity that comes with clothing. The above reactions reflect that the relevance of clothing to ID is a double-edged sword. Traditionally, clothing is a helpful characteristic for ReID, where people wearing the same clothes are likely to have the same ID. However, entangled representation of clothing and ID leads the statistical-based neural networks to converge towards shallow clothing features that can be easily distinguished. This statistical association gives theReID model a faulty perception that there is a strong rela- tion between clothing and ID, which would undermine the ultimate prediction for seeking robust and sensible results. Recent years have witnessed numerous deep learning at- tempts to model discriminative clues for person distinguish- able learning. However, plenty of misleading information exists in these attempts, as some non-ID areas ( e.g., clothing This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1472 and background) may correlate highly with ID. As shown in Fig. 1(a), conventional ReID methods focus on image regions with distinct discrimination characteristics [ 33,46], leading to the entanglement of clothing and ID-intrinsic clues. In CC-ReID, this phenomenon will lead to biasedclassification, misleading the ReID model to focus on the non-ID areas that appear to be ID-related. As shown in Fig. 1(b), clothing may mislead the classifier by giving high scores to person images with similar colors or patterns, but ignoring the faces and details that matter. To this end, if clothing bias can be captured and removed from the exist-ing model, it will enhance the contribution of genuinely ID-relevant features to ID discrimination. Lacking practical tools to alleviate clothing bias makes it challenging to correct the misleading attention of the current ReID model. Even knowing that clothing is a critical influ- encing factor, it is not apparent how to intervene in clothing directly in the representation space. Not to mention thatrough negligence on clothing will damage the integrity of person representation, e.g., the mask-based [ 12,16,41] and gait-based methods [ 6,17] try to obtain cloth-agnostic rep- resentations by forcibly covering up or neglecting clothing information. While effective, these straightforward methods lose a plethora of semantic information and overlook the factual relation between clothing and real ID. Causal inference is a recently emerging theory [ 19] widely used to extract causality [ 10] and explore the true association between two events [ 25]. Thanks to causal in- ference, we can revisit the issue of clothing bias from a causal perspective. The ladder of causality [ 7] divides cogni- tive abilities into three levels from low to high: association, intervention, and counterfactual analysis. Many aforemen- tioned research works explore CC-ReID from the surface level of data association, while more advanced cognitive is not covered. Intervention allows us to incorporate clothing knowledge into the model prediction and eliminate the cor- responding effects, in contrast to counterfactual data, which are difficult to obtain under strict variable control. Therefore, this paper attempts to start with intervention by examining the perturbation of clothing on the results and removing such perturbation from model predictions. Through the causalintervention, we attempt to remove the effect of clothing without destroying semantic integrity and further optimizing the learned discriminative features. To bring the theoretical intervention into practice, we de- sign a dual-branch model to capture clothing bias and ID clues separately and strip clothing inference from ID repre- sentation learning to simulate the entire intervention process. The clothing branch represents the model’s perception of clothing, breaking the false association between clothing and ID brought by the entangled representation. Subsequently, while maintaining semantic integrity, this paper achieves bias elimination and improves the robustness of ID representa-tion by constantly mitigating the influence of clothing on ID classification. Further, to improve the accuracy of clothing bias distillation, as clothing has top-middle-bottom charac- teristics, this paper adopts pyramid matching strategy [ 8]t o enhance the partial feature representation of clothing. Ad- ditionally, we introduce two learning objectives explicitly designed to encourage clothing mitigation. A knowledge transfer objective is adopted to strengthen the perception of clothing bias entangled with ID-intrinsic representation. A bias elimination objective is utilized to cooperate with the causal auto-intervention for ID-intrinsic feature extraction. Our contributions can be summarized threefold: •We propose a novel causality-based Auto-Intervention Model (AIM) for Cloth-Changing person Re- IDentification (CC-ReID). The proposed AIM guarantees that the learned representation is unaffectedby clothing bias. To the best of our knowledge, AIM is the first model to introduce causality into CC-ReID. •A dual-branch model is proposed to simulate the causal intervention. Clothing bias is gradually stripped from the entangled ID-clothing representation without de-stroying semantic integrity, which optimizes the ID- intrinsic feature learning. •We comprehensively demonstrate how clothing bias affects the current ReID model and highlight the signif- icance of causal inference in CC-ReID. The experimen- tal results on two CC-ReID datasets, PRCC-ReID [ 38] and LTCC-ReID [ 26], show that AIM outperforms state- of-the-art methods.

Yang_Semantic_Human_Parsing_via_Scalable_Semantic_Transfer_Over_Multiple_Label_CVPR_2023

Abstract This paper presents Scalable Semantic Transfer (SST), a novel training paradigm, to explore how to leverage the mutual benefits of the data from different label domains (i.e. various levels of label granularity) to train a powerful hu- man parsing network. In practice, two common applica- tion scenarios are addressed, termed universal parsing and dedicated parsing , where the former aims to learn homoge- neous human representations from multiple label domains and switch predictions by only using different segmentation heads, and the latter aims to learn a specific domain pre- diction while distilling the semantic knowledge from other domains. The proposed SST has the following appealing benefits: (1)it can capably serve as an effective train- ing scheme to embed semantic associations of human body parts from multiple label domains into the human repre- sentation learning process; (2)it is an extensible semantic transfer framework without predetermining the overall rela- tions of multiple label domains, which allows continuously adding human parsing datasets to promote the training. (3) the relevant modules are only used for auxiliary training and can be removed during inference, eliminating the extra reasoning cost. Experimental results demonstrate SST can effectively achieve promising universal human parsing per- formance as well as impressive improvements compared to its counterparts on three human parsing benchmarks (i.e., PASCAL-Person-Part, ATR, and CIHP). Code is available athttps://github.com/yangjie-cv/SST .

1. Introduction Human parsing, which aims to assign pixel-wise cate- gory predictions for human body parts, has played a criti- cal role in human-oriented visual content analysis, editing, and generation, e.g., virtual try-on [7], human motion trans- fer [14], and human activity recognition [8]. Prior methods *Corresponding author.have chronically been dominated by developing powerful network architectures [11, 18, 24, 25, 32]. However, these network designs usually serve as general components for human parsing in a specific labeling system, and how to use the data from multiple label domains to further improve the accuracy of the parsing network is still an open issue. Recent studies [12, 20] have proposed using multiple la- bel domains to train a universal parsing network by con- structing graph neural networks that learn semantic associ- ations from different label domains. Graphonomy [10, 20] investigates cross-domain semantic coherence through se- mantic aggregation and transfer on explicit graph structures. Grapy-ML [12] extends this method by stacking three levels of human part graph structures, from coarse to fine granu- larity. Despite these graph-based methods consider prior human body structures, such as the spatial associations of human parts and the semantic associations of different la- bel domains as in Fig. 1-(a), they have limitations in their ability to dynamically add new label domains due to their dependence on pre-determined graph structures. Addition- ally, explicit graph reasoning incurs extra inference costs, making it impractical in some scenarios. Therefore, there is a need for a more general, scalable, and simpler manner to harness the mutual benefits of data from multiple label domains to enhance given human parsing networks. In this work, we propose Scalable Semantic Transfer (SST), a novel training paradigm that builds plug-and-play modules to effectively embed semantic associations of hu- man body parts from multiple label domains into the given parsing network. The proposed SST scheme can be eas- ily applied to two common scenarios, i.e., universal pars- ing and dedicated parsing. Similar to [12, 20], universal parsing aims to enforce the parsing network to learn homo- geneous human representation from multiple label domains with the help of the SST scheme in the training phase. It achieves different levels of human parsing prediction for an arbitrary image in the inference phase , as illustrated in Fig- ure 1-(b). In contrast, dedicated parsing aims to optimize a parsing network for a specific label domain by employ- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19424 Figure 1. (a) The Motivation of SST: 1) Intra-Domain Hierarchical Semantic Structure (Upper): The spatial correlated two parts are connected by a white line; 2) Cross-Domain Semantic Mapping and Consistency (Lower): We regularize the semantic consistency (orange line) between original category representations (solid line) and mapped ones (dot line) for two datasets (green for dataset-1 and blue for dataset-2). (b) The Training and Inference Phase of Universal Parsing. (c) The Training and Inference Phase of Dedicated Parsing. ing the SST scheme to distill the semantic knowledge from other label domains in the training phase, and finally real- izes more accurate prediction for the specific label domain in the inference phase, as shown in Fig. 1-(c). The proposed SST has several appealing benefits: (1) Generality. It is a general training scheme that effectively incorporates prior knowledge of human body parts into the representation learning process of a given parsing network. It enables the integration of both intra-domain spatial cor- relations and inter-domain semantic associations, making it suitable for training both universal and dedicated parsing networks. (2) Scalability. It is a scalable framework that enables the addition of new datasets for more effective train- ing. In practice, universal parsing enables the parsing net- work to acquire a more robust human representation by ex- panding the label domain pool, while dedicated parsing fa- cilitates the transfer of rich semantic knowledge from more datasets to enhance the parsing network for a specific label domain. (3) Simplicity. All modules in SST are plug-and- play ones that are only used for auxiliary training and can be dropped out during inference, eliminating unnecessary model complexity in real-world applications. The main contributions are summarized as follows: 1) We propose a novel training paradigm called Scalable Semantic Transfer (SST) that effectively leverages the benefits of data from different label domains to train powerful universal and dedicated parsing networks. 2) We introduce three plug-and-play modules in SST that incorporate prior knowledge of human body parts into the training process of a given parsing network, which can be removed during inference. 3) Extensive experiments demonstrate the effectiveness of SST on both universal and dedicated parsing tasks, achieving impressive results on three human parsing datasets compared with the counterparts.

Zhang_Backdoor_Defense_via_Deconfounded_Representation_Learning_CVPR_2023

Abstract Deep neural networks (DNNs) are recently shown to be vulnerable to backdoor attacks, where attackers embed hid- den backdoors in the DNN model by injecting a few poi- soned examples into the training dataset. While exten- sive efforts have been made to detect and remove back- doors from backdoored DNNs, it is still not clear whether a backdoor-free clean model can be directly obtained from poisoned datasets. In this paper, we first construct a causal graph to model the generation process of poisoned data and find that the backdoor attack acts as the confounder, which brings spurious associations between the input im- ages and target labels, making the model predictions less reliable. Inspired by the causal understanding, we pro- pose the Causality-inspired Backdoor Defense (CBD), to learn deconfounded representations for reliable classifi- cation. Specifically, a backdoored model is intentionally trained to capture the confounding effects. The other clean model dedicates to capturing the desired causal effects by minimizing the mutual information with the confounding representations from the backdoored model and employ- ing a sample-wise re-weighting scheme. Extensive exper- iments on multiple benchmark datasets against 6 state-of- the-art attacks verify that our proposed defense method is effective in reducing backdoor threats while maintaining high accuracy in predicting benign samples. Further anal- ysis shows that CBD can also resist potential adaptive at- tacks. The code is available at https://github.com/ zaixizhang/CBD .

1. Introduction Recent studies have revealed that deep neural networks (DNNs) are vulnerable to backdoor attacks [18, 38, 53], *Qi Liu is the corresponding author. Trigger Pattern ImageDog Turtle Cat Predicted LabelDNN Classifier (a) Illustration of Backdoor Attack (b) Causal Graph X YBFigure 1. (a) A real example of the backdoor attack. The back- doored DNN classifies the “turtle” image with a trigger pattern as the target label “dog”. (b) The causal graph represents the causalities among variables: Xas the input image, Yas the la- bel, and Bas the backdoor attack. Besides the causal effect of X onY(X→Y), the backdoor attack can attach trigger patterns to images ( B→X), and change the labels to the targeted label (B→Y). Therefore, as a confounder , the backdoor attack B opens a spurious path between XandY(X←B→Y). where attackers inject stealthy backdoors into DNNs by poi- soning a few training data. Specifically, backdoor attack- ers attach the backdoor trigger ( i.e.,a particular pattern) to some benign training data and change their labels to the attacker-designated target label. The correlations between the trigger pattern and target label will be learned by DNNs during training. In the inference process, the backdoored model behaves normally on benign data while its predic- tion will be maliciously altered when the backdoor is ac- tivated. The risk of backdoor attacks hinders the applica- tions of DNNs to some safety-critical areas such as auto- matic driving [38] and healthcare systems [14]. On the contrary, human cognitive systems are known to be immune to input perturbations such as stealthy trigger patterns induced by backdoor attacks [17]. This is because humans are more sensitive to causal relations than the sta- tistical associations of nuisance factors [29,60]. In contrast, deep learning models that are trained to fit the poisoned 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12228 datasets can hardly distinguish the causal relations and the statistical associations brought by backdoor attacks. Based on causal reasoning, we can identify causal relation [50,52] and build robust deep learning models [70, 71]. Therefore, it is essential to leverage causal reasoning to analyze and mitigate the threats of backdoor attacks. In this paper, we focus on the image classification tasks and aim to train backdoor-free models on poisoned datasets without extra clean data. We first construct a causal graph to model the generation process of backdoor data where nuisance factors ( i.e.,backdoor trigger patterns) are consid- ered. With the assistance of the causal graph, we find that the backdoor attack acts as the confounder and opens a spu- rious path between the input image and the predicted label (Figure 1). If DNNs have learned the correlation of such a spurious path, their predictions will be changed to the target labels when the trigger is attached. Motivated by our causal insight, we propose Causality- inspired Backdoor Defense (CBD) to learn deconfounded representations for classification. As the backdoor attack is stealthy and hardly measurable, we cannot directly block the backdoor path by the backdoor adjustment from causal inference [52]. Inspired by recent advances in disentan- gled representation learning [20, 36, 66], we instead aim to learn deconfounded representations that only preserve the causality-related information. Specifically in CBD, two DNNs are trained, which focus on the spurious correlations and the causal effects respectively. The first DNN is de- signed to intentionally capture the backdoor correlations with an early stop strategy. The other clean model is then trained to be independent of the first model in the hidden space by minimizing mutual information. The information bottleneck strategy and sample-wise re-weighting scheme are also employed to help the clean model capture the causal effects while relinquishing the confounding factors. After training, only the clean model is used for downstream clas- sification tasks. In summary, our contributions are: • From a causal perspective, we find the backdoor at- tack acts as the confounder that causes spurious corre- lations between the input images and the target label. • With the causal insight, we propose a Causality- inspired Backdoor Defense (CBD), which learns de- confounded representations to mitigate the threat of poisoning-based backdoor attacks. • Extensive experiments with 6 representative backdoor attacks are conducted. The models trained using CBD are of almost the same clean accuracy as they were directly trained on clean data and the average backdoor attack success rates are reduced to around 1 %, which verifies the effectiveness of CBD. • We explore one potential adaptive attack against CBD,which tries to make the backdoor attack stealthier by adversarial training. Experiments show that CBD is robust and resistant to such an adaptive attack.

Zhai_Feature_Representation_Learning_With_Adaptive_Displacement_Generation_and_Transformer_Fusion_CVPR_2023

Abstract Micro-expressions are spontaneous, rapid and subtle fa- cial movements that can neither be forged nor suppressed. They are very important nonverbal communication clues, but are transient and of low intensity thus difficult to rec- ognize. Recently deep learning based methods have been developed for micro-expression (ME) recognition using fea- ture extraction and fusion techniques, however, targeted feature learning and efficient feature fusion still lack fur- ther study according to the ME characteristics. To address these issues, we propose a novel framework Feature Repre- sentation Learning with adaptive Displacement Generation and Transformer fusion (FRL-DGT), in which a convolu- tional Displacement Generation Module (DGM) with self- supervised learning is used to extract dynamic features from onset/apex frames targeted to the subsequent ME recogni- tion task, and a well-designed Transformer Fusion mecha- nism composed of three Transformer-based fusion modules (local, global fusions based on AU regions and full-face fu- sion) is applied to extract the multi-level informative fea- tures after DGM for the final ME prediction. The extensive experiments with solid leave-one-subject-out (LOSO) eval- uation results have demonstrated the superiority of our pro- posed FRL-DGT to state-of-the-art methods.

1. Introduction As a subtle and short-lasting change, micro-expression (ME) is produced by unconscious contractions of facial muscles and lasts only 1/25th to 1/5th of a second, as illus- trated in Figure 1, revealing a person’s true emotions under- neath the disguise [8,35]. The demands for ME recognition technology are becoming more and more extensive [2, 18], *Corresponding author. Figure 1. A video sequence depicting the order of which onset, apex and offset frames occur. Sample frames are from a “surprise” sequence in CASME II. Our goal is to design a novel feature rep- resentation learning method based on an onset-apex frame pair for facial ME recognition. (Images from CASME II ©Xiaolan Fu) including multimedia entertainment, film-making, human- computer interaction, affective computing, business nego- tiation, teaching and learning, etc. Since MEs have invol- untary muscle movements with short duration and low in- tensity in nature, the research of ME is attractive but diffi- cult [1, 22]. Therefore, it is crucial and desired to extract robust feature representations to conduct ME analyses. A lot of feature representation methods are already avail- able including those relying heavily on hand-crafted fea- tures with expert experiences [4, 11, 25] and deep learning techniques [33, 36, 40]. However, the performance of deep learning networks is still restricted for ME classification, mainly due to the complexity of ME and insufficient train- ing data [28, 47]. Deep learning methods can automatically extract optimal features and offer an end-to-end classifica- tion, but in the existing solutions, dynamic feature extrac- tion is only taken as a data preprocessing strategy. It is not integrated with the subsequent neural network, thus failing to adapt the generated dynamic features to a specific train- ing task, leading to redundancy or missing features. Such shortcoming motivates us to design a dynamic feature ex- tractor to adapt the subsequent ME recognition task. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22086 In this paper, we propose a novel end-to-end feature rep- resentation learning framework named FRL-DGT, which is constructed with a well-designed adaptive Displacement Generation Module (DGM) and a subsequent Transformer Fusion mechanism composed of the Transformer-based lo- cal, global, and full-face fusion modules, as illustrated in Figure 2, for ME feature learning to model global informa- tion while focusing on local features. Our FRL-DGT only requires a pair of images, i.e., the onset and the apex frames, which are extracted from the frame sequence. Unlike the previous methods which extract optical flow or dynamic imaging features, our DGM and corresponding loss functions are designed to generate the displacement be- tween expression frames, using a convolution module in- stead of the traditional techniques. The DGM is involved in training with the subsequent ME classification module, and therefore its parameters can be tuned based on the feed- back from classification loss to generate more targeted dy- namic features adaptively. We shall emphasize that the la- beled training data for ME classification is very limited and therefore the supervised data for our DGM is insufficient. To handle this case, we resort to a self-supervised learning strategy and sample sufficient additional random pairs of image sequence as the extra training data for the DGM, so that it is able to fully extract the necessary dynamic features adaptively for the subsequent ME recognition task. Regarding fusing the dynamic features extracted from DGM, we first adopt the AU (Action Unit) region partition- ing method from FACS (Facial Action Coding System) [49] to get 9 AUs, and then crop the frames and their displace- ments into blocks based on the 9 AUs and the full-face re- gion as input to the Transformer Fusion. We argue that the lower layers in the Transformer Fusion should encode and fuse different AU region features in a more targeted way, while the higher layers can classify MEs based on the infor- mation of all AUs. We propose a novel fusion layer with at- tentions as a linear fusion before attention mechanism [45], aiming at a more efficient and accurate integration of the embedding vectors. The fusion layers are interleaved with Transformer’s basic blocks to form a new multi-level fusion module for classification to ensure it to better learn global information and long-term dependencies of ME. To summarize, our main contributions are as follows: • We propose a novel end-to-end network FRL-DGT which fully explores AU regions from onset-apex pair and the displacement between them to extract compre- hensive features via Transformer Fusion mechanism with the Transformer-based local, global, and full-face feature fusions for ME recognition. • Our DGM is well-trained with self-supervised learn- ing, and makes full use of the subsequent classifica- tion supervision information in the training phase, sothat the trained DGM model is able to generate more targeted ME dynamic features adaptively. • We present a novel fusion layer to exploit the linear fusion before attention mechanism in Transformer for fusing the embedding features at both local and global levels with simplified computation. • We demonstrate the effectiveness of each module with ablation study and the outperformance of the proposed FRL-DGT to the SOTA methods with extensive exper- iments on three popular ME benchmark datasets.

Zhang_Ref-NPR_Reference-Based_Non-Photorealistic_Radiance_Fields_for_Controllable_Scene_Stylization_CVPR_2023

Abstract Current 3D scene stylization methods transfer textures and colors as styles using arbitrary style references, lack- ing meaningful semantic correspondences. We introduce Reference-Based Non-Photorealistic Radiance Fields (Ref- NPR) to address this limitation. This controllable method stylizes a 3D scene using radiance fields with a single styl- ized 2D view as a reference. We propose a ray registra- tion process based on the stylized reference view to ob- tain pseudo-ray supervision in novel views. Then we ex- ploit semantic correspondences in content images to fill oc- cluded regions with perceptually similar styles, resulting in non-photorealistic and continuous novel view sequences. Our experimental results demonstrate that Ref-NPR out- performs existing scene and video stylization methods re- garding visual quality and semantic correspondence. The code and data are publicly available on the project page at https://ref-npr.github.io .

1. Introduction In the past decade, there has been a rising demand for stylizing and editing 3D scenes and objects in various fields, including augmented reality, game scene design, and digitalartwork. Traditionally, professionals achieve these tasks by creating 2D reference images and converting them into styl- ized 3D textures. However, establishing direct cross-modal correspondence is challenging and often requires significant time and effort to obtain stylized texture results similar to the 2D reference schematics. A critical challenge in the 3D stylization problem is to ensure that stylized results are perceptually similar to the given style reference. Benefiting from radiance fields [2, 10, 28, 29, 37, 46], recent novel-view stylization meth- ods [5,8,15,16,30,45] greatly facilitated style transfer from an arbitrary 2D style reference to 3D implicit representa- tions. However, these methods do not provide explicit con- trol over the generated results, making it challenging to specify the regions where certain styles should be applied and ensure the visual quality of the results. On the other hand, reference-based video stylization methods allow for the controllable generation of stylized novel views with bet- ter semantic correspondence between content and style ref- erence, as demonstrated in works like [17, 38]. However, these methods may diverge from the desired style when styl- izing a frame sequence with unseen content, even with the assistance of stylized keyframes. To address the aforementioned limitations, we propose This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4242 a new paradigm for stylizing 3D scenes using a single stylized reference view. Our approach, called Reference- Based Non-Photorealistic Radiance Fields (Ref-NPR), is a controllable scene stylization method that takes advantage of volume rendering to maintain cross-view consistency and establish semantic correspondence in transferring style across the entire scene. Ref-NPR utilizes stylized views from radiance fields as references instead of arbitrary style images to achieve both flexible controllability and multi-view consistency. A reference-based ray registration process is designed to project the 2D style reference into 3D space by utilizing the depth rendering of the radiance field. This process provides pseudo-ray supervision to maintain geometric and percep- tual consistency between stylized novel views and the styl- ized reference view. To obtain semantic style correspon- dence in occluded regions, Ref-NPR performs template- based feature matching, which uses high-level semantic fea- tures as implicit style supervision. The correspondence in the content domain is utilized to select style features in the given style reference, which are then used to transfer style globally, especially in occluded regions. By doing so, Ref- NPR generates the entire stylized 3D scene from a single stylized reference view. Ref-NPR produces visually appealing stylized views that maintain both geometric and semantic consistency with the given style reference, as presented in Fig. 1. The generated stylized views are perceptually consistent with the refer- ence while also exhibiting high visual quality across various datasets. We have demonstrated that Ref-NPR, when using the same stylized view as reference, outperforms state-of- the-art scene stylization methods [30, 45] both qualitatively and quantitatively. In summary, our paper makes three contributions. Firstly, we introduce a new paradigm for stylizing 3D scenes that allows for greater controllability through the use of a stylized reference view. Secondly, we propose a novel approach called Ref-NPR, consisting of a reference- based ray registration process and a template-based feature matching scheme to achieve geometrically and percep- tually consistent stylizations. Finally, our experiments demonstrate that Ref-NPR outperforms state-of-the-art scene stylization methods such as ARF and SNeRF both qualitatively and quantitatively. More comprehensive results and a demo video can be found in the supplementary material and on our project page.

Zhang_Real-Time_Controllable_Denoising_for_Image_and_Video_CVPR_2023

Abstract Controllable image denoising aims to generate clean samples with human perceptual priors and balance sharp- ness and smoothness. In traditional filter-based denoising methods, this can be easily achieved by adjusting the fil- tering strength. However, for NN (Neural Network)-based models, adjusting the final denoising strength requires per- forming network inference each time, making it almost im- possible for real-time user interaction. In this paper, we in- troduce Real-time Controllable Denoising (RCD), the first deep image and video denoising pipeline that provides a fully controllable user interface to edit arbitrary denois- ing levels in real-time with only one-time network inference. Unlike existing controllable denoising methods that require multiple denoisers and training stages, RCD replaces the last output layer (which usually outputs a single noise map) of an existing CNN-based model with a lightweight module that outputs multiple noise maps. We propose a novel Noise Decorrelation process to enforce the orthogonality of the noise feature maps, allowing arbitrary noise level control through noise map interpolation. This process is network- free and does not require network inference. Our experi- ments show that RCD can enable real-time editable image and video denoising for various existing heavy-weight mod- els without sacrificing their original performance.

1. Introduction Image and video denoising are fundamental problems in computational photography and computer vision. With the development of deep neural networks [12, 26, 49, 59], model-based denoising methods have achieved tremendous success in generating clean images and videos with superior denoising scores [4,55,57]. However, it should be noted that the improvement in reconstruction accuracy (e.g., PSNR, SSIM) is not always accompanied by an improvement in visual quality, which is known as the Perception-Distortion trade-off [6]. In traditional denoising approaches, we can easily adjust the denoising level by tuning related control *Equal contribution. Perception DistortionReal-time tuning for user preference GT Initial Results T uning A BC D E Intensity PerceptionNoisy SampleFigure 1. Real-time controllable denoising allows users further tuning the restored results to achieve Perception-Distortion trade- off. A-B: tuning with changing denoising intensity. C-E: tuning without changing denoising intensity. parameters and deriving our preferred visual results. How- ever, for typical deep network methods, we can only restore the degraded image or video to a fixed output with a prede- termined restoration level. In recent years, several modulation methods have been proposed to generate continuous restoration effects be- tween two pre-defined denoising levels. These methods can be categorized into two kinds: interpolation-based meth- ods [17,24,50,51], which use deep feature interpolation lay- ers, and condition-network-based methods, which import an extra condition network for denoising control [9, 25, 39]. Essentially, both types of methods are designed based on the observation that the outputs of the network change con- tinuously with the modulation of features/filters. This ob- servation enables deep denoising control, but it also intro- duces several limitations. First, there is a lack of explain- ability , as the relationship between the control parameters (how to modulate features) and the control operation (how the network outputs are changed) is unclear [24]. This in- dicates that black-box operators (network layers) must be used to encode them. Second, the use of control parame- ters as network inputs requires entire network propagation each time control parameters change, resulting in a lack of efficiency . Lastly, current modulation methods often re- quire an explicit degradation level during training, which is hard to obtain for real-world samples . As a result, cur- rent controllable denoising methods only focus on synthetic This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14028 ......Noisy Image NoisesParam 0 Denoising NetworkNoisy Image ...... 10 20 30 40 50Editable Noises Clean Image 0Contr ol NetworksParam 0 Param 1 Param 2 Clean Image 0Clean Image 1Clean Image 2Black-box operator: "Translating" control parameters to networksContr ol Parameters Contr ol Parameters Conventional Contr ollable DenoisingParam 0 Param 1 Param 2Param 0 Param 1 Param 2 RCD (ours)Noise LevelFigure 2. Comparison of pipelines between conventional control- lable denoising and our RCD. RCD achieves real-time noise con- trol by manipulating editable noises directly. noise benchmarks. Furthermore, both interpolation-based and condition-network-based methods have their own draw- backs. Interpolation-based methods often require multi- ple training stages, including pretraining two basic models (start level and end level). On the other hand, condition- network-based methods are strenuous to jointly optimize the base network and the condition network. In this paper, we research on the problem: Can we achieve real-time controllable denoising that abandons the auxiliary network and requires no network forward propa- gation for changing restoration effects at test time? Towards this goal, we propose Real-time Control- lable Denoising method (RCD), a lightweight pipeline for enabling rapid denoising control to achieve Perception- Distortion Balance (See Fig. 1). Our RCD can be plugged into any noise-generate-based restoration methods [11, 46, 54, 55] with just a few additional calculations. Specifi- cally, we replace the last layer of an existing denoising network (which usually outputs a single noise map) with a lightweight module that generates multiple noise maps with different noise levels. We utilize a novel Noise Decorrela- tion process to enforce the orthogonality of the noise distri- bution of these noise maps during training. As a result, we can attain arbitrary denoising effects by simple linear inter- polation of these noise maps. Since this process does not require network inference, it makes real-time user interac- tion possible even for heavy denoising networks. Fig. 2 illustrates the fundamental differences between our RCD approach and conventional controllable denoising methods. In contrast to traditional methods that rely on con- trol networks, the RCD pipeline generates editable noises of varying intensities/levels, providing explicit control by ex- ternal parameters and enabling network-free, real-time de- noising editing. Real-time editing capabilities offered by RCD create new opportunities for numerous applications that were previously impossible using conventional tech- niques, such as online video denoising editing, even during playback (e.g., mobile phone camera video quality tuning for ISP tuning engineers), as well as deploying controllable denoising on edge devices and embedded systems. Since the editing stage of RCD only involves image interpolation,users can edit their desired results on low-performance de- vices without the need for GPUs/DSPs. Moreover, unlike previous methods that only support changing noise levels, RCD allows users to adjust denois- ing results at a specific noise level by providing a new in- terface to modify the noise generation strategy. RCD is also the first validated method for controllable denoising on real-world benchmarks. It is noteworthy that existing con- trollable methods typically require training data with fixed- level noise to establish their maximum and minimum noise levels, which makes them unsuitable for most real-world benchmarks comprising data with varying and unbalanced noise levels. Our main contributions can be summarized as follows: • We propose RCD, a controllable denoising pipeline that firstly supports real-time denoising control (> 2000×speedup compared to conventional controllable methods) and larger control capacity (more than just intensity) without multiple training stages [24] and auxiliary networks [50]. • RCD is the first method supporting controllable de- noising on real-world benchmarks. • We propose a general Noise Decorrelation technique to estimate editable noises. • We achieve comparable or better results on widely- used real/synthetic image-denoising and video- denoising datasets with minimal additional computa- tional cost.

Yan_A_Unified_HDR_Imaging_Method_With_Pixel_and_Patch_Level_CVPR_2023

Abstract Mapping Low Dynamic Range (LDR) images with differ- ent exposures to High Dynamic Range (HDR) remains non- trivial and challenging on dynamic scenes due to ghosting caused by object motion or camera jitting. With the success of Deep Neural Networks (DNNs), several DNNs-based methods have been proposed to alleviate ghosting, they can- not generate approving results when motion and saturation occur. To generate visually pleasing HDR images in various cases, we propose a hybrid HDR deghosting network, called HyHDRNet, to learn the complicated relationship between reference and non-reference images. The proposed HyH- DRNet consists of a content alignment subnetwork and a Transformer-based fusion subnetwork. Specifically, to ef- fectively avoid ghosting from the source, the content align- ment subnetwork uses patch aggregation and ghost atten- tion to integrate similar content from other non-reference images with patch level and suppress undesired compo- nents with pixel level. To achieve mutual guidance between patch-level and pixel-level, we leverage a gating module to sufficiently swap useful information both in ghosted and saturated regions. Furthermore, to obtain a high-quality HDR image, the Transformer-based fusion subnetwork uses a Residual Deformable Transformer Block (RDTB) to adap- tively merge information for different exposed regions. We examined the proposed method on four widely used public HDR image deghosting datasets. Experiments demonstrate that HyHDRNet outperforms state-of-the-art methods both quantitatively and qualitatively, achieving appealing HDR visualization with unified textures and colors.

1. Introduction Natural scenes cover a very broad range of illumina- tion, but standard digital camera sensors can only measure *†The first three authors contributed equally to this work. This work was partially supported by NSFC (U19B2037, 61901384), Natu- ral Science Basic Research Program of Shaanxi (2021JCW-03, 2023-JC- QN-0685), the Fundamental Research Funds for the Central Universi- ties (D5000220444), and National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology. Corre- sponding author: Jinqiu Sun. Figure 1. Our approach produces high-quality HDR images, lever- aging both patch-wise aggregation and pixel-wise ghost atten- tion. The two modules provide complementary visual information: patch aggregation recovers patch-level content of the complex dis- torted regions and ghost attention provides pixel-level alignment. a limited dynamic range. Images captured by cameras of- ten have saturated or under-exposed regions, which lead to terrible visual effects due to severely missing details. High Dynamic Range (HDR) imaging has been developed to ad- dress these limitations, and it can display richer details. A common way of HDR imaging is to fuse a series of differ- ently exposed Low Dynamic Range (LDR) images. It can recover a high-quality HDR image when both the scene and the camera are static, however, it suffers from ghosting arti- facts on dynamic objects or hand-held camera scenarios. Several methods have been proposed to alleviate these problems, including alignment-based methods [1, 7, 20], rejection-based methods [3, 8, 16, 18, 33] and patch-based methods [5, 13, 19]. The alignment-based methods employ global ( e.g., homographies) or non-rigid alignment ( e.g., optical flow) to rectify the content of motion regions, but they are error-prone and cause ghosts due to saturation and occlusion. The rejection-based methods attempt to remove the motion components from the exposed images and re- place them with the content of the reference image. Al- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22211 though these methods achieve good quality for static scenes, they discard the misalignment regions, which causes insuf- ficient content in moving regions. The patch-based meth- ods utilize patch-level alignment to transfer and fuse simi- lar content. While these patch-based methods achieve better performance, they suffer from high computational costs. With the rise of Deep Neural Networks (DNNs), many works directly learn the complicated mapping between LDR and HDR using a CNN. In general, these models fol- low the paradigm of alignment before fusion. The non-end- to-end DNN-based methods [6, 22] first align LDR images with optical flow or homographies, and then fuse aligned images to generate HDR images. The alignment approaches are error-prone and inevitably cause ghosting artifacts when complex foreground motions occur. Based on the attention- based end-to-end method AHDRNet [25, 26] which per- forms spatial attention to suppress motion and saturation, several methods [2,12,27,30,31] have been proposed to re- move ghosting artifacts. The spatial attention module pro- duces attention maps and element-wise multiplies with non- reference features, thus the model removes motion or satu- rated regions and highlights more informative regions. However, the success of these methods relies on no- ticeable variations between reference and non-reference frames. These methods perform well in the marginal areas, even if there is a misalignment in the input images. Unluck- ily, spatial attention produces unsatisfactory results when motion and saturation are present simultaneously (see Fig- ure 1). The reason can be attributed to that spatial attention uses element-wise multiplication, which only considers the features in the same positions. For example, in the reference frame of Figure 1 ( i.e., LDR with EV=0), the information in the over-exposed regions is unavailable, spatial attention can only rely on the non-saturated information of the same position ( i.e., moving regions) in the non-reference frame due to element-wise multiplication. Therefore, recovering the content of the moving and saturated regions is challeng- ing. Finally, this limitation of spatial attention causes obvi- ous ghosting artifacts in these complex cases. To generate high-quality HDR images in various cases, we propose a Hybrid HDR deghosting Network, named Hy- HDRNet, to establish the complicated alignment and fusion relationship between reference and non-reference images. The proposed HyHDRNet comprises a content alignment subnetwork and a Transformer-based fusion subnetwork. For the content alignment subnetwork, inspired by patch- based HDR imaging methods [5, 19], we propose a novel Patch Aggregation (PA) module, which calculates the sim- ilarity map between different patches and selectively ag- gregates useful information from non-reference LDR im- ages, to remove ghosts and generate content of saturation and misalignment. While the traditional patch-based HDR imaging methods have excellent performance but have thefollowing drawbacks: 1) low patch utilization ratio caused by reusing the same patches, which leads to insufficient content during fusion, 2) structural destruction of images when transfering patches, 3) high computational complex- ity in full resolution. To this end, our Patch Aggregation mechanism 1) aggregates multiple patches which improves the patch utilization ratio 2) aggregates patches instead of exchanging them to maintain structural information, 3) cal- culates a similarity map within a window to reduce compu- tational complexity. These advantages promote the network to remedy the content of saturated and motion regions(See Figure 9), other patch-based HDR imaging methods cannot achieve this goal. In a word, our PA module (patch level) discovers and aggregates similar patches within a large re- ceptive field according to the similarity map, thus it can re- cover the content inside the distorted regions. To further avoid ghosting, we also employ a ghost attention module (pixel level) as a complementary branch for the PA module, and propose a gating module to achieve mutual guidance of these two modules in the content alignment subnetwork. In addition, unlike previous methods using DNN structure in the feature fusion stage which has static weights and only merges the local information, we propose a Transformer- based fusion subnetwork that uses Residual Deformable Transformer Block (RDTB) to model long-range dependen- cies of different regions. The RDTB can dynamically ad- just weights and adaptively merge information in different exposure regions. The experiments demonstrate that our proposed method achieves state-of-the-art performance on public datasets. The main contributions of our work can be summarized as follows: • We propose a hybrid HDR deghosting network to ef- fectively integrate the advantage of patch aggregation and ghost attention using a gating strategy. • We first introduce the patch aggregation module which selectively aggregates useful information from non- reference LDR images to remove ghosts and generate content for saturation. • A novel residual deformable Transformer block is pro- posed, which can adaptively fuse a large range of in- formation to generate high-quality HDR images. • We carry out both qualitative and quantitative experi- ments, which show that our method achieves state-of- the-art results over four public benchmarks.

Zhang_SadTalker_Learning_Realistic_3D_Motion_Coefficients_for_Stylized_Audio-Driven_Single_CVPR_2023

Abstract Generating talking head videos through a face image and a piece of speech audio still contains many challenges. i.e., unnatural head movement, distorted expression, and identity modification. We argue that these issues are mainly caused by learning from the coupled 2D motion fields. On the other hand, explicitly using 3D information also suffers problems of stiff expression and incoherent video. We present SadTalker, which generates 3D motion coefficients (head pose, expression) of the 3DMM from audio and implicitly modulates a novel 3D-aware face render for talking head generation. To learn the realistic motion coefficients, we *Equal Contribution †Corresponding Authorexplicitly model the connections between audio and differ- ent types of motion coefficients individually. Precisely, we present ExpNet to learn the accurate facial expression from audio by distilling both coefficients and 3D-rendered faces. As for the head pose, we design PoseVAE via a conditional VAE to synthesize head motion in different styles. Finally, the generated 3D motion coefficients are mapped to the un- supervised 3D keypoints space of the proposed face render to synthesize the final video. We conducted extensive experi- ments to demonstrate the superiority of our method in terms of motion and video quality.1 1The code and demo videos are available at https://sadtalker. github.io . This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8652

1. Introduction Animating a static portrait image with speech audio is a challenging task and has many important applications in the fields of digital human creation, video conferences, etc. Previous works mainly focus on generating lip mo- tion [2, 3, 28, 29, 49] since it has a strong connection with speech. Recent works also aim to generate a realistic talking face video containing other related motions, e.g., head pose. Their methods mainly introduce 2D motion fields by land- marks [50] and latent warping [37, 38]. However, the quality of the generated videos is still unnatural and restricted by the preference pose [16, 49], month blur [28], identity modi- fication [37, 38], and distorted face [37, 38, 47]. Generating a natural-looking talking head video contains many challenges since the connections between audio and different motions are different. i.e., the lip movement has the strongest connection with audio, but audio can be talked via different head poses and eye blink. Thus, previous facial landmark-based methods [2, 50] and 2D flow-based audio to expression networks [37, 38] may generate the distorted face since the head motion and expression are not fully disentan- gled in their representation. Another popular type of method is the latent-based face animation [3, 16, 28, 49]. Their meth- ods mainly focus on the specific kind of motions in talking face animation and struggle to synthesize high-quality video. Our observation is that the 3D facial model contains a highly decoupled representation and can be used to learn each type of motion individually. Although a similar observation has been discussed in [47], their methods also generate inaccu- rate expressions and unnatural motion sequences. From the above observation, we propose SadTalker, a Stylized Audio- Driven Talk ing-head video generation sys- tem through implicit 3D coefficient modulation. To achieve this goal, we consider the motion coefficients of the 3DMM as the intermediate representation and divide our task into two major components. On the one hand, we aim to generate the realistic motion coefficients ( e.g., head pose, lip motion, and eye blink) from audio and learn each motion individu- ally to reduce the uncertainty. For expression, we design a novel audio to expression coefficient network by distilling the coefficients from the lip motion only coefficients from [28] and the perceptual losses (lip-reading loss [1], facial landmark loss) on the reconstructed rendered 3D face [5]. For the stylized head pose, a conditional V AE [6] is used to model the diversity and life-like head motion by learning the residual of the given pose. After generating the realistic 3DMM coefficients, we drive the source image through a novel 3D-aware face render. Inspired by face-vid2vid [40], we learn a mapping between the explicit 3DMM coefficients and the domain of the unsupervised 3D keypoint. Then, the warping fields are generated through the unsupervised 3D keypoints of source and driving and it warps the reference im- age to generate the final videos. We train each sub-networkof expression generation, head poses generation and face renderer individually and our system can be inferred in an end-to-end style. As for the experiments, several metrics show the advantage of our method in terms of video and motion methods. The main contribution of this paper can be summarized as: •We present SadTalker, a novel system for a stylized audio-driven single image talking face animation using the generated realistic 3D motion coefficients. •To learn the realistic 3D motion coefficient of the 3DMM model from audio, ExpNet and PoseV AE are presented individually. •A novel semantic-disentangled and 3D-aware face ren- der is proposed to produce a realistic talking head video. •Experiments show that our method achieves state-of- the-art performance in terms of motion synchronization and video quality.

Xu_Toward_RAW_Object_Detection_A_New_Benchmark_and_a_New_CVPR_2023

Abstract In many computer vision applications (e.g., robotics and autonomous driving), high dynamic range (HDR) data isnecessary for object detection algorithms to handle a vari-ety of lighting conditions, such as strong glare. In this pa- per , we aim to achieve object detection on RAW sensor data, which naturally saves the HDR information from image sen- sors without extra equipment costs. We build a novel RAWsensor dataset, named ROD, for Deep Neural Networks (DNNs)-based object detection algorithms to be applied toHDR data. The ROD dataset contains a large amount of an-notated instances of day and night driving scenes in 24-bit dynamic range. Based on the dataset, we first investigatethe impact of dynamic range for DNNs-based detectors and demonstrate the importance of dynamic range adjustmentfor detection on RAW sensor data. Then, we propose a sim-ple and effective adjustment method for object detection onHDR RAW sensor data, which is image adaptive and jointly optimized with the downstream detector in an end-to-endscheme. Extensive experiments demonstrate that the per-formance of detection on RAW sensor data is significantly superior to standard dynamic range (SDR) data in different situations. Moreover , we analyze the influence of texture in- formation and pixel distribution of input data on the perfor-mance of the DNNs-based detector . Code and dataset will be available at https://gitee.com//mindspore/ models/tree/master/research/cv/RAOD .

1. Introduction Real-world scenes are complex and of high dynamic range (HDR), especially in extreme situations like the directlight of other vehicles. In many computer vision applica- tions, such as autonomous driving and robotics, HDR data is important and necessary for making safety-critical deci-sions [ 34] since it extends the captured luminance. For in- stance, images may easily get over-exposed in brighter areas *Equal contribution.†Corresponding author. This work was done when Ruikang Xu was a research intern at Huawei Noah’s Ark Lab.on standard dynamic range (SDR) images, but there may be important information in corresponding raw regions. To ob- tain the HDR data, recent works use additional cameras andeven unconventional sensors, such as neuromorphic cam- eras and infrared cameras [ 13,25], which inevitably brings extra costs. In this paper, we make the first effort to achieve object detection on RAW sensor data, which naturally stores HDR information without any additional burden. RAW sensor data is generated from the image sensor, and is the input data of the image signal processor (ISP), rendering SDR images suitable for human perception andunderstanding. RAW sensor data is naturally HDR and saveall information from image sensors. However, datasets from the RAW domain are difficult to collect, store and annotate. And, to the best of our knowledge, there is no large-scaleHDR RAW sensor dataset available for object detection.Existing RAW sensor datasets are no more than 14-bit andnot large enough for practical applications. For instance,PASCALRAW dataset [ 27] is of only 12-bit, which is not wide enough to handle complex lighting conditions. To fill this gap, we create a novel RAW sensor dataset for objectdetection on the driving scene, named as ROD, which con-sists of 25k annotated RAW sensor data in a 24-bit dynamicrange on day and night scenarios. On the other hand, most Deep Neural Networks (DNNs)- based object detection algorithms are designed for the com- mon SDR data, which only records information in the 8-bitdynamic range. Hence, we first investigate the impact of dy-namic range for DNNs-based detection algorithms and ex- perimentally find that directly applying these DNNs-baseddetectors on HDR RAW sensor data results in significant performance degradation, and it gets worse when the dy-namic range increases. Then, we analyze the key compo-nent of the ISP system and demonstrate the importance ofdynamic range adjustment for RAW detection. In this paper, we propose an adjustment method for the effective detection on RAW sensor data, which is jointly op-timized with the downstream detection network in an end- to-end scheme. Note that our proposed method is trained together with the detector from scratch only using object an- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13384 notations as the supervision. To effectively exploit the HDR information from RAW sensor data, we devise an image- adaptive processing network to regulate RAW sensor datawith learnable transformation functions. Specifically, we design two modules to adjust the dynamic range of RAW sensor data by image-level and pixel-level information. In addition, our proposed method is lightweight and computa-tionally efficient. Extensive experimental results on the proposed ROD dataset demonstrate that the performance of object detec-tion on RAW sensor data is significantly superior to de-tection on SDR data in different scenarios. Our methodalso outperforms recent state-of-the-art neural ISP meth-ods [ 23,41]. Comprehensive ablation experiments show that our proposed method effectively improves the perfor-mance of DNNs-based object detection algorithms on HDR RAW sensor data. Furthermore, we analyze the influence of texture information and pixel distribution of the input datafor the performance of the downstream detection network. In summary, the main contributions are as follows:•We build a novel RAW sensor dataset for object detec- tion on HDR RAW sensor data, which contains 25k drivingscenes on both day and night scenarios. •We propose a simple and effective adjustment method for detection on HDR RAW sensor data, which is jointlyoptimized with the detector in an end-to-end manner. •Extensive experiments demonstrate that object detec- tion on HDR RAW sensor data significantly outperformsthat on SDR data in different situations. It also shows that our method is effective and computationally efficient.

Yu_Mind_the_Label_Shift_of_Augmentation-Based_Graph_OOD_Generalization_CVPR_2023

Abstract Out-of-distribution (OOD) generalization is an impor- tant issue for Graph Neural Networks (GNNs). Recent works employ different graph editions to generate aug- mented environments and learn an invariant GNN for gen- eralization. However, the label shift usually occurs in aug- mentation since graph structural edition inevitably alters the graph label. This brings inconsistent predictive rela- tionships among augmented environments, which is harm- ful to generalization. To address this issue, we propose LiSA , which generates label-invariant augmentations to fa- cilitate graph OOD generalization. Instead of resorting to graph editions, LiSA exploits Label-invariant Subgraphs of the training graphs to construct Augmented environments. Specifically, LiSA first designs the variational subgraph generators to extract locally predictive patterns and con- struct multiple label-invariant subgraphs efficiently. Then, the subgraphs produced by different generators are col- lected to build different augmented environments. To pro- mote diversity among augmented environments, LiSA fur- ther introduces a tractable energy-based regularization to enlarge pair-wise distances between the distributions of en- vironments. In this manner, LiSA generates diverse aug- mented environments with a consistent predictive relation- ship and facilitates learning an invariant GNN. Extensive experiments on node-level and graph-level OOD bench- marks show that LiSA achieves impressive generalization performance with different GNN backbones. Code is avail- able on https://github.com/Samyu0304/LiSA .

1. Introduction Learning from graph-structured data is a fundamental problem in various applications, such as 3D vision [50], *Corresponding Authorknowledge graph reasoning [65], and social network analy- sis [32]. Recently, the Graph Neural Networks (GNNs) [33] have become a de facto standard in developing deep learn- ing systems on graphs [10], showing superior performance on point cloud classification [18], recommendation system [56], biochemistry [29] and so on. Despite their remarkable success, these models heavily rely on the i.i.d. assumption that the training and testing data are independently drawn from an identical distribution [9, 39]. When tested on out- of-distribution (OOD) graphs ( i.e. larger graphs), GNN usu- ally suffers from unsatisfactory performances and unstable prediction results. Hence, handling the distribution shift for GNNs has received increasing attention. Many solutions have been proposed to explore the OOD generalization problem in Euclidean space [47], such as invariant learning [3, 14, 38], group fairness [34], and distribution-robust optimization [49]. Recent works mainly resort to learning an invariant classifier that per- forms equally well in different training environments [3, 16, 37, 38]. However, the study of its counterpart problem for non-Euclidean graphs is comparatively lacking. One challenge is the environmental scarcity of graph-structured data [39, 53]. Inspired by the data augmentation literature [48, 52], some pioneering works propose to generate aug- mented training environments by applying different graph edition policies to the training graphs [55,57]. After training in these environments, the GNN is expected to have better OOD generalization ability. Nevertheless, the graph labels may change during the graph edition since they are sensi- tive to graph structural modifications. This causes the label shift problem of augmented graphs. For example, methods of graph adversarial attack usually seek to modify the graph structure to permute the model prediction [9]. Moreover, a small structural modification can drastically influence the biochemical property of molecule or protein graphs [30]. We formalize the impact of the label shift in augmen- tations on generalization using a unified structure equa- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11620 tion model [1]. Our analysis indicates that the label shift causes inconsistent predictive relationships among the aug- mented environments. This misguides the GNN to out- put a perturbed prediction rather than the invariant pre- diction, making the learned GNN hard to generalize (see Section 3 for more details). Thus, it is crucial to gener- ate label-invariant augmentations for graph OOD general- ization. However, designing label-invariant graph edition is nontrivial or even brings extensive computation, since it requires learning class-conditional distribution for discrete and irregular graphs. In this work, we propose a novel label- invariant subgraph augmentation method, dubbed LiSA , for the graph OOD generalization problem. For an input graph, LiSA first designs the variational subgraph generators to identify locally predictive patterns ( i.e. important nodes or edges for the graph label) and generate multiple label- invariant subgraphs. These subgraphs capture prediction- relevant information with different structures, and thus con- struct augmented environments with a consistent predictive relationship. To promote diversity among the augmenta- tions, we propose a tractable energy-based regularization to enlarge the pair-wise distances between the distributions of augmented environments. With the augmentations pro- duced by LiSA, a GNN classifier is learned to be invariant across these augmented environments. The GNN predictor and variational subgraph generators are jointly optimized with a bi-level optimization scheme [61]. LiSA is model- agnostic and is flexible in handling both graph-level and node-level distribution shifts. Extensive experiments indi- cate that LiSA enjoys satisfactory performance gain over the baselines on 7 graph classification datasets and 4 node classification datasets. Our contributions are as follows: • We propose a model-agnostic label-invariant subgraph augmentation (LiSA) framework to generate aug- mented environments with consistent predictive rela- tionships for graph OOD generalization. • We propose the variational subgraph generator to dis- cover locally crucial patterns to construct the label- invariant subgraphs efficiently. • To further promote diversity, we further propose an energy-based regularization to enlarge pair-wise dis- tances between the distributions of different aug- mented environments. • Extensive experiments on node-level and graph-level tasks indicate that LiSA enjoys satisfactory perfor- mance gain over the baselines on various backbones.

Zhang_Quantum-Inspired_Spectral-Spatial_Pyramid_Network_for_Hyperspectral_Image_Classification_CVPR_2023

Abstract Hyperspectral image (HSI) classification aims at assign- ing a unique label for every pixel to identify categories of different land covers. Existing deep learning models for HSIs are usually performed in a traditional learning paradigm. Being emerging machines, quantum computers are limited in the noisy intermediate-scale quantum (NISQ) era. The quantum theory offers a new paradigm for de- signing deep learning models. Motivated by the quan- tum circuit (QC) model, we propose a quantum-inspired spectral-spatial network (QSSN) for HSI feature extraction. The proposed QSSN consists of a phase-prediction module (PPM) and a measurement-like fusion module (MFM) in- spired from quantum theory to dynamically fuse spectral and spatial information. Specifically, QSSN uses a quantum representation to represent an HSI cuboid and extracts joint spectral-spatial features using MFM. An HSI cuboid and its phases predicted by PPM are used in the quantum represen- tation. Using QSSN as the building block, we further pro- pose an end-to-end quantum-inspired spectral-spatial pyra- mid network (QSSPN) for HSI feature extraction and classi- fication. In this pyramid framework, QSSPN progressively learns feature representations by cascading QSSN blocks and performs classification with a softmax classifier. It is the first attempt to introduce quantum theory in HSI pro- cessing model design. Substantial experiments are con- ducted on three HSI datasets to verify the superiority of the proposed QSSPN framework over the state-of-the-art meth- ods.

1. Introduction Hyperspectral images (HSIs) are always represented by three-order tensors that collected by remote sensors to record the characteristics reflected by land covers. Each HSI contains two spatial dimensions and one spectral di- *Corresponding authors. FLOPs/M7580859095AA (%) QSSPN HiT RvT SSAN SSRN SSFTTFigure 1. Performance comparison between the proposed QSSPN framework and existing methods. Average accuracies on Indian Pines dataset are presented. mension to reflect abundant spectral information and spatial context, making it different from color images. HSI classifi- cation aims to determine land-cover categories of areas rep- resented by every pixel according to rich spatial and spectral information [21]. It has a broad range of applications, in- cluding target detection [42], mining [1] and agriculture [8]. Based on the recorded HSIs, samples in different categories may be extremely imbalanced in some data sets. This makes HSI classification very challenging [20, 49]. In recent decades, a large number of HSI classifica- tion models have been proposed. Existing models can be divided into traditional learning methods and deep learn- ing methods. Traditional classification models are simply adopted for HSIs, such as support vector machine (SVM) [27] and K-nearest neighbours (KNN) [19]. To achieve bet- ter results, advanced traditional learning models, e.g., ex- tended morphological profile (EMP) [3] and extended mul- tiattribute profile (EMAP) [7], first conduct feature extrac- tion and then perform classification. These methods are two-step learning models that cannot achieve satisfactory performance. Deep learning models are promising to si- multaneously conduct feature extraction and classification This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9925 in an end-to-end manner [12, 34]. Benefiting from the de- veloped technique, deep learning models for HSI classi- fication are in blossom. A wide variety of models rely on convolutional neural networks (CNNs). Most existing models are designed to separately extract spatial and spec- tral features in different branches [18, 47]. To enrich the development of deep learning models, attention mechanic [26, 46], graph convolution network [45] and transformer learning [36, 43, 48] have been explored in recent works. These works make significant improvement in architecture design and classification performance. However, they sepa- rately aggregate spatial and spectral information in different modules, resulting in model redundancy and inefficient ex- ploration of the correlation between spatial and spectral in- formation. Besides, they are all developed from traditional learning paradigm. Different from traditional computers, quantum comput- ers are emerging machines to perform quantum algorithms [2]. Quantum computing utilizes quantum theory to process data in quantum device [14, 28]. There are several models of quantum computation, such as quantum circuits, quan- tum annealing and adiabatic quantum computation. It has been proved that quantum computers outperform classical computers in solving certain problems [11, 33]. For ex- ample, Shor’s algorithm can be used to solve the integer factorization problem much faster than algorithms running on the classical computer [33]. In the noisy intermediate- scale quantum (NISQ) era [31], quantum computers cannot perform complex quantum algorithms for practical appli- cations. Fortunately, quantum computation provides a new mathematical formalism for computing. As a new learn- ing paradigm, quantum machine learning (QML) adopts quantum computation to enhance classical machine learn- ing models [4, 32]. Recently, quantum theory has been adopted in classical algorithms and deep learning models, with an expectation of improving computation efficiency and outcome quality [9, 38]. Motivated by quantum theory, in this paper, we first introduce a quantum-inspired spectral-spatial net- work (QSSN) for HSI feature extraction. The proposed QSSN includes a phase-prediction module (PPM) and a measurement-like fusion module (MFM). Instead of fus- ing spatial and spectral information in different modules, we extract joint spatial-spectral features in the same op- eration. A small data cuboid taken from an HSI is rep- resented by quantum-inspired state representation, and the amplitudes in a state vector are values in normalized input cuboid. The corresponding phases are predicted by PPM, dynamically modulating the spatial and spectral correla- tions. MFM simultaneously aggregates spatial and spectral information to generate a feature cuboid. Additionally, we design a quantum-inspired spectral-spatial pyramid network (QSSPN) framework with multiple cascaded QSSN blocksand a softmax classifier for HSI classification. Deep pyra- mid structure of QSSPN gradually decreases the channels of feature cuboids and extracts robust and expressive fea- tures from original data for classification. This is a simple but efficient end-to-end framework taking advantage from both quantum theory and deep neural network. The contributions of this article are summarized as fol- lows. • As far as we know, QSSN and QSSPN are the first proposed quantum-inspired method for HSI feature ex- traction and classification. It shows that this quantum- inspired framework is promising. • Motivated by QC in quantum theory, we propose a QSSN for HSI feature extraction. QSSN aggre- gates spatial-spectral information simultaneously us- ing quantum-like representations and operations. • Based on QSSN, we develop a QSSPN for HSI classi- fication. In QSSPN, QSSN blocks stacked in a pyra- mid manner that can extract robust and discrimina- tive features for classification. As shown in Fig. 1, QSSPN achieves the best classification accuracy with low model complexity. The rest of this paper is organized as follows: Sec. 2 reviews the related work about deep learning-based HSI classification and quantum-inspired deep learning methods. Sec. 3 introduces the detailed framework of QSSN and de- scribes the proposed QSSPN for HSI classification. Sec. 4 presents the experimental results and analysis. Finally, Sec. 5 draws the conclusion.

Xu_Seeing_Electric_Network_Frequency_From_Events_CVPR_2023

Abstract Most of the artificial lights fluctuate in response to the grid’s alternating current and exhibit subtle variations in terms of both intensity and spectrum, providing the poten- tial to estimate the Electric Network Frequency (ENF) from conventional frame-based videos. Nevertheless, the perfor- mance of Video-based ENF (V-ENF) estimation largely re- lies on the imaging quality and thus may suffer from signif- icant interference caused by non-ideal sampling, motion, and extreme lighting conditions. In this paper, we show that the ENF can be extracted without the above limita- tions from a new modality provided by the so-called event camera, a neuromorphic sensor that encodes the light in- tensity variations and asynchronously emits events with ex- tremely high temporal resolution and high dynamic range. Specifically, we first formulate and validate the physical mechanism for the ENF captured in events, and then pro- pose a simple yet robust Event-based ENF (E-ENF) estima- tion method through mode filtering and harmonic enhance- ment. Furthermore, we build an Event-Video ENF Dataset (EV-ENFD) that records both events and videos in diverse scenes. Extensive experiments on EV-ENFD demonstrate that our proposed E-ENF method can extract more accu- rate ENF traces, outperforming the conventional V-ENF by a large margin, especially in challenging environments with object motions and extreme lighting conditions. The code and dataset are available at https://github.com/ xlx-creater/E-ENF .

1. Introduction Light flicker is a nuisance in video [33] but can be ex- ploited to detect and estimate the Electric Network Fre- *Equal contribution. †Corresponding authors. The research was partially supported by the National Natural Sci- ence Foundation of China under Grants 62271354 and the Natural Sci- ence Foundation of Hubei Province, China under Grants 2022CFB084 and 2021CFB467. Event Stream Event Camera Estimated ENFStatic Scenes Dynamic Scenes Extreme Lighting(a) Processing Pipeline for E-ENF Extraction (b) Video Sequences (c) Event Streams (d) ENF EstimatesFigure 1. Illustrations of (a) the processing pipeline for E-ENF extraction and comparative experimental results under (b) different recording environments, in terms of (c) recorded event streams and (d) extracted V-ENF and E-ENF. quency (ENF) [10, 11], which is the power line transmis- sion frequency in a grid, unlocking the potential of mul- timedia forensic verification [7, 9], electrical load monitor- ing [23,24], recording device identification [1,28], etc. This is enabled by the discovery that the ENF fluctuates slightly and randomly around the nominal value ( 50or60Hz) and consistently across the entire intra-grid nodes [13]. Existing Video-based ENF (V-ENF) extraction methods attempt to restore the illumination fluctuations by averag- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18022 ing the pixel intensities of every frame over the video se- quences [11]. Based on this, V-ENF estimates can be im- proved by using a higher sample rate in the rolling shut- ter [26,28] or an advanced frequency estimator [3,6,14,19]. However, due to the inherent characteristics of conventional cameras, existing V-ENF extraction methods still face a se- ries of challenges, which are summarized as follows. •Non-ideal sampling. For global shutter videos, the sample rate equals to the camera’s frame rate, usu- ally at 25or30fps, which is smaller than the Nyquist rate to record light flicker, resulting in frequency alias- ing for ENF estimation [10]. Although rolling shutter can increase the sample rate through the line exposure mechanism, it also introduces the inter-frame idle time within which no signal can be sampled [28,31], violat- ing the uniform sampling condition. •Motion. Motion in the video leads to the abrupt change of pixel intensity, which is inconsistent with the flicker pattern and produces strong interference. Note that motion can be caused not only by object move- ment ( e.g., walking person) but also by camera move- ment ( e.g., hand-held camera shake). •Extreme lighting conditions. Imaging quality of over- or under-exposed videos (even under good light- ing conditions) can be substantially degraded, in which the flicker information can be totally lost [23]. Mean- while, under extreme low-light conditions ( e.g., night scene with insufficient lighting), one has to boost the ISO for appropriate exposure, inevitably bringing in ISO noises, which can severely deteriorate the ENF. Generally, the actual content in video recordings can have diverse scenes and objects, which become interference with the task of ENF extraction from the content. Note that putting a video camera still against a white wall illu- minated by a light source is an ideal condition for V-ENF extraction [10, 26]. However, to the very opposite, real- world recordings hardly satisfy this condition. To date, re- searchers and practitioners are still working intensively to tackle the above problems [14, 27, 29]. In this paper, we show that the above-mentioned prob- lems can be effectively solved via the proposed use of the so-called event camera, a new modality for recording. The event camera is a neuromorphic sensor that encodes the light intensity variations and asynchronously emits events with extremely high temporal resolution and high dynamic range [8, 21]. Different from V-ENF, the proposed Event-based ENF (E-ENF) extraction approach collects the illumination changes and converts them into an event stream, based on which the ENF traces can be eventually estimated, as shown in Fig. 1(a). Thanks to the high temporal resolution andhigh dynamic range of the event camera, E-ENF can pro- vide a sufficient sample rate and extract reliable ENF traces even under harsh conditions, e.g., motion interference and extreme lighting, as shown in Fig. 1(b-d). Since no prior work has focused on extracting ENF from events, we hereby attempt to provide the first proof-of- concept study theoretically and experimentally demonstrat- ing why the ENF can exist in event streams and how it can be reliably extracted therein, in comparison with the con- ventional V-ENF approach. The main contributions of this paper are three-fold: • Based on the event-sensing mechanism, we formulate the process of the ENF fluctuation, reflected by light flicker, being recorded and converted into an event stream, validating the ENF capture in events. • We propose the first method that effectively extracts the ENF traces from events, featuring a uniform- interval temporal sampling algorithm, a majority- voting spatial sampling algorithm, and a harmonic se- lection algorithm, respectively. • We construct and open-source the first Event-Video hybrid ENF Dataset (EV-ENFD), containing both events and videos recorded in real-world environments with motion and extreme lighting conditions. The code and dataset are available at https://github. com/xlx-creater/E-ENF .

Yoo_Towards_End-to-End_Generative_Modeling_of_Long_Videos_With_Memory-Efficient_Bidirectional_CVPR_2023

Abstract Autoregressive transformers have shown remarkable success in video generation. However, the transformers are prohibited from directly learning the long-term dependency in videos due to the quadratic complexity of self-attention, and inherently suffering from slow inference time and er- ror propagation due to the autoregressive process. In this paper, we propose Memory-efficient Bidirectional Trans- former (MeBT) for end-to-end learning of long-term depen- dency in videos and fast inference. Based on recent ad- vances in bidirectional transformers, our method learns to decode the entire spatio-temporal volume of a video in par- allel from partially observed patches. The proposed trans- former achieves a linear time complexity in both encoding and decoding, by projecting observable context tokens into a fixed number of latent tokens and conditioning them to decode the masked tokens through the cross-attention. Em- powered by linear complexity and bidirectional modeling, our method demonstrates significant improvement over the autoregressive transformers for generating moderately long videos in both quality and speed. Videos and code are avail- able at https://sites.google.com/view/mebt-cvpr2023.

1. Introduction Modeling the generative process of videos is an impor- tant yet challenging problem. Compared to images, gener- ating convincing videos requires not only producing high- quality frames but also maintaining their semantic struc- tures and dynamics coherently over long timescale [11, 16, 27, 36, 38, 54]. Recently, autoregressive transformers on discrete repre- sentation of videos have shown promising generative mod- eling performances [11,32,51,53]. Such methods generally involve two stages, where the video frames are first turned into discrete tokens through vector quantization, and then their sequential dynamics are modeled by an autoregressive transformer. Powered by the flexibility of discrete distri-butions and the expressiveness of transformer architecture, these methods demonstrate impressive results in learning and synthesizing high-fidelity videos. However, autoregressive transformers for videos suffer from critical scaling issues in both training and inference. During training, due to the quadratic cost of self-attention, the transformers are forced to learn the joint distribution of frames entirely from short videos ( e.g., 16 frames [11, 53]) and cannot directly learn the statistical dependencies of frames over long timescales. During inference, the mod- els are challenged by two issues of autoregressive gener- ative process – its serial process significantly slows down the inference speed, and perhaps more importantly, autore- gressive prediction is prone to error propagation where the prediction error of the frames accumulates over time. To (partially) address the issues, prior works proposed improved transformers for generative modeling of videos, which are categorized as the following: (a)Employing sparse attention to improve scaling during training [12, 16, 51],(b)Hierarchical approaches that employ separate mod- els in different frame rates to generate long videos with a smaller computation budget [11, 16], and (c)Remov- ing autoregression by formulating the generative process as masked token prediction and training a bidirectional trans- former [12, 13]. While each approach is effective in ad- dressing specific limitations in autoregressive transformers, none of them provides comprehensive solutions to afore- mentioned problems – (a, b) still inherits the problems in autoregressive inference and cannot leverage the long-term dependency by design due to the local attention window, and(c)is not appropriate to learn long-range dependency due to the quadratic computation cost. We believe that an alternative that jointly resolves all the issues would provide a promising approach towards powerful and efficient video generative modeling with transformers. In this work, we propose an efficient transformer for video synthesis that can fully leverage the long-range de- pendency of video frames during training, while being able to achieve fast generation and robustness to error propaga- tion. We achieve the former with a linear complexity ar- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22888 chitecture that still imposes dense dependencies across all timesteps, and the latter by removing autoregressive seri- alization through masked generation with a bidirectional transformer. While conceptually simple, we show that ef- ficient dense architecture and masked generation are highly complementary, and when combined together, lead to sub- stantial improvements in modeling longer videos compared to previous works both in training and inference. The con- tributions of this paper are as follows: • We propose Memory-efficient Bidirectional Trans- former (MeBT) for generative modeling of video. Un- like prior methods, MeBT can directly learn long- range dependency from training videos while enjoying fast inference and robustness in error propagation. • To train MeBT for moderately long videos, we propose a simple yet effective curriculum learning that guides the model to learn short- to long-term dependencies gradually over training. • We evaluate MeBT on three challenging real-world video datasets. MeBT achieves a performance compet- itive to state-of-the-arts for short videos of 16 frames, and outperforms all for long videos of 128 frames while being considerably more efficient in memory and computation during training and inference.

Zhang_Accelerating_Dataset_Distillation_via_Model_Augmentation_CVPR_2023

Abstract Dataset Distillation (DD), a newly emerging field, aims at generating much smaller but efficient synthetic training datasets from large ones. Existing DD methods based on gradient matching achieve leading performance; however, they are extremely computationally intensive as they re- quire continuously optimizing a dataset among thousands of randomly initialized models. In this paper, we assume that training the synthetic data with diverse models leads to better generalization performance. Thus we propose twomodel augmentation techniques, i.e. using early-stage models andparameter perturbation to learn an informative synthetic set with significantly reduced training cost. Exten- sive experiments demonstrate that our method achieves up to 20×speedup and comparable performance on par with state-of-the-art methods.

1. Introduction Dataset Distillation (DD) [3, 48] or Dataset Condensa- tion [55, 56], aims to reduce the training cost by generat- ing a small but informative synthetic set of training exam- ples; such that the performance of a model trained on the small synthetic set is similar to that trained on the origi- nal, large-scale dataset. Recently, DD has become an in- creasingly more popular research topic, and has been ex- plored in a variety of contexts, including federated learn- ing [17, 42], continual learning [33, 40], neural architecture search [43, 57], medical computing [25, 26] and graph neu- ral networks [21, 30]. DD has been typically cast as a meta-learning prob- lem [16] involving bilevel optimization. For instance, Wang et al . [48] formulate the network parameters as a function of the learnable synthetic set in the inner-loop *Equal contribution. †Corresponding author: Dongkuan Xu. 20x 1x 0.75x1x 5x 10x 0.0 0.4 0.8 1.2 GPU Hours for Dataset Condensation (hours in log10 scale)40506070Test Accuracy (%)DMOurs5 Ours10Ours20TMIDC DSA DCFigure 1. Performances of condensed datasets for training ConvNet-3 v.s. GPU hours to learn the 10 images per class con- densed CIFAR-10 datasets with a single RTX-2080 GPU. Ours 5, Ours 10, and Ours 20accelerates the training speed of the state-of- the-art method IDC [22] 5×,10×, and 20×faster. optimization; then optimize the synthetic set by minimiz- ing classification loss on the real data in the outer-loop. This recursive computation hinders its application to real- world large-scale model training, which involves thousands to millions of gradient descent steps. Several methods have been proposed to improve the DD method by introducing ridge regression loss [2, 36], trajectory matching loss [3], etc. To avoid unrolling the recursive computation graph, Zhao et al. [57] propose to learn synthetic set by matching gradients generated by real and synthetic data when training deep networks. Based on this surrogate goal, several meth- ods have been proposed to improve the informativeness or compatibility of synthetic datasets from other perspectives, ranging from data augmentation [55], contrastive signal- ing [24], resolution reduction [22], and bit encoding [41]. Although model training on a small synthetic set is fast, the dataset distillation process is typically expensive. For instance, the state-of-the-art method IDC [22] takes ap- proximately 30 hours to condense 50,000 CIFAR-10 im- 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11950 ages into 500 synthetic images with a single RTX-2080 GPU, which is equivalent to the time it takes to train 60 ConvNet-3 models on the original dataset. Furthermore, the distillation time cost will rapidly increase for large-scale datasets e.g. ImageNet-1K, which prevents its application in computation-limited environments like end-user devices. Prior work [56] on reducing the distillation cost results in significant regression from the state-of-the-art performance. In this paper, we aim to speed up the dataset distillation process, while preserving even improving the testing per- formance over state-of-the-art methods. Prior works are computationally expensive as they fo- cus on generalization ability such that the learned synthetic set is useful to train many different networks as opposed to a targeted network. This requires optimizing the syn- thetic set over thousands of differently initialized networks. For example, IDC [22] learns the synthetic set over 2000 randomly initialized models, while the trajectory matching method (TM) [3] optimizes the synthetic set for 10000 dis- tillation steps with 200 pre-trained expert models. Dataset distillation, which learns the synthetic data that is generaliz- able to unseen models, can be considered as an orthogonal approach to model training which learns model parameters that are generalizable to unseen data. Similarly, training the synthetic data with diverse models leads to better general- ization performance. This intuitive idea leads to the follow- ing research questions: Question 1. How to design the candidate pool of models to learn synthetic data, for instance, consisting of randomly initialized, early-stage or well-trained models? Prior works [3, 22, 48, 57] use models from all training stages. The underlying assumption is that models from all training stages have similar importance. Zhao et al . [56] show that synthetic sets with similar generalization perfor- mance can be learned with different model parameter dis- tributions, given an objective function in the form of feature distribution matching between real and synthetic data. In this paper, we take a closer look at this problem and show that learning synthetic data on early-stage models is more efficient for gradient/parameter matching based dataset dis- tillation methods. Question 2. Can we learn a good synthetic set using only a few models? Our goal is to learn a synthetic set with a small number of (pre-trained) models to minimize the computational cost. However, using fewer models leads to poor generalization ability of the synthetic set. Therefore, we propose to apply parameter perturbation on selected early-stage models to incorporate model diversity and improve the generalization ability of the learned synthetic set. In a nutshell, we propose two model augmentation techniques to accelerate the training speed of dataset dis- tillation, namely using early-stage models andparameterperturbation to learn an informative synthetic set with sig- nificantly less training cost. As illustrated in Fig. 1., our method achieves up to 20 ×speedup and comparable per- formance on par with state-of-the-art DD methods.

Zhang_Learning_Emotion_Representations_From_Verbal_and_Nonverbal_Communication_CVPR_2023

Abstract Emotion understanding is an essential but highly chal- lenging component of artificial general intelligence. The absence of extensive annotated datasets has significantly impeded advancements in this field. We present Emotion- CLIP , the first pre-training paradigm to extract visual emo- tion representations from verbal and nonverbal communi- cation using only uncurated data. Compared to numerical labels or descriptions used in previous methods, commu- nication naturally contains emotion information. Further- more, acquiring emotion representations from communica- tion is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emo- tion cues using sentiment-guided contrastive learning. Ex- tensive experiments validates the effectiveness and transfer- ability of EmotionCLIP . Using merely linear-probe evalua- tion protocol, EmotionCLIP outperforms the state-of-the- art supervised visual emotion recognition methods and ri- vals many multimodal approaches across various bench- marks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related do- mains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.

1. Introduction If artificial intelligence (AI) can be equipped with emo- tional intelligence (EQ), it will be a significant step toward developing the next generation of artificial general intelli- gence [46, 92]. The combination of emotion and intelli- gence distinguishes humans from other animals. The ability to understand, use, and express emotions will significantly facilitate the interaction of AI with humans and the environ- ment [20, 48–50], making it the foundation for a wide vari- ety of HCI [3], robotics [11], and autonomous driving [31] applications. *equal contribution Conversation: - Dad. Who is Dede? - Jesus. She was the love of my life and I was too stupid to realized it. I lost her because of something so dumb.0 1 1 0 0Categorical Label: Shocked Regretful Description: An old man is chatting with his son while eating at a restaurant. Figure 1. Emotions emerge naturally in human communica- tion through verbal and nonverbal cues. The rich semantic de- tails within the expression can hardly be represented by human- annotated categorical labels and descriptions in current datasets. Artificial emotional intelligence (AEI) research is still in its nascency [30, 73]. The recent emergence of pre- trained models in CV [10, 24, 62] and NLP [7, 16, 33, 68] domains has ushered in a new era of research in related subjects. By training on large-scale unlabeled data in a self- supervised manner, the model learns nontrivial representa- tions that generalize to downstream tasks [42,62]. Unfortu- nately, such a technique remains absent from AEI research. The conventional approaches in visual emotion understand- ing have no choice but to train models from scratch, or leverage models from less-relevant domains [27, 66], suf- fering from data scarcity [29, 45]. The lack of pre-trained models greatly limits the development of AEI research. Research in neuroscience and psychology offers insights for addressing this problem. Extending from the capabil- ities that have been coded genetically, humans learn emo- tional expressions through daily interaction and communi- cation as early as when they are infants. It has been shown that both vision [58] and language [41] play crucial roles in this learning process. By absorbing and imitating expres- sions from others, humans eventually master the necessary This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18993 feelings to comprehend emotional states by observing and analyzing facial expressions, body movements, contextual environments, etc. Inspired by how humans comprehend emotions, we pro- pose a new paradigm for emotion understanding that learn directly from human communication. The core of our idea is to explore the consistency between verbal and nonverbal affective cues in daily communication. Fig. 1 shows how communication reveals emotion. Our method that learns from communication is not only aligned with the human learning process but also has several advantages: 1)Our method bypasses the problems in emotion data collection by leveraging uncurated data from daily com- munication. Existing emotion understanding datasets are mainly annotated using crowdsourcing [29, 45, 77]. For image classification tasks, it is straightforward for annota- tors to agree on an image’s label due to the fact that the label is determined by certain low-level visual character- istics. However, crowdsourcing participants usually have lower consensus on producing emotion annotations due to the subjectivity and subtlety of affective labels [90]. This phenomenon makes it extremely difficult to collect accurate emotion annotations on a large scale. Our approach does not rely on human annotations, allowing us to benefit from nearly unlimited web data. 2)Our use of verbal expressions preserves fine-grained semantics to the greatest extent possible. Limited by the data collection strategy, existing datasets usually only con- tain annotations for a limited number of emotion categories, which is far from covering the space of human emotional expression [86]. Moreover, the categorical labels com- monly used in existing datasets fail to precisely represent the magnitude or intensity of a certain emotion. 3)Our approach provides a way to directly model ex- pressed emotion. Ideally, AEI should identify the individ- ual’s emotional state, i.e., the emotion the person desires to express. Unfortunately, it is nearly impossible to col- lect data on this type of “expressed emotion” on a large scale. Instead, the current practice is to collect data on “per- ceived emotion” to approximate the person’s actual emo- tional state, which inevitably introduces noise and bias to labels. In general, learning directly from how humans express themselves is a promising alternative that gives a far broader source of supervision and a more comprehensive represen- tation. This strategy is closely analogous to the human learning process and provides an efficient solution for ex- tracting emotion representations from uncurated data. We summarize our main contributions as follows: • We introduce EmotionCLIP, the first vision-language pre-training paradigm using uncurated data to the vi- sual emotion understanding domain. • We propose two techniques to guide the model to cap-ture salient emotional expressions from human verbal and nonverbal communication. • Extensive experiments and analysis demonstrate the superiority and transferability of our method on vari- ous downstream datasets in emotion understanding.

Yuan_You_Are_Catching_My_Attention_Are_Vision_Transformers_Bad_Learners_CVPR_2023

Abstract Vision Transformers (ViTs), which made a splash in the field of computer vision (CV), have shaken the dominance of convolutional neural networks (CNNs). However, in the process of industrializing ViTs, backdoor attacks have brought severe challenges to security. The success of ViTs benefits from the self-attention mechanism. However, com- pared with CNNs, we find that this mechanism of capturing global information within patches makes ViTs more sensi- tive to patch-wise triggers. Under such observations, we delicately design a novel backdoor attack framework for ViTs, dubbed BadViT, which utilizes a universal patch-wise trigger to catch the model’s attention from patches bene- ficial for classification to those with triggers, thereby ma- nipulating the mechanism on which ViTs survive to confuse itself. Furthermore, we propose invisible variants of BadViT to increase the stealth of the attack by limiting the strength of the trigger perturbation. Through a large number of ex- periments, it is proved that BadViT is an efficient backdoor attack method against ViTs, which is less dependent on the number of poisons, with satisfactory convergence, and is transferable for downstream tasks. Furthermore, the risks inside of ViTs to backdoor attacks are also explored from the perspective of existing advanced defense schemes.

1. Introduction Transformers, which are all-powerful in the field of nat- ural language processing (NLP) [6, 13, 57], have recently set off a wave of frenetic research in computer vision (CV). Thanks to the self-attention mechanism, vision transform- ers (ViTs) have broken the perennial domination of convo- lutional neural networks (CNNs) [17], and have been de- *Corresponding author.veloped in hot areas like image classification [36, 53, 66], object detection [3, 7] and semantic segmentation [46, 64]. The architecture optimization researches of ViTs are also continuously improving the performance and efficiency [26, 40, 45, 53], and providing vitality for advancing the de- ployment of ViTs in industry. Unfortunately, manifold threats in deep learning pose se- vere challenges to ViTs. For instance, adversarial attacks [10, 11, 25, 28, 31, 32, 49, 50, 76] confuse the deep model to make wrong predictions by adding subtle perturbations to the input. In addition, backdoor attacks are also extremely threatening to deep models [22, 27, 48, 51, 68]. More and more deep learning tasks are “outsourced” training or di- rectly fine-tuning on pre-trained models [22, 75], allowing attackers to implant backdoors into the model by establish- ing a strong association between the trigger and the attack behavior. In response, a growing number of researchers have paid attention to the security of ViTs under adversarial attacks [1,4,20,39,44] and backdoor attacks [16,37,47,73]. However, previous ViT backdoor works have not system- atically compared with CNN to elucidate the vulnerability source of ViTs to backdoor attacks, and have not consid- ered balancing attack concealment and attack benefit, so that triggers can be easily detected by the naked eye. To fill this gap, we systemically discuss the robustness of ViTs and CNNs under basic backdoor attacks with differ- ent trigger settings and find that ViTs seem to be more vul- nerable to patch-wise triggers rather than image-blending triggers. Delving into the essence of ViTs, images are di- vided into patches as tokens to calculate attentions, which can capture more interaction information between patches at the global level than CNNs [44]. Thus the patch-wise perturbation has been shown to sufficiently affect the self- attention mechanism of ViTs [20] and make ViTs weaker learners than CNNs. Inspired by this, a natural and interest- ing question is whether backdoor attacks with the patch- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24605 wise trigger are resultful in ViTs. Accordingly, we pro- pose BadViT, a well-designed backdoor attack framework against ViTs. Through the optimization process, the univer- sal adversarial patch is generated as a trigger, which can be better caught by the model through the self-attention mech- anism of ViTs to tighten the connection between the trig- ger and the target class. To achieve invisible attacks, we limit the perturbation strength of the adversarial patch-wise trigger and adopt the blending strategy [12] instead of past- ing. Moreover, we adopt a ViTs backdoor defense Patch- Drop [16], and two state-of-the-art defenses in CNNs, Neu- ral Cleanse [59] and FinePruning [33], to explore the vul- nerability of ViTs against our BadViT. Our main contributions are as follows: • We explore the robustness of ViTs compared with CNNs against backdoor attacks with different triggers. • We propose our BadViT as well as its invisible version and verify the validity through abundant experiments. • We show the effect of BadViTs under three advanced defense methods, and further discuss the characteris- tics of ViTs under backdoor attacks through patch pro- cessing, reverse engineering, and pruning. We believe this paper will offer readers a new under- standing of the robustness of ViTs against backdoor attacks, and provide constructive insights into the follow-up ViTs system optimization and backdoor defense efforts.

Zhang_DeSTSeg_Segmentation_Guided_Denoising_Student-Teacher_for_Anomaly_Detection_CVPR_2023

Abstract Visual anomaly detection, an important problem in com- puter vision, is usually formulated as a one-class classifi- cation and segmentation task. The student-teacher (S-T) framework has proved to be effective in solving this chal- lenge. However, previous works based on S-T only empir- ically applied constraints on normal data and fused multi- level information. In this study, we propose an improved model called DeSTSeg, which integrates a pre-trained teacher network, a denoising student encoder-decoder, and a segmentation network into one framework. First, to strengthen the constraints on anomalous data, we intro- duce a denoising procedure that allows the student net- work to learn more robust representations. From synthet- ically corrupted normal images, we train the student net- work to match the teacher network feature of the same im- ages without corruption. Second, to fuse the multi-level S-T features adaptively, we train a segmentation network with rich supervision from synthetic anomaly masks, achieving a substantial performance improvement. Experiments on the industrial inspection benchmark dataset demonstrate that our method achieves state-of-the-art performance, 98.6% on image-level AUC, 75.8% on pixel-level average preci- sion, and 76.4% on instance-level average precision.

1. Introduction Visual anomaly detection (AD) with localization is an essential task in many computer vision applications such as industrial inspection [24, 36], medical disease screening [27, 32], and video surveillance [18, 20]. The objective of these tasks is to identify both corrupted images and anoma- lous pixels in corrupted images. As anomalous samples oc- cur rarely, and the number of anomaly types is enormous, it is unlikely to acquire enough anomalous samples with all possible anomaly types for training. Therefore, AD tasks were usually formulated as a one-class classification andsegmentation, using only normal data for model training. The student-teacher (S-T) framework, known as knowl- edge distillation, has proven effective in AD [3, 9, 26, 31, 33]. In this framework, a teacher network is pre-trained on a large-scale dataset, such as ImageNet [10], and a student network is trained to mimic the feature represen- tations of the teacher network on an AD dataset with nor- mal samples only. The primary hypothesis is that the stu- dent network will generate different feature representations from the teacher network on anomalous samples that have never been encountered in training. Consequently, anoma- lous pixels and images can be recognized in the inference phase. Notably, [26, 31] applied knowledge distillation at various levels of the feature pyramid so that discrepan- cies from multiple layers were aggregated and demonstrated good performance. However, there is no guarantee that the features of anomalous samples are always different between S-T networks because there is no constraint from anoma- lous samples during the training. Even with anomalies, the student network may be over-generalized [22] and output similar feature representations as those by the teacher net- work. Furthermore, aggregating discrepancies from multi- level in an empirical way, such as sum or product, could be suboptimal. For instance, in the MVTec AD dataset under the same context of [31], we observe that for the category of transistor , employing the representation from the last layer, with 88.4% on pixel-level AUC, outperforms that from the multi-level features, with 81.9% on pixel-level AUC. To address the problem mentioned above, we propose DeSTSeg , illustrated in Fig. 1, which consists of a denois- ing student network, a teacher network, and a segmenta- tion network. We introduce random synthetic anomalies into the normal images and then use these corrupted im- ages1for training. The denoising student network takes a corrupted image as input, whereas the teacher network takes the original clean image as input. During training, 1All samples shown in this paper are licensed under the CC BY-NC-SA 4.0. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3914 Figure 1. Overview of DeSTSeg. Synthetic anomalous images are generated and used during training. In the first step (a), the student network with synthetic input is trained to generate similar feature representations as the teacher network from the clean image. In the second step (b), the element-wise product of the student and teacher networks’ normalized outputs are concatenated and utilized to train the segmentation network. The segmentation output is the predicted anomaly score map. the feature discrepancy between the two networks is min- imized. In other words, the student network is trained to perform denoising in the feature space. Given anomalous images as input to both networks, the teacher network en- codes anomalies naturally into features, while the trained denoising student network filters anomalies out of feature space. Therefore, the two networks are reinforced to gen- erate distinct features from anomalous inputs. For the ar- chitecture of the denoising student network, we decided to use an encoder-decoder network for better feature denoising instead of adopting an identical architecture as the teacher network. In addition, instead of using empirical aggrega- tion, we append a segmentation network to fuse the multi- level feature discrepancies in a trainable manner, using the generated binary anomaly mask as the supervision signal. We evaluate our method on a benchmark dataset for sur- face anomaly detection and localization, MVTec AD [2]. Extensive experimental results show that our method out- performs the state-of-the-art methods on image-level, pixel- level, and instance-level anomaly detection tasks. We also conduct ablation studies to validate the effectiveness of our proposed components. Our main contributions are summarized as follows. (1)We propose a denoising student encoder-decoder, which is trained to explicitly generate different feature representa- tions from the teacher with anomalous inputs. (2) We em- ploy a segmentation network to adaptively fuse the multi- level feature similarities to replace the empirical inference approach. (3) We conduct extensive experiments on the benchmark dataset to demonstrate the effectiveness of our method for various tasks.

Zhang_Inversion-Based_Style_Transfer_With_Diffusion_Models_CVPR_2023

Abstract The artistic style within a painting is the means of ex- pression, which includes not only the painting material, colors, and brushstrokes, but also the high-level attributes, including semantic elements and object shapes. Previous arbitrary example-guided artistic image generation meth- ods often fail to control shape changes or convey elements. Pre-trained text-to-image synthesis diffusion probabilistic models have achieved remarkable quality but often require extensive textual descriptions to accurately portray the at- tributes of a particular painting. The uniqueness of an art- work lies in the fact that it cannot be adequately explained with normal language. Our key idea is to learn the artis- tic style directly from a single painting and then guide the synthesis without providing complex textual descriptions. Specifically, we perceive style as a learnable textual de- scription of a painting. We propose an inversion-based style transfer method (InST), which can efficiently and accurately learn the key information of an image, thus capturing and transferring the artistic style of a painting. We demonstrate the quality and efficiency of our method on numerous paint- ings of various artists and styles. Codes are available at https://github.com/zyxElsa/InST . *Corresponding author: [email protected]

1. Introduction If a photo speaks 1000 words, then every painting tells a story. A painting contains the engagement of an artist’s own creation. The artistic style of a painting can be the per- sonalized textures and brushstrokes, the portrayed beautiful moment, or some particular semantic elements. All these artistic factors are difficult to describe with words. There- fore, when we wish to utilize a favorite painting to create new digital artworks that can imitate the original idea of the artist, the task turns into example-guided artistic image gen- eration. Generating artistic image(s) from given example(s) has attracted great interest in recent years. A typical task is style transfer [1, 6, 8, 17, 34, 55, 59], which can create a new artistic image from an arbitrary input pair of natural images and painting image, by combining the content of the nat- ural image and the style of the painting image. However, particular artistic factors such as object shape and semantic elements are difficult to be transferred (see Figures 2(b) and 2(e)). Text guided stylization [14, 16, 29, 38] produces an artistic image from a natural image and a text prompt, but the text prompt for the target style can usually be a rough description only of the material (e.g., “oil”, “watercolor” or “sketch”), art movement(e.g. “Impressionism” or “Cu- bism”), artist (e.g., “Vincent van Gogh” or “Claude Monet”) or a famous artwork (e.g., “Starry Night” or “The Scream”). This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10146 ‘[C]’ ‘A painting of a girl reads a book in the style of [C]’ Target Style CNN-Based Style Transfer Model Style Image Text-to-Image Generative Model‘A painting of a girl reads a book in the style of Tony Toscani’ Text-to-Image Generative Model Text-to-Image Generative Model (a) Text-to-Image Generative ModelRandom NoiseRandom NoiseRandom Noise Vit-Based Style Transfer Model Style Image (b) (c) (d) (e) (f)‘A painting of a girl reads a book in the style of Modernism’ ’Figure 2. Concept differences between text-to-image synthesis [50], classical style transfer [7, 59] and our InST. Diffusion-based methods [9,22,23,26,36,54] generate high- quality and diverse artistic images on the basis of a text prompt, with or without image examples. In addition to the input image, a detailed auxiliary textual input is required to guide the generation process if we want to reproduce some vivid contents and styles. However, the creative idea of a specific painting may still be difficult to reproduce in the result. In this paper, we propose a novel example-guided artis- tic image generation framework (i.e., inversion-based style transfer, InST) which related to style transfer and text-to- image synthesis, to mitigate all the above problems. Given only a single input painting image, our method can learn and transfer its style to a natural image with a very simple text prompt (see Figures 1 and 2(f)). The resulting image exhibits very similar artistic attributes to those of the orig- inal painting, including material, brushstrokes, colors, ob- ject shapes and semantic elements, without losing diversity. Furthermore, the content of the resulting image can also be controlled by providing a text description (see Figure 2(c)). To achieve this goal, we need to obtain the representa- tion of the image style, which refers to the set of attributes that appear in the high-level textual description of the im- age. We define the textual descriptions as “new words” that do not exist in the normal language and obtain the embed- dings via inversion method. We exploit the recent success of diffusion models [42,50] and inversion [2,15]. We adapt diffusion models in our work as the backbone to be inverted and as a generator in image-to-image and text-to-image syn- thesis. Specifically, we propose an efficient and accurate textual inversion based on the attention mechanism, which can quickly learn key features from an image, and a stochas- tic inversion to maintain the semantic of the content image. We use CLIP [39] image encoder and learn key information of the image through multi-layer cross-attention. Taking an artistic image as a reference, the attention-based inversion module is fed with its image embedding and then gives its textual embedding. The diffusion models conditioning onthe textual embedding can produce new images with the learned style of the reference. To demonstrate the effectiveness of InST, we conduct comprehensive experiments on numerous images of various artists and styles. The experiments show that InST produces outstanding results, generating artistic images that imitate the style attributes to a high degree and with a content that is consistent with that of the input natural images or text de- scriptions. InST demonstrates much improved visual qual- ity and artistic consistency compared with state-of-the-art approaches. These outcomes indicate the generality, preci- sion, and adaptability of the proposed method.

Yang_Pruning_Parameterization_With_Bi-Level_Optimization_for_Efficient_Semantic_Segmentation_on_CVPR_2023

Abstract With the ever-increasing popularity of edge devices, it is necessary to implement real-time segmentation on the edge for autonomous driving and many other applications. Vi- sion Transformers (ViTs) have shown considerably stronger results for many vision tasks. However, ViTs with the full- attention mechanism usually consume a large number of computational resources, leading to difficulties for real- time inference on edge devices. In this paper, we aim to de- rive ViTs with fewer computations and fast inference speed to facilitate the dense prediction of semantic segmentation on edge devices. To achieve this, we propose a pruning pa- rameterization method to formulate the pruning problem of semantic segmentation. Then we adopt a bi-level optimiza- tion method to solve this problem with the help of implicit gradients. Our experimental results demonstrate that we can achieve 38.9 mIoU on ADE20K val with a speed of 56.5 FPS on Samsung S21, which is the highest mIoU under the same computation constraint with real-time inference.

1. Introduction Inspired by the extraordinary performance of Deep Neu- ral Networks (DNNs), DNNs have been applied to various tasks. In this work, we focus on semantic segmentation, which aims to assign a class label to each pixel of an image to perform a dense prediction. It plays an important role in many real-world applications, such as autonomous driving. However, as a dense prediction task, segmentation models usually have complicated multi-scale feature fusion struc- tures with large feature sizes, leading to tremendous mem- ory and computation overhead with slow inference speed. To reduce the memory and computation cost, certain lightweight CNN architectures [21, 40, 66] are designed for efficient segmentation. Besides CNNs, inspired by the re- cent superior performance of vision transformers (ViTs) *These authors contributed equally to this work. Figure 1. Comparison of accuracy versus FPS on ADE20K. [17], some works [10, 11, 72] adopt ViTs in segmentation tasks to explore self-attention mechanism with the global receptive field. However, it is still difficult for ViTs to re- duce the computation cost of the dense prediction for seg- mentation with large feature sizes. With the wide spread of edge devices such as mobile phones, it is essential to perform real-time inference of seg- mentation on edge devices in practice. To facilitate mo- bile segmentation, the state-of-the-art work TopFormer [68] adopts a token pyramid transformer to produce scale-aware semantic features with tokens from various scales. It signif- icantly outperforms CNN- and ViT-based networks across different semantic segmentation datasets and achieves a good trade-off between accuracy and latency. However, it only partially optimizes the token pyramid module, which costs most of the computations and latency. In this work, we propose a pruning parameterization method with bi-level optimization to further enhance the performance of TopFormer. Our objective is to search for a suitable layer width for each layer in the token pyramid This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15402 module, which is the main cost of computations and latency (over 60%). To achieve this, we first formulate the problem with pruning parameterization to build a pruning framework with a soft mask as a representation of the pruning policy. With this soft mask, we further adopt thresholding to con- vert the soft mask into a binary mask so that the model is trained with actual pruned weights to obtain pruning results directly. This is significantly different from other meth- ods [27, 47] to train with unpruned small non-zero weights and use fine-tuning to mitigate the performance degradation after applying pruning. Besides, to update the soft mask as long as the pruning policy, we adopt to straight though es- timator (STE) method to make the soft mask differentiable. Thus, we can build the pruning parameterization framework with minimal overhead. Based on this framework, we need to search the best- suited layer width for each layer in the token pyramid mod- ule. It is non-trivial to perform the search. As the to- ken pyramid module needs to extract multi-scale informa- tion from multiple spatial resolutions, the large hierarchical search space leads to difficulties of convergence. To resolve this problem, we adopt a bi-level optimization method. In the outer optimization, we try to obtain the pruning pol- icy based on the pruning parameters (the soft mask). In the inner optimization, the optimized model weights with the best segmentation performance under this soft mask can be obtained. Compared with a typical pruning method, our work incorporates the implicit gradients with second-order derivatives to further guide the update of the soft mask and achieve better performance. Our experimental results demonstrate that we can achieve 38.9 mIoU (mean class- wise intersection-over-union) on the ADE20K dataset with a speed of 56.5 FPS on Samsung S21, which is the highest mIoU under the same computation constraint with real-time inference speed. As demonstrated in Figure 1, our models can achieve a better tradeoff between the mIoU and speed. We summarize our contributions below, • We propose a pruning parameterization method to build a pruning framework with a soft mask. We further use a threshold-based method to convert the soft mask into the binary mask to perform actual pruning during model training and inference. Besides, STE is adopted to update the soft mask efficiently through gradient descent opti- mizers. • To solve the pruning problem formulated with the frame- work of pruning parameterization, we propose a bi-level optimization method to utilize implicit gradients for bet- ter results. We show that the second-order derivatives in the implicit gradients can be efficiently obtained through first-order derivatives, saving computations and memory. • Our experimental results demonstrate that we can achieve the highest mIoU under the same computation constraint on various datasets. Specifically, we can achieve 38.9mIoU on the ADE20K dataset with a real-time inference speed of 56.5 FPS on the Samsung S21.

Zhang_CompletionFormer_Depth_Completion_With_Convolutions_and_Vision_Transformers_CVPR_2023

Abstract Given sparse depths and the corresponding RGB images, depth completion aims at spatially propagating the sparse measurements throughout the whole image to get a dense depth prediction. Despite the tremendous progress of deep- learning-based depth completion methods, the locality of the convolutional layer or graph model makes it hard for the network to model the long-range relationship between pixels. While recent fully Transformer-based architecture has reported encouraging results with the global recep- tive field, the performance and efficiency gaps to the well- developed CNN models still exist because of its deteriora- tive local feature details. This paper proposes a Joint Con- volutional Attention and Transformer block (JCAT), which deeply couples the convolutional attention layer and Vision Transformer into one block, as the basic unit to construct our depth completion model in a pyramidal structure. This hybrid architecture naturally benefits both the local connec- tivity of convolutions and the global context of the Trans- former in one single model. As a result, our Completion- Former outperforms state-of-the-art CNNs-based methods on the outdoor KITTI Depth Completion benchmark and indoor NYUv2 dataset, achieving significantly higher effi- ciency (nearly 1/3 FLOPs) compared to pure Transformer- based methods. Code is available at https://github. com/youmi-zym/CompletionFormer .

1. Introduction Active depth sensing has achieved significant gains in performance and demonstrated its utility in numerous ap- plications, such as autonomous driving and augmented re- ality. Although depth maps captured by existing commer- cial depth sensors ( e.g., Microsoft Kinect [ 23], Intel Re- alSense [ 11]) or depths points within the same scanning line of LiDAR sensors are dense, the distance between valid/correct depth points could still be far owing to thesensor noise, challenging conditions such as transparent, shining, and dark surfaces, or the limited number of scan- ning lines of LiDAR sensors. To address these issues, depth completion [ 2,16,26,31], which targets at completing and reconstructing the whole depth map from sparse depth mea- surements and a corresponding RGB image ( i.e., RGBD), has gained much attention in the latest years. For depth completion, one key point is to get the depth affinity among neighboring pixels so that reliable depth la- bels can be propagated to the surroundings [ 2,3,8,16,26]. Based on the fact that the given sparse depth could be highly sparse due to noise or even no measurement be- ing returned from the depth sensor, it requires depth com- pletion methods to be capable of 1) detecting depth out- liers by measuring the spatial relationship between pixels in both local and global perspectives; 2) fusing valid depth values from close or even extremely far distance points. All these properties ask the network for the potential to capture both local and global correlations between pixels. Current depth completion networks collect context infor- mation with the widely used convolution neural networks (CNNs) [ 2,3,8,16,26,29,37,51] or graph neural net- work [ 42,49]. However, both the convolutional layer and graph models can only aggregate within a local region, e.g. square kernel in 3×3for convolution and kNN-based neigh- borhood for graph models [ 42,49], making it still tough to model global long-range relationship, in particular within the shallowest layers of the architecture. Recently, Guide- Former [ 31] resorts fully Transformer-based architecture to enable global reasoning. Unfortunately, since Vision Trans- formers project image patches into vectors through a single step, this causes the loss of local details, resulting in ignor- ing local feature details in dense prediction tasks [ 28,43]. For depth completion, the limitations affecting pure CNNs or Transformer based networks also manifest, as shown in Fig.1. Despite anydistance the reliable depth points could be distributed at, exploring an elegant integration of these two distinct paradigms, i.e. CNNs and Transformer, has not This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18527 (a) RGB (b) Pure CNNs (d) CompletionFormer (c) Pure Transformer Figure 1. Comparison of attention maps of pure CNNs, Vision Transformer, and the proposed CompletionFormer with joint CNNs and Transformer structure. The pixel highlighted with a yellow cross in RGB image (a) is the one we want to observe how the network predicts it. Pure CNNs architecture (b) activates discriminative local regions ( i.e., the region on the fire extinguisher), whereas pure Transformer based models (c) activate globally yet fail on local details. In contrast, our full CompletionFormer (d) can retain both the local details and global context. been studied for depth completion yet. In this work, we propose CompletionFormer, a pyrami- dal architecture coupling CNN-based local features with Transformer-based global representations for enhanced depth completion. Generally, there are two gaps we are facing: 1) the content gap between RGB and depth in- put; 2) the semantic gap between convolution and Trans- former. As for the multimodal input, we propose embed- ding the RGB and depth information at the early network stage. Thus our CompletionFormer can be implemented in an efficient single-branch architecture as shown in Fig. 2 and multimodal information can be aggregated throughout the whole network. Considering the integration of convolu- tion and Transformer, previous work has explored from sev- eral different perspectives [ 6,12,25,28,43] on image classi- fication and object detection. Although state-of-the-art per- formance has been achieved on those tasks, high computa- tion cost [ 12] or inferior performance [ 6,12] are observed when these networks are directly adapted to depth com- pletion task. To promise the combination of self-attention and convolution still being efficient, and also effective, we embrace convolutional attention and Transformer into one block and use it as the basic unit to construct our network in a multi-scale style. Specifically, the Transformer layer is inspired by Pyramid Vision Transformer [ 39], which adopts spatial-reduction attention to make the Transformer layer much more lightweight. As for the convolution-related part, the common option is to use plain convolutions such as In- verted Residual Block [ 32]. However, the huge semantic gap between convolution and the Transformer and the lost local details by Transformer require the convolutional lay- ers to increase its own capacity to compensate for it. Fol- lowing this rationale, we further introduce spatial and chan- nel attention [ 40] to enhance convolutions. As a result, without any extra module to bridge the content and semantic gaps [ 12,28,31], every convolution and Transformer layer in the proposed block can access the local and global fea- tures. Hence, information exchange and fusion happen ef- fectively at every block of our network. To summarize, our main contributions are as follows: • We propose integrating Vision Transformer with con-volutional attention layers into one block for depth completion, enabling the network to possess both local and global receptive fields for multi-modal information interaction and fusion. In particular, spatial and chan- nel attention are introduced to increase the capacity of convolutional layers. • Taking the proposed J oint C onvolutional A ttention and Transformer (JCAT) block as the basic unit, we intro- duce a single-branch network structure, i.e. Comple- tionFormer. This elegant design leads to a compara- ble computation cost to current CNN-based methods while presenting significantly higher efficiency when compared with pure Transformer based methods. • Our CompletionFormer yields substantial improve- ments to depth completion compared to state-of-the-art methods, especially when the provided depth is very sparse, as often occurs in practical applications.

Yu_How_To_Prevent_the_Continuous_Damage_of_Noises_To_Model_CVPR_2023

Abstract Deep learning with noisy labels is challenging and in- evitable in many circumstances. Existing methods reduce the impact of mislabeled samples by reducing loss weights or screening, which highly rely on the model’s superior discriminative power for identifying mislabeled samples. However, in the training stage, the trainee model is im- perfect and will wrongly predict some mislabeled samples, which cause continuous damage to the model training. Con- sequently, there is a large performance gap between exist- ing anti-noise models trained with noisy samples and mod- els trained with clean samples. In this paper, we put forward a Gradient Switching Strategy (GSS) to prevent the contin- uous damage of mislabeled samples to the classifier. Theo- retical analysis shows that the damage comes from the mis- leading gradient direction computed from the mislabeled samples. The trainee model will deviate from the correct optimization direction under the influence of the accumu- lated misleading gradient of mislabeled samples. To ad- dress this problem, the proposed GSS alleviates the damage by switching the gradient direction of each sample based on the gradient direction pool, which contains all-class gradi- ent directions with different probabilities. During training, each gradient direction pool is updated iteratively, which assigns higher probabilities to potential principal direc- tions for high-confidence samples. Conversely, uncertain samples are forced to explore in different directions rather than mislead model in a fixed direction. Extensive experi- ments show that GSS can achieve comparable performance with a model trained with clean data. Moreover, the pro- posed GSS is pluggable for existing frameworks. This idea of switching gradient directions provides a new perspective for future noisy-label learning.

1. Introduction Recently, Deep Neural Networks (DNNs) have achieved breakthrough results across various computer vision tasks [9, 14–16, 20, 34, 36, 49, 50]. The high performance *Li Sun is the corresponding author. Figure 1. Performance comparison of existing methods and the proposed GSS on CIFAR-10 with 40% noisy labels. The red dashed line denotes the upper limit, which is the accuracy of mod- els trained with completely clean labels. of DNNs requires a large amount of labeled data, but it is hard to guarantee label quality in many circumstances. As a matter of fact, many benchmark datasets inevitably contain noisy labels according to investigation results in [35]. Various types of researches are proposed to address the noisy-label problem. The mainstream types are robust loss function [18, 27, 44, 54] and sample screening [30, 37, 40]. These methods deal with mislabeled samples in essen- tially similar ways, that is, by decreasing the weights of low-confidence samples, which highly rely on the trainee model’s discriminative power of identifying mislabeled samples. However, during training the trainee model is imperfect and will miss many mislabeled samples, which will continuously damage the model. That is why there is a large performance gap between existing anti-noise models trained with noisy samples and models trained with clean samples. As shown in Fig. 1, existing anti-noise meth- ods (denoted by solid lines) have 1.61%∼9.81% accu- racy gaps compared with the reference upper limit (red dash line), which denotes the performance of models trained with clean samples. It raises an important question: how to prevent the continuous damage of noises to model train- ing? The theoretical analysis in Section 4 shows that the noise damage comes from the misleading gradient direc- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12054 tions caused by noises. Therefore, it is a viable solution to handle the continuous damage of noises to model training by eliminating the impact of misleading gradient directions. In this paper, we put forward a Gradient Switching Strat- egy (GSS) to prevent the continuous damage of mislabeled samples to the model training. The core idea is assigning a random gradient direction to cancel out the negative im- pact of mislabeled samples, especially for uncertain sam- ples which could continuously generate a misleading gra- dient in a single direction. For high-confidence samples, the model will be optimized using their potential principal directions with a larger probability. As the model’s dis- criminative power grows over training time, parts of uncer- tain samples will become high-confidence samples, which in turn optimizes the model with their potential principal di- rections. Finally, the model will be well-trained with almost all samples in the dataset step by step. Specifically, we devise a gradient direction pool for each sample, which contains all-class gradient directions with different probabilities. The probabilities of different gra- dient directions are determined based on the original noisy label, predictions, and partial randomness. In the training stage, for uncertain samples, the probabilities of different gradient directions are dominated by randomness. The mul- tiple random gradient directions prevent a fixed misdirec- tion from continuously damaging the training. The high-confidence samples consist of two groups: the predictions are consistent with original labels (consistent sample), and the predictions are not consistent with original labels (non-consistent sample). For consistent samples, the gradient direction of the original label (potential principal direction) has a higher probability than those for the remain- ing gradient directions. For non-consistent samples, two highest probabilities correspond to the gradient directions of the original label and model prediction. The model ex- plores two gradient directions and determines the potential principal direction during training. In summary, the poten- tial principal directions of high-confidence samples guide the optimization of the model. Experiment results demonstrate that the proposed GSS can effectively prevent the damage of mislabeled samples to the model training. The proposed GSS is pluggable for existing frameworks for noisy-label learning, which can achieve 1.23%∼9.22% accuracy improvement than SOTA for high noise rates. Additionally, the model with GSS trained on noisy samples can achieve comparable perfor- mance with models trained with clean samples. Overall, our contributions are summarized as follows: • This paper is the first to clarify the continuous damage of the mislabeled samples to model training. Theoreti- cal analysis shows the continuous damage comes from the misleading gradient direction derived from misla- beled samples, which provides a new perspective forfuture noisy-label learning research. • We propose the Gradient Switching Strategy (GSS) to prevent the continuous gradient damage of mislabeled samples to the model training. A gradient direction pool containing gradient directions of all classes with dynamic probabilities for each sample is devised to al- leviate the impact of uncertain samples and optimize the model with the potential principal direction. • Detailed theoretical analysis and extensive experimen- tal results show that the proposed GSS can effectively prevent damage of mislabeled samples. Through com- bining GSS with existing anti-noise learning methods, the final classification performance can achieve up to 1.23%∼9.22% accuracy improvement over SOTA on datasets with severe noise, some of which are even comparable to the model trained with clean samples.

Zhang_3D-Aware_Object_Goal_Navigation_via_Simultaneous_Exploration_and_Identification_CVPR_2023

Abstract Object goal navigation (ObjectNav) in unseen environ- ments is a fundamental task for Embodied AI. Agents in ex- isting works learn ObjectNav policies based on 2D maps, scene graphs, or image sequences. Considering this task happens in 3D space, a 3D-aware agent can advance its ObjectNav capability via learning from fine-grained spa- tial information. However, leveraging 3D scene representa- tion can be prohibitively unpractical for policy learning in this floor-level task, due to low sample efficiency and expen- sive computational cost. In this work, we propose a frame- work for the challenging 3D-aware ObjectNav based on two straightforward sub-policies. The two sub-polices, namely corner-guided exploration policy and category-aware iden- tification policy, simultaneously perform by utilizing on- line fused 3D points as observation. Through extensive ex- periments, we show that this framework can dramatically improve the performance in ObjectNav through learning from 3D scene representation. Our framework achieves the best performance among all modular-based methods on the Matterport3D and Gibson datasets, while requiring (up to 30x) less computational cost for training. The code will be released to benefit the community.1

1. Introduction As a vital task for intelligent embodied agents, object goal navigation (ObjectNav) [38, 49] requires an agent to find an object of a particular category in an unseen and unmapped scene. Existing works tackle this task through end-to-end reinforcement learning (RL) [27, 36, 47, 51] or modular-based methods [9, 14, 35]. End-to-end RL based methods take as input the image sequences and directly output low-level navigation actions, achieving competitive *Joint first authors †Corresponding author: [email protected] 1Homepage: https://pku-epic.github.io/3D-Aware-ObjectNav/ : sofa : chair: cushion Looking for a chair A C B A C B C Figure 1. We present a 3D-aware ObjectNav framework along with simultaneous exploration and identification policies: A→B, the agent was guided by an exploration policy to look for its target; B→C, the agent consistently identified a target object and finally called STOP. performance while suffering from lower sample efficiency and poor generalizability across datasets [3, 27]. Therefore, we favor modular-based methods, which usually contain the following modules: a semantic scene mapping module that aggregates the RGBD observations and the outputs from semantic segmentation networks to form a semantic scene map; an RL-based goal policy module that takes as input the semantic scene map and learns to online update a goal location; finally, a local path planning module that drives the agent to that goal. Under this design, the semantic ac- curacy and geometric structure of the scene map are crucial to the success of object goal navigation. We observe that the existing modular-based methods mainly construct 2D maps [8, 9], scene graphs [34, 56] or neural fields [43] as their scene maps. Given that objects lie in 3D space, these scene maps are inevitably deficient in leveraging 3D spatial information of the environment com- prehensively and thus have been a bottleneck for further improving object goal navigation. In contrast, forming a 3D scene representation naturally offers more accurate, spa- tially dense and consistent semantic predictions than its 2D counterpart, as proved by [12, 31, 45]. Hence, if the agent could take advantage of the 3D scene understanding and This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6672 form a 3D semantic scene map, it is expected to advance the performance of ObjectNav. However, leveraging 3D scene representation would bring great challenges to ObjectNav policy learning. First, building and querying fine-grained 3D representation across a floor-level scene requires extensive computational cost, which can significantly slow down the training of RL [7,55]. Also, 3D scene representation induces considerably more complex and high-dimensional observations to the goal policy than its 2D counterpart, leading to a lower sam- ple efficiency and hampering the navigation policy learn- ing [22, 57]. As a result, it is demanding to design a frame- work to efficiently and effectively leverage powerful 3D in- formation for ObjectNav. To tackle these challenges, we propose a novel frame- work composed of an online semantic point fusion module for 3D semantic scene mapping and two parallel policy net- works in charge of scene exploration and object identifica- tion, along with a local path planning module. Our online semantic point fusion module extends a highly efficient on- line point construction algorithm [53] to enable online se- mantic fusion and spatial semantic consistency computation from captured RGBD sequences. This 3D scene construc- tion empowers a comprehensive 3D scene understanding for ObjectNav. Moreover, compared to dense voxel-based methods [7, 55], our point-based fusion algorithm are more memory-efficient [40,46] which makes it practically usable for floor-level navigation task. (See Figure 1) Moreover, to ease the learning of navigation policy, we further propose to factorize the navigation policy into two sub-policies, namely exploration and identification. The two policies simultaneously perform to roll out an explo- ration goal and an identified object goal (if exist), respec- tively. Then the input for the local path planning module will switch between these two goals, depending on whether there exists an identified target object. More specifically, we propose a corner-guided exploration policy which learns to predict a long-term discrete goal at one of the four corners of the bounding box of the scene. These corner goals ef- ficiently drive the agent to perceive the surroundings and explore regions where the target object is possibly settled. And for identification, a category-aware identification pol- icy is proposed to dynamically learn a discrete confidence threshold to identify the semantic predictions for each cat- egory. Both of these policies are trained by RL in low- dimensional discrete action space. Through experiments, the simultaneous two-policy mechanism and discrete action space design dramatically reduce the difficulty in learning for 3D-aware ObjectNav and achieve better performance than existing modular-based navigation strategies [26, 35]. Through extensive evaluation on the public benchmarks, we demonstrate that our method performs online 3D-awareObjectNav at 15 FPS while achieving the state-of-the- art performance on navigation efficiency. Moreover, our method outperforms all other modular-based methods in both efficiency and success rate with up to 30x times less computational cost. Our main contributions include: • We present the first 3D-aware framework for Object- Nav task. • We build an online point-based construction and fusion algorithm for efficient and comprehensive understand- ing of floor-level 3D scene representation. • We propose a simultaneous two-policy mechanism which mitigates the problem of low sample efficiency in 3D-aware ObjectNav policy learning.

Yu_V2X-Seq_A_Large-Scale_Sequential_Dataset_for_Vehicle-Infrastructure_Cooperative_Perception_and_CVPR_2023

Abstract Utilizing infrastructure and vehicle-side information to track and forecast the behaviors of surrounding traf- fic participants can significantly improve decision-making and safety in autonomous driving. However, the lack of real-world sequential datasets limits research in this area. To address this issue, we introduce V2X-Seq, the first large-scale sequential V2X dataset, which includes data frames, trajectories, vector maps, and traffic lights captured from natural scenery. V2X-Seq comprises two parts: the sequential perception dataset, which includes more than 15,000 frames captured from 95 scenarios, and the trajectory forecasting dataset, which contains about 80,000 infrastructure-view scenarios, 80,000 vehicle-view scenarios, and 50,000 cooperative-view scenarios cap- tured from 28 intersections’ areas, covering 672 hours of data. Based on V2X-Seq, we introduce three new tasks for vehicle-infrastructure cooperative (VIC) autonomous driv- ing: VIC3D Tracking, Online-VIC Forecasting, and Offline- VIC Forecasting. We also provide benchmarks for the intro- duced tasks. Find data, code, and more up-to-date informa- tion at https://github.com/AIR-THU/DAIR-V2X-Seq.

1. Introduction Although single-vehicle autonomous driving has made significant advancements in recent years, it still faces sig- nificant safety challenges due to its limited perceptual field and inability to accurately forecast the behaviors of traffic participants. These challenges hinder autonomous vehicles from making well-informed decisions and driving safer. A promising solution to address these challenges is to leverage infrastructure information via Vehicle-to-Everything (V2X) *Corresponding author. Work done while at AIR. For any questions or discussions, please email [email protected]. 1021010WiWhRXWTUajecWRU\WiWhLaUge-VcaOeTUajecWRU\V2X VieZ SiQgOe-ageQW VieZDAIR-V2X-COPV2VV2X-SeT AUgRYeUVe nXScenceVKITTIWa\mRV2X-Sim 103Sim.Real ASRllRScaSeHighDNGSIMWIBAM10101RRSe3DFigure 1. Autonomous driving datasets. V2X-Seq is the first large- scale, real-world, and sequential V2X dataset. The green circle de- notes the real-world dataset, and the pink triangle denotes the sim- ulated dataset. The abscissa represents the number of sequences. communication, which has been shown to significantly ex- pand perception range and enhance autonomous driving safety [1, 38]. However, current research primarily focuses on utilizing infrastructure data to improve the perception ability of autonomous driving, particularly in the context of frame-by-frame 3D detection. To enable well-informed decision-making for autonomous vehicles, it is critical to also incorporate infrastructure data to track and predict the behavior of surrounding traffic participants. To accelerate the research on cooperative sequential per- ception and forecasting, we release a large-scale sequential V2X dataset, V2X-Seq. All elements of this dataset were captured and generated from real-world scenarios. Com- pared with DAIR-V2X [38], which focuses on 3D object detection tasks, V2X-Seq is specifically designed for track- ing and trajectory forecasting tasks. The V2X-Seq dataset is divided into two parts: the sequential perception dataset This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5486 Table 1. Comparison with the public autonomous driving dataset. ’-’ denotes that the information is not provided. ’Real/Sim.’ indi- cates whether the data was collected from the real world or a simulator. V2X view includes multi-vehicle cooperative view and vehicle- infrastructure cooperative view. V2X-Seq is the first large-scale sequential V2X dataset and focuses on vehicle-infrastructure cooperative view. All data elements, including the traffic light signals, are captured and generated from the real world. Dataset Year Real/Sim. ViewWith With With With Tracked TotalScenesTrajectory 3D Boxes Maps Traffic Light Objects/Scene Time (hour) KITTI [12] 2012 Real Single-vehicle ! ! # # 43.67 1.5 50 nuScenes [3] 2019 Real Single-vehicle ! ! ! # 75.75 5.5 1,000 Waymo Motion [10, 28] 2021 Real Single-vehicle ! ! ! ! - 574 103,354 Argoverse [5] 2019 Real Single-vehicle ! # ! # 50.03 320 324,557 ApolloScape [16, 25] 2019 Real Single-vehicle ! # # # 50.6 2.5 103 HighD [17] 2018 Real Drone ! # ! # - 16.5 5,940 WIBAM [14] 2021 Real Infrastructure # # # # 0 0.25 0 NGSIM [29] 2016 Sim. Infrastructure ! # # # - 1.5 540 V2X-Sim 2.0 [21] 2022 Sim. V2X ! ! # # - 0.3 100 OPV2V [35] 2021 Sim. V2X ! ! # # 26.5 0.2 73 Cooper(inf) [1] 2019 Sim. V2X ! ! # # 30 - <100 DAIR-V2X-C [38] 2021 Real V2X # ! ! # 0 0.5 100 V2X-Seq/Perception 2023 Real V2X ! ! ! # 110 0..43 95 V2X-Seq/Forecasting 2023 Real V2X ! ! ! ! 101 583 210,000 and the trajectory forecasting dataset. The sequential per- ception dataset comprises 15,000 frames captured from 95 scenarios, which include infrastructure images, infrastruc- ture point clouds, vehicle-side images, vehicle-side point clouds, 3D detection/tracking annotations, and vector maps. The trajectory forecasting dataset comprises 210,000 sce- narios, including 50,000 cooperative-view scenarios, that were mined from 672 hours of data collected from 28 in- tersection areas. To our knowledge, V2X-Seq is the first sequential V2X dataset that includes such a large-scale sce- narios, making it an ideal resource for developing and test- ing cooperative perception and forecasting algorithms. Based on the V2X-Seq dataset, we introduce three novel tasks for vehicle-infrastructure cooperative perception and forecasting. The first task is VIC3D Tracking, which aims to cooperatively locate, identify, and track 3D objects us- ing sequential sensor inputs from both the vehicle and in- frastructure. The second task is Online-VIC trajectory fore- casting, which focuses on accurately predicting future be- havior of target agents by utilizing past infrastructure tra- jectories, ego-vehicle trajectories, real-time traffic lights, and vector maps. The third task is Offline-VIC trajectory forecasting, which involves extracting relevant knowledge from previously collected infrastructure data to facilitate vehicle-side forecasting. These proposed tasks are accom- panied by rich benchmarks. Additionally, we propose an intermediate-level framework, FF-Tracking, to effectively solve the VIC3D Tracking task. The main contributions are organized as follows: • We release the V2X-Seq dataset, which constitutes the first large-scale sequential V2X dataset. All data are cap- tured and generated from the real world. • Based on the V2X-Seq dataset, we introduce three tasksfor the vehicle-infrastructure cooperative autonomous driving community. To enable a fair evaluation of these tasks, we have carefully designed a set of benchmarks. • We propose a middle fusion method, named FF-Tracking, for solving VIC3D Tracking and our proposed method can efficiently overcome the latency challenge.

Zhang_Transferable_Adversarial_Attacks_on_Vision_Transformers_With_Token_Gradient_Regularization_CVPR_2023

Abstract Vision transformers (ViTs) have been successfully de- ployed in a variety of computer vision tasks, but they are still vulnerable to adversarial samples. Transfer-based attacks use a local model to generate adversarial samples and di- rectly transfer them to attack a target black-box model. The high efficiency of transfer-based attacks makes it a severe security threat to ViT-based applications. Therefore, it is vi- tal to design effective transfer-based attacks to identify the deficiencies of ViTs beforehand in security-sensitive scenar- ios. Existing efforts generally focus on regularizing the in- put gradients to stabilize the updated direction of adversar- ial samples. However, the variance of the back-propagated gradients in intermediate blocks of ViTs may still be large, which may make the generated adversarial samples focus on some model-specific features and get stuck in poor lo- cal optima. To overcome the shortcomings of existing ap- proaches, we propose the Token Gradient Regularization (TGR) method. According to the structural characteristics of ViTs, TGR reduces the variance of the back-propagated gradient in each internal block of ViTs in a token-wise man- ner and utilizes the regularized gradient to generate adver- sarial samples. Extensive experiments on attacking both ViTs and CNNs confirm the superiority of our approach. Notably, compared to the state-of-the-art transfer-based at- tacks, our TGR offers a performance improvement of 8.8% on average.

1. Introduction Transformers have been widely deployed in the natu- ral language processing, achieving state-of-the-art perfor- mance. Vision transformer (ViT) [5] first adapts the trans- former structure to the computer vision, and manifests ex- cellent performance. Afterward, diverse variants of ViTs have been proposed to further improve its performance *Corresponding author. Figure 1. Illustration of our Token Gradient Regularization (TGR) method. The red-colored entry represents the back-propagated gradient with extreme values. The back-propagated gradients cor- responding to one token in the internal blocks of ViTs are called the token gradients. Since we regularize the back-propagated gradients in a token-wise manner, we eliminate the token gradi- ents (marked with crosses) where extreme gradients locate during back-propagation to reduce the gradient variance. We then use the regularized gradients to generate adversarial samples. [2, 26] and broaden its application to different computer vi- sion tasks [42, 43], which makes ViTs a well-recognized successor for convolutional neural networks (CNNs). Un- fortunately, recent studies have shown that ViTs are still vulnerable to adversarial attacks [1, 22], which add human- imperceptible noise to a clean image to mislead deep learn- ing models. It is thus of great importance to understand DNNs [13, 27, 28, 35] and devise effective attacking meth- ods to identify their deficiencies before deploying them in safety-critical applications [16, 17, 37]. Adversarial attacks can be generally partitioned into two categories. The first category is the white-box attack, where attackers can obtain the structures and weights of the tar- get models for generating adversarial samples. The second one is the black-box attack, where attackers cannot fetch the information of the target model. Among different black- box attacks, transfer-based methods employ white-box at- tacks to attack a local source model and directly transfer This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16415 the generated adversarial sample to attack the target black- box model. Due to their high efficiency and applicability, transfer-based attacks pose a serious threat to the security of ViT-based applications in practice. Therefore, in this work, we focus on transfer-based attacks on ViTs. There are generally two branches of transfer-based at- tacks in the literature [15]. The first one is based on input transformation, which aims to combine the input gradients of multiple transformed images to generate transferable per- turbations. Complementary to such methods, the second branch is based on gradient regularization, which modifies the back-propagated gradient to stabilize the update direc- tion of adversarial samples and escape from poor local op- tima. For example, Variance Tuning Method (VMI) [29] tunes the input gradient to reduce the variance of input gra- dients. However, the variance of the back-propagated gra- dients in intermediate blocks of ViTs may still be large, which may make the generated adversarial samples focus on some model-specific features with extreme gradient val- ues. As a result, the generated adversarial samples may still get stuck in poor local optima and possess limited transfer- ability across different models. To address the weaknesses of existing gradient regularization-based approaches, we propose the Token Gradient Regularization (TGR) method for transferable ad- versarial attacks on ViTs. According to the architecture of ViTs, TGR reduces the variance of the back-propagated gra- dient in each internal block of ViTs and utilizes the regular- ized gradient to generate adversarial samples. More specifically, ViTs crop one image into small patches and treat these patches as a sequence of input tokens to fit the architecture of the transformer. The output tokens of internal blocks in ViTs correspond to the extracted inter- mediate features. Therefore, we view token representations as basic feature units in ViTs. We then examine the back- propagated gradients of the classification loss with respect to token representations in each internal block of ViTs, which we call token gradient in this work. As illustrated in Figure 1, we directly eliminate the back-propagated gra- dients with extreme values in a token-wise manner until ob- taining the regularized input gradients, which we used to update the adversarial samples. Consequently, we can re- duce the variance of the back-propagated gradients in in- termediate blocks of ViTs and produce more transferable adversarial perturbations. We conducted extensive experiments on the ImageNet dataset to validate the effectiveness of our proposed attack method. We examined the transferability of our generated adversarial samples to different ViTs and CNNs. Notably, compared with the state-of-the-art benchmarks, our pro- posed TGR shows a significant performance improvement of 8.8% on average. We summarize the contributions of this work as below:• We propose the Token Gradient Regularization (TGR) method for transferable adversarial attacks on ViTs. According to the architectures of ViTs, TGR regu- larizes the back-propagated gradient in each internal block of ViTs in a token-wise manner and utilizes the regularized gradient to generate adversarial samples. • We conducted extensive experiments to validate the ef- fectiveness of our approach. Experimental results con- firm that, on average, our approach can outperform the state-of-the-art attacking method with a significant margin of 8.8% on attacking ViT models and 6.2% on attacking CNN models. • We showed that our method can be combined with other compatible attack algorithms to further enhance the transferability of the generated adversarial sam- ples. Our method can also be extended to use CNNs as the local source models.

Zeng_3D-Aware_Facial_Landmark_Detection_via_Multi-View_Consistent_Training_on_Synthetic_CVPR_2023

Abstract Accurate facial landmark detection on wild images plays an essential role in human-computer interaction, enter- tainment, and medical applications. Existing approaches have limitations in enforcing 3D consistency while detect- ing 3D/2D facial landmarks due to the lack of multi-view in-the-wild training data. Fortunately, with the recent ad- vances in generative visual models and neural rendering, we have witnessed rapid progress towards high quality 3D image synthesis. In this work, we leverage such approaches to construct a synthetic dataset and propose a novel multi- view consistent learning strategy to improve 3D facial land- mark detection accuracy on in-the-wild images. The pro- posed 3D-aware module can be plugged into any learning- based landmark detection algorithm to enhance its accu- racy. We demonstrate the superiority of the proposed plug- *This work was done when Libing Zeng and Wentao Bao were interns at OPPO US Research Center, InnoPeak Technology, Inc.in module with extensive comparison against state-of-the- art methods on several real and synthetic datasets.

1. Introduction Accurate and precise facial landmark plays a signif- icant role in computer vision and graphics applications, such as face morphing [54], facial reenactment [58], 3D face reconstruction [17, 18, 30], head pose estimation [38], face recognition [1, 10, 13, 19, 32, 41, 71], and face genera- tion [11, 21, 60, 69]. In these applications, facial landmark detection provides great sparse representation to ease the burden of network convergence in different training stages and is often used as performance evaluation metric. For in- stance, as a facial prior, it provides good initialization for subsequent training [66, 67, 69, 76], good intermediate rep- resentation to bridge the gap between different modalities for content generation [11,27,51,79], loss terms which reg- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12747 ularize the facial expression [11, 52], or evaluation metrics to measure the facial motion quality [53, 73, 78]. The aforementioned applications require the estimated facial landmarks to be accurate even with significantly var- ied facial appearance under different identities, facial ex- pressions, and extreme head poses. Tremendous efforts have been devoted to address this problem [15, 22–24, 29, 34,40,56,63,74,75,77,82,84]. These approaches often rely on manually annotated large-scale lab-controlled or in-the- wild image datasets [4,34] to handle various factors such as arbitrary facial expressions, head poses, illumination, facial occlusions, etc. However, even with the high cost of human labeling, consistent andaccurate manual annotation of landmarks re- mains challenging [22, 23, 34]. It is very difficult, if not im- possible, to force a person to annotate the facial landmark keypoints at the same pixel locations for faces of different poses, let alone different annotators under different labeling environments. Such annotation inconsistency and inaccu- racy in training images are often the killing factor to learn an accurate landmark localization model. This is particu- larly a major problem in non-frontal faces where annotation becomes extremely challenging. As shown in Fig. 1(a) a small annotation variation in view #1, results in a signifi- cant inaccuracy in view #2. This multi-view inconsistency and inaccuracy can ultimately lead to poor landmark de- tection accuracy, especially for facial images with extreme head pose. To mitigate this annotation inconsistency and inaccuracy issue, we propose to learn facial landmark detection by en- forcing multi-view consistency during training. Given the images of the same facial identity captured with different head poses, instead of detecting facial landmark at each sep- arate facial image, we propose a multi-view consistency su- pervision to locate facial landmark in a holistic 3D-aware manner. To enforce multi-view consistency, we introduce self-projection consistency loss and multi-view landmark loss in training. We also propose an annotation genera- tion procedure to exploit the merits of lab-controlled data (e.g., multi-view images, consistent annotations) and in-the- wild data ( e.g., wide range of facial expressions, identities). Thanks to this synthetic data, our method does not rely on human annotation to obtain the accurate facial landmark locations. Therefore, it alleviates the problem of learning from inaccurate and inconsistent annotations. We formulate our solution as a plug-in 3D aware module, which can be incorporated into any facial landmark detec- tor and can boost a pre-trained model with higher accuracy and multi-view consistency. We demonstrate the effective- ness of our approach through extensive experiments on both synthetic and real datasets. The main contributions of our work are as follows: • We show, for the first time, how to combine the meritsof lab captured face image data ( e.g., multi-view) and the in-the-wild face image datasets ( e.g., appearance diversity). Using our proposed approach we produce a large-scale synthetic, but realistic, multi-view face dataset, titled DAD-3DHeads-Syn. • We propose a novel 3D-aware optimization module, which can be plugged into any learning-based facial landmark detection methods. By refining an existing landmark detection algorithm using our optimization module, we are able to improve its accuracy and multi- view consistency. • We demonstrate the performance improvements of our module built on top multiple baseline methods on sim- ulated dataset, lab-captured datasets, and in-the-wild datasets.

Zhang_Boosting_Verified_Training_for_Robust_Image_Classifications_via_Abstraction_CVPR_2023

Abstraction Zhaodi Zhang1,2,4, Zhiyi Xue2, Yang Chen2, Si Liu3, Yueling Zhang2, Jing Liu1, Min Zhang2 1Shanghai Key Laboratory of Trustworthy Computing 2East China Normal University 3ETH Z ¨urich4Chengdu Education Research Institute Abstract This paper proposes a novel, abstraction-based, certified training method for robust image classifiers. Via abstrac- tion, all perturbed images are mapped into intervals before feeding into neural networks for training. By training on intervals, all the perturbed images that are mapped to the same interval are classified as the same label, rendering the variance of training sets to be small and the loss landscape of the models to be smooth. Consequently, our approach significantly improves the robustness of trained models. For the abstraction, our training method also enables a sound and complete black-box verification approach, which is or- thogonal and scalable to arbitrary types of neural networks regardless of their sizes and architectures. We evaluate our method on a wide range of benchmarks in different scales. The experimental results show that our method outperforms state of the art by (i) reducing the verified errors of trained models up to 95.64%; (ii) totally achieving up to 602.50x speedup; and (iii) scaling up to larger models with up to 138 million trainable parameters. The demo is available at https://github.com/zhangzhaodi233/ABSCERT.git.

1. Introduction The robustness of image classifications based on neu- ral networks is attracting more attention than ever due to their applications to safety-critical domains such as self- driving [51] and medical diagnosis [43]. There has been a considerable amount of work on training robust neural networks against adversarial perturbations [1, 6–8, 10, 12, 28,32,56–59,62]. Conventional defending approaches aug- ment the training set with adversarial examples [5, 15, 27, 32,37,41,54,55]. They target only specific adversaries, de- pending on how the adversarial samples are generated [7], but cannot provide robustness guarantees [2, 11, 26]. Recent approaches attempt to train certifiably robust models with guarantees [2, 11, 14, 24, 26, 29, 49, 63]. They rely on the robustness verification results of neural networks in the training process. Most of the existing verification ap- proaches are based on symbolic interval propagation (SIP)[11,14,14,24,26], by which intervals are symbolically input into neural networks and propagated on a layer basis. There are, however, mainly three obstacles for these approaches to be widely adopted: (i) adding the verification results to the loss function for training brings limited improvement to the robustness of neural networks due to overestimation intro- duced in the verification phase; (ii) they are time-consuming due to the high complexity of the verification problem per se, e.g., NP-complete for the simplest ReLU-based fully connected feedforward neural networks [19]; and (iii) the verification is tied to specific types of neural networks in terms of network architectures and activation functions. To overcome the above obstacles, we propose a novel, abstraction-based approach for training verified robust im- age classifications whose inputs are numerical intervals . Regarding (i), we abstract each pixel of an image into an interval before inputting it into the neural network. The in- terval is numerically input to the neural network by assign- ing to two input neurons its lower and upper bounds, re- spectively. This guarantees that all the perturbations to the pixel in this interval do not alter the classification results, thereby improving the robustness of the network. More- over, this imposes no overestimation in the training phase. To address the challenge (ii), we use forward propagation and back propagation only to train the network without ex- tra time overhead. Regarding (iii), we treat the neural net- works as black boxes since we deal only with the input layer and do not change other layers. Hence, being agnostic to the actual neural network architectures, our approach can scale up to fairly large neural networks. Additionally, we iden- tify a crucial hyper-parameter, namely abstraction granu- larity , which corresponds to the size of intervals used for training the networks. We propose a gradient descent-based algorithm to refine the abstraction granularity for training a more robust neural network

Zang_Discovering_the_Real_Association_Multimodal_Causal_Reasoning_in_Video_Question_CVPR_2023

Abstract Video Question Answering (VideoQA) is challenging as it requires capturing accurate correlations between modal- ities from redundant information. Recent methods focus on the explicit challenges of the task, e.g. multimodal feature extraction, video-text alignment and fusion. Their frameworks reason the answer relying on statistical evi- dence causes, which ignores potential bias in the multi- modal data. In our work, we investigate relational structure from a causal representation perspective on multimodal data and propose a novel inference framework. For vi- sual data, question-irrelevant objects may establish simple matching associations with the answer. For textual data, the model prefers the local phrase semantics which may de- viate from the global semantics in long sentences. There- fore, to enhance the generalization of the model, we dis- cover the real association by explicitly capturing visual features that are causally related to the question seman- tics and weakening the impact of local language seman- tics on question answering. The experimental results on two large causal VideoQA datasets verify that our pro- posed framework 1) improves the accuracy of the existing VideoQA backbone, 2) demonstrates robustness on com- plex scenes and questions. The code will be released at https://github.com/Chuanqi-Zang/Discovering- the-Real-Association .

1. Introduction Video Question Answering (VideoQA) aims to un- derstand visual information and describe it in language question-answer format, which is a natural cognitive capa- bility for humans. Computer Vision (CV) and Natural Lan- guage Processing (NLP) as the base models for VideoQA have shown significant progress due to the successful appli- cation of deep learning, such as action classification [25], object detection [10], instance segmentation [31], and large- scale pre-trained language model [6,19]. These tremendous *Wei Liang is the corresponding author. Question: What will happen if the girl sprains? Answer: The girl will stop. Reason: A.There are a lot people here , and can find someone to help at any time. B.The girl can’t exercise because of a sprain and needs to rest .Question: What will happen if the power is cut off? Answer: A.[person_1] will stop singing karaoke. B.[person_1] and [person_2] both cannot work. TV , indoor singing karaoke Someone, help Reason Figure 1. Two samples in VideoQA dataset. They exhibit spurious reasoning processes of B2A [26] that rely on statistical patterns, including visually spurious object-relationship associations (top) and textually unilateral semantic representations (bottom). advances in basic applications fuel the confidence in fine- grained multimodal analysis and reasoning, not only fea- ture extraction, but also fine-grained general causality esti- mation, which is critical for a robust cognitive system. Recent VideoQA methods usually explore multimodal relational knowledge by sophisticated structured architec- ture, such as memory-augmented model [9], hierarchical model [20], topological model [18], and transformer-based model [37]. Although experiments validate their feature fusion capabilities, we find that these methods concentrate on statistical association based on multimodal data, ignor- ing the real stable association. They usually use a gener- ally constrained approach with Empirical Risk Minimiza- tion (ERM), which tends to over-rely on the co-occurrence bias of repeated objects and words in the collected observa- tional data, and bypasses the impact of complete semantics at the sentence level. This mechanism reduces the robust- ness of the model on new data, even in the test set which has similar distributions to the training set. For example, as shown in Fig. 1 (top), two people are playing ”cricket” indoors, and there are other objects in the room, including a TV that is on. If relying on the statisti- cal relationship, the model may be confused by the two im- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19027 portant visual factors of ”indoor” and ”television”, and mis- judge the concerned event, that is, ”sing karaoke”. Based on the clues provided by the question, the concerned event is related to the action that is taking place, e.g. ”two men play- ing cricket”. This requires the model to accurately judge the objects involved in the event and infer the answer. In Fig. 1 (bottom), we show the predicted deviation caused by statistical relationships in the text. When inferring the rea- sons for answer selection, existing models intensively rely on correlations between local words and video content, e.g. ”someone; help”, ignoring unreasonable inferences from other parts of the sentence, e.g. ”a lot”. From these observations, we summarize two causal chal- lenges for the VideoQA task. 1) Irrelevant objects con- founder. The visual information related to the question is usually causally related to finite objects in the video. When other objects or background are considered, they are sub- ject to data bias and become confounders, misleading the model to select the negative candidate. 2) Keywords con- founder. The semantics expressed by textual information is represented by the overall sentence. Some long sentences contain partially sensible keywords. When just focusing on local keywords semantics, the model falls into spurious causal inferences, reducing the robustness of the model. To address the above causal challenges and improve the robustness of the model, we propose a Multimodal Causal Reasoning (MCR) framework for video-text data. In this framework, causal features and confounding features are decoupled separately in visual and textual modes through two training strategies. To explicitly decouple the influence of different objects and scenes, MCR extracts fine-gained object-level appearance features and motion dynamics by spatial Region of Interest (ROI) Align [13] on global vi- sual features. Among them, causally related objects are se- lected based on the correlation between object features and question semantics. In addition to the visual feature, we also model object coordinate information, category infor- mation, and global object interaction information to pro- vide spatio-temporal relation representations for accurate causal attribute classification. For textual confounders, we adopt a strategy to reduce the impact of keywords on causal- ity. MCR relies on the correlation between word encod- ing and question-visual co-embedding to select keywords which have a crucial impact on the prediction results. These keywords provide negative representations for successive deductive answers. Therefore, we combine these keywords with other candidate answers to generate difficult negative samples to improve the recognition ability of the model. During training, visual intervention and textual interven- tion are iteratively optimized. Multimodal causal relation- ships are gradually established which improves the robust- ness and reusability of the model. We summarize our contributions as: (1) We discover twonew types of causal challenges for both visual data and tex- tual data. (2) We propose an object-level causal relationship extraction strategy to establish the real association between objects and language semantics, and a keyword broadcast- ing strategy to cut off the spurious influence of local textual information. (3) We achieve state-of-the-art performance on two latest large causal VideoQA datasets.

Yang_Video_Event_Restoration_Based_on_Keyframes_for_Video_Anomaly_Detection_CVPR_2023

Abstract Video anomaly detection (VAD) is a significant computer vision problem. Existing deep neural network (DNN) based VAD methods mostly follow the route of frame reconstruc-tion or frame prediction. However , the lack of mining andlearning of higher-level visual features and temporal con-text relationships in videos limits the further performance of these two approaches. Inspired by video codec theory,we introduce a brand-new VAD paradigm to break throughthese limitations: First, we propose a new task of video event restoration based on keyframes. Encouraging DNN toinfer missing multiple frames based on video keyframes so as to restore a video event, which can more effectively mo- tivate DNN to mine and learn potential higher-level visual features and comprehensive temporal context relationshipsin the video. To this end, we propose a novel U-shaped Swin Transformer Network with Dual Skip Connections ( USTN- DSC ) for video event restoration, where a cross-attention and a temporal upsampling residual skip connection are in-troduced to further assist in restoring complex static and dynamic motion object features in the video. In addition, we propose a simple and effective adjacent frame differ-ence loss to constrain the motion consistency of the video sequence. Extensive experiments on benchmarks demon- strate that USTN-DSC outperforms most existing methods,validating the effectiveness of our method.

1. Introduction Video anomaly detection (V AD) is a hot but challenging research topic in the field of computer vision. One of the most challenging aspects comes from the scarcity of anoma-lous samples, which hinders us from learning anomalousbehavior patterns from the samples. As a result, it is hardfor supervised methods to show their abilities, as unsuper- vised methods are by far the most widely explored direc-tion [ 1,2,4,8,10,19,25–27,31,33,37,41–43,47–51]. To †Corresponding authors.IBBB B B B PI IBBB B B B PIN o n e II IN o n eN o n eN o n eN o n eN o n eN o n e II IN o n eN o n eN o n eN o n eN o n e (a) video codec (b) video event restoration Figure 1. Video codec vs. video event restoration. Compared to video decoding based on explicit motion information, our pro- posed video event restoration task encourages DNN to automat- ically mine and learn the implied spatio-temporal relationships from several frames with key appearance and temporal transition information to restore the entire video event. determine whether an abnormal event occurs under unsu- pervised, a general approach is to model a regular patternbased on normal samples, leaving samples that are incon-sistent with this pattern as anomalies. Under the unsupervised framework, reconstruction and prediction-based approaches are two of the most represen- tative paradigms [ 10,19] for V AD in the current era of deep neural networks (DNN). The reconstruction-based ap- proaches typically use deep autoencoder (DAE) to learn to reconstruct an input frame and regard a frame with large reconstruction errors as anomalous. However, the power-ful generalization capabilities of DAE enable well recon- struction of anomalous as well [ 8]. This is attributed to the simple task of reconstructing input frames, where DAEonly memorizes low-level details rather than understand-ing high-level semantics [ 17]. The idea of the prediction- based approach assumes that anomalous events are un-predictable and then builds a model that can predict fu-ture frames according to past frames, and poor predictionsof future frames indicate the presence of anomalies [ 19]. The prediction-based approaches further consider short-term temporal relationships to model appearance and mo- tion patterns. However, due to the appearance and motion variations of adjacent frames being minimal, the predic- tor can predict the future frame better based on the past few frames, even for anomalous samples. The inherent lack of capability of these two modeling paradigms in min- ing higher-level visual features and comprehensive tempo- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14592 ral context relationships limits their further performance im- provement. In this paper, we aim to explore better methods for mining complex spatio-temporal relationships in the video.Inspired by video codec theory [ 15,18], we propose a brand-new V AD paradigm: video event restoration based on keyframes for V AD. In the video codec [ 18], three types of frames are utilized, namely I-frame, P-frame, and B-frame.I-frame contains the complete appearance information ofthe current frame, which is called a keyframe. P-frame con- tains the motion difference information from the previous frame, and B-frame contains the motion difference betweenthe previous and next frames. Based on these three types of frames containing explicit appearance and motion rela- tive relationships, the video can be decoded. Inspired by this process, our idea sprang up: It should be also theoreti-cally feasible if we give keyframes that contain only implicit relative relations between appearance and motion, and thenencourage DNN to explore and mine the potential spatio- temporal variation relationships therein to infer the missing multiple frames for video event restoration. To motivate DNN to explore and learn spatio-temporal evolutionary re-lationships in the video actively, we do not provide frameslike P-frames or B-frames that contain explicit motion in- formation as input cues. This task is extremely challengingcompared to reconstruction and prediction-based tasks, be- cause DNN must learn to infer the missing multiple framesbased on keyframes only. This requires a strong regularity and temporal correlation of the event in the video for a better restoration. On the contrary, video restoration will be poor for anomalous events that are irregular and random. Underthis assumption, our proposed keyframes-based video event restoration task can be exactly applicable to V AD. Fig. 1 compares the video codec and video event restoration task with an illustration. To perform this challenging task, we propose a novel U-shaped Swin Transformer Network with Dual Skip Connections ( USTN-DSC ) for video event restoration based on keyframes. USTN-DSC follows the classic U-Net [ 36] architecture design, where its backbone consists of multiple layers of swin transformer (ST) [ 21] blocks. Fur- thermore, to cope with the complex motion patterns in the video so as to better restore the video event, we build dual skip connections in USTN-DSC. Specifically, we introduce a cross-attention and a temporal upsampling residual skip connection to further assist in restoring complex dynamicand static motion object features in the video. In addition, to ensure that the restored video sequence has the consistencyof temporal variation with the real video sequence, we pro- pose a simple and effective adjacent frame difference (AFD) loss. Compared with the commonly used optical flow con- straint loss [ 19], AFD loss is simpler to be calculated while having comparable constraint effectiveness.The main contributions are summarized as follows: • We introduce a brand-new video anomaly detection paradigm that is to restore the video event based onkeyframes, which can more effectively mine and learn higher-level visual features and comprehensive tempo-ral context relationships in the video. • We introduce a novel model called USTN-DSC for video event restoration, where a cross-attention and atemporal upsampling residual skip connection are in- troduced to further assist in restoring complex dynamicand static motion object features in the video. • We propose a simple and effective AFD loss to con- strain the motion consistency of the video sequence. • USTN-DSC outperforms most existing methods on three benchmarks, and extensive ablation experiments demonstrate the effectiveness of USTN-DSC.

Zeng_PEFAT_Boosting_Semi-Supervised_Medical_Image_Classification_via_Pseudo-Loss_Estimation_and_CVPR_2023

Abstract Pseudo-labeling approaches have been proven beneficial for semi-supervised learning (SSL) schemes in computer vi- sion and medical imaging. Most works are dedicated to finding samples with high-confidence pseudo-labels from the perspective of model predicted probability. Whereas this way may lead to the inclusion of incorrectly pseudo- labeled data if the threshold is not carefully adjusted. In addition, low-confidence probability samples are frequently disregarded and not employed to their full potential. In this paper, we propose a novel Pseudo-loss Estimation andFeature Adversarial Training semi-supervised frame- work, termed as PEFAT, to boost the performance of multi- class and multi-label medical image classification from the point of loss distribution modeling and adversarial train- ing. Specifically, we develop a trustworthy data selec- tion scheme to split a high-quality pseudo-labeled set, in- spired by the dividable pseudo-loss assumption that clean data tend to show lower loss while noise data is the oppo- site. Instead of directly discarding these samples with low- quality pseudo-labels, we present a novel regularization approach to learn discriminate information from them via injecting adversarial noises at the feature-level to smooth the decision boundary. Experimental results on three med- ical and two natural image benchmarks validate that our PEFAT can achieve a promising performance and surpass other state-of-the-art methods. The code is available at https://github.com/maxwell0027/PEFAT.

1. Introduction Deep learning has achieved remarkable success in vari- ous computer vision tasks [4–6,13,15,19]. This success has ∗Equal contribution.†Y . Xia is the corresponding author. This work was supported in part by the National Natural Science Foundation of China under Grants 62171377, and in part by the Key Research and Development Program of Shaanxi Province, China, under Grant 2022GY-084. (a) Illustration of traditional SSL methods (top) and our method (bottom). (b) Probability distribution of labeled data (left) and validation data (right), when using the warm-upped model on ISIC2018 dataset. Figure 1. (a) shows the main difference between our method and other SSL methods, our method selects high-quality pseudo- labeled data by pseudo-loss estimation, and also injects feature- level adversarial noises for better unlabeled data mining; (b) in- dicates the phenomenon that clean pseudo-labeled set is hard to collect when using probability-based threshold, which is mainly attributed to the over-confident prediction. also made practical applications more accessible, including medical image analysis (MIA) [10,21,30,31,35,39]. How- ever, unlike computer vision, annotating a large-scale med- ical image dataset requires expert knowledge and is time- consuming and costly. Alternatively, unlabeled data can be collected from clinical sites in a more available way, thereby mitigating the cost of data annotation by leveraging these unlabeled data. Semi-supervised Learning (SSL) has drawn a lot of at- tention due to its superior performance, by only leveraging limited labeled data and a vast number of unlabeled data. Under the SSL setting, it is critical to mine adequate infor- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15671 mation from unlabeled data. In the existing SSL methods, pseudo-labeling [14,22,29,37,38] and consistency regular- ization [12, 24, 25] are the mainstream. The former focuses on finding confident pseudo-labels for re-training, and the latter aims to improve the robustness of the model by keep- ing one logical distribution similar to the other. However, most SSL methods encounter two issues. First, unreliable pseudo-labels are a problem with the threshold- selecting data method based on predicted probability as it often introduces numerous incorrect pseudo-labels due to confirmation bias. As illustrated in Figure 1b, unlabeled data with both correct and incorrect pseudo-labels follow similar probability distributions. Second, informative unla- beled samples are underutilized as unselected data with low probabilities typically cluster around the decision boundary. Recent studies [1, 17] have found that neural networks tend to fit clean data first and then memorize noise data during training, resulting in lower loss for clean data and higher loss for noise data in early stages of training. Furthermore, some works [9, 25] have investigated the effects of adver- sarial training under semi-supervised settings, which show potential for learning from low-quality pseudo-labeled data. All these findings pave the way towards solving the afore- mentioned problems. In this paper, we propose a novel SSL method called Pseudo-loss Estimation and Feature Adversarial Training (PEFAT) for multi-class and multi-label medical image classification. First, we introduce a new estimation scheme for reliable pseudo-labeled data selection from the perspec- tive of pseudo-loss distribution. It is motivated by our argu- ment that there is a dividable loss distribution between cor- rect and incorrect pseudo-labeled data. Specifically, we first warm up the model on training data with contrastive learn- ing, in order to learn unbiased representation. Then we set up a two-component Gaussian Mixture Model (GMM) [28] to learn prior loss distribution on labeled data. In the procedure of pseudo-labeled data selection, we feed cross pseudo-loss to the fitted GMM and obtain the trustworthy pseudo-labeled data with posterior probability. Second, we propose a feature adversarial training (FAT) strategy that in- jects adversarial noises in the feature-level to smooth the de- cision boundary, aiming at for further utilizing the rest uns- elected but informative data. Although FAT is originally de- signed for the rest data, it can also be applied to the selected pseudo-labeled data. Based on the technics above, our PE- FAT successfully boosts the classification performance in MIA from the point of trustworthy pseudo-labeled data se- lection and adversarial consistency regularization. To summarize, our main contributions are three-fold. (1) Different from previous works that select pseudo-labeled data with a probability threshold, we present a new se- lection approach from the perspective of the loss distribu- tion, which exhibits superior ability in high-quality pseudo-labeled data collection. (2) We propose a new adversarial regularization strategy to fully leverage the rest unlabeled but informative data, which benefits the model in decision boundary smoothing and better representation learning. (3) Extensive experimental results on three public medical im- age datasets and two natural image datasets demonstrate the superiority of the proposed PEFAT, which significantly sur- passes other advanced SSL methods.

Yang_K3DN_Disparity-Aware_Kernel_Estimation_for_Dual-Pixel_Defocus_Deblurring_CVPR_2023

Abstract The dual-pixel (DP) sensor captures a two-view image pair in a single snapshot by splitting each pixel in half. The disparity occurs in defocus blurred regions between the two views of the DP pair, while the in-focus sharp regions have zero disparity. This motivates us to propose a K3DN framework for DP pair deblurring, and it has three modules: i) a disparity-aware deblur module. It estimates a disparity feature map, which is used to query a trainable kernel set to estimate a blur kernel that best describes the spatially-varying blur. The kernel is constrained to be symmetrical per the DP formulation. A simple Fourier transform is performed for deblurring that follows the blur model; ii) a reblurring regularization module. It reuses the blur kernel, performs a simple convolution for reblurring, and regularizes the estimated kernel and disparity feature unsupervisedly, in the training stage; iii) a sharp region preservation module. It identifies in-focus regions that correspond to areas with zero disparity between DP images, aims to avoid the introduction of noises during the deblurring process, and improves image restoration performance. Experiments on four standard DP datasets show that the proposed K3DN outperforms state-of-the-art methods, with fewer parameters and flops at the same time.

1. Introduction Dual-pixel (DP) sensors are widely employed in digital single-lens reflex cameras (DSLR) and smartphone cam- eras. Each pixel of a DP sensor is divided into two photodi- odes to provide two sub-views of a scene, namely, left and right views, to assist in auto-focusing. Recently, researchers have used the two-view DP pair to benefit several computer vision tasks, especially defocus deblurring [ 1,2,24]. Specif- ically, blurred pixels in the left and right view of the DP pair exhibit an amount of shift, termed DP disparity, which pro- vides information for the blur kernel estimation (the key for defocus deblurring [ 14,16,19]). †Corresponding author.⇤Equal contribution.010203024252627 KPAC(21)IFAN(21) DEMENet(19)DPDNet(20)RDPD(21)DDDNet(21)BAMBNet(22)Restormer(22)OursOursDeepRFT(21)DRBNet (22) Params (M)PSNR (dB)01K2K3K4K5KOursOursDeepRFT(21)DRBNet (22)IFAN(21)KPAC(21)RDPD(21)BAMBNet(22)Restormer(22) DDDNet(21)DPDNet(20) DEMENet(19)Flops (G)Figure 1. Comparison of our method and state-of-the-art methods on the real DPD-blur dataset [ 1]. The Blue andredcycles sep- arately denote our tiny and lightweight models. We show PSNRs (dB) with respect to model parameters (M) and Flops (G). The numbers in the brackets beside methods denote publication years. Best viewed in color on the screen. Existing DP-based defocus deblurring methods are gen- erally divided into two categories: end-to-end-based and disparity (defocus map)-based methods. End-to-end based methods [ 1,2,11,17,19,24] directly restore an all-in-focus image from a defocus blurred DP pair without consider- ing DP domain knowledge for network design. Therefore, it is challenging for these methods to tackle spatially-varying and large blur. Disparity-based methods [ 10,11,15,23] esti- mate the DP disparity (or defocus map) to aid the restoration of an all-in-focus image. However, these methods either re- quire extra data (ground-truth DP disparity as supervision) [15], neglect DP disparity knowledge when warping [ 10], heavily depend on the number of network branches [ 11], or rely on pre-calibrated kernel sets and additional priors [ 23] for training. Our method, K3DN, belongs to the disparity-based group, and it has high performance on real DP datasets (see Fig.1for an example). K3DN follows the mathematic in- dications of DP and blur models, while not requiring pre- calibrated kernel sets and extra data. K3DN consists of three modules: i) a disparity-aware deblur module; ii) a re- blurring regularization module; iii) a sharp region preserva- tion module. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13263 The first component is a disparity-aware deblur module that estimates spatially-varying kernels for deblurring. We define a trainable kernel set, and kernels in the set can model diverse blur patterns. Each kernel satisfies the horizontally symmetric constraint [ 2,16], following the DP image for- mulation. Given the input DP image pair, we first estimate a disparity feature map, and then use the disparity feature map to query the kernel set to estimate a DP blur kernel that best describes the spatially-varying blur. With the blur ker- nel, we simply perform a Fourier transform for deblurring that follows the blur model. The second component we developed is a reblurring reg- ularization module. It reuses the DP blur kernel, takes an all-in-focus image as input, and performs a simple convolu- tion to generate a blurred image. By enforcing the similarity between the blurred image and the input DP image, we reg- ularize the proposed deblur module. Our third component is a sharp region preservation mod- ule. Not all regions of DP pairs are blurred. We observe that previous works neglect to preserve pixel values or features from in-focus sharp regions. By mining (based on the DP model) similarities between features before and after our deblur module, we preserve features from all-in-focus re- gions and improve feature quality, leading to better image restoration performance. Please note that all intermediate parts ( e.g., the train- able kernel set and disparity estimator) do not require ex- tra ground-truth information ( e.g., pre-calibrated kernel sets and disparity) for training. We only need ground-truth all- in-focus images for end-to-end training. Our contributions are summarized below: •A simple K3DN framework for DP image pair defocus deblurring; •A disparity-aware deblur module. It has a trainable kernel set, and estimates a disparity feature map to query the set to estimate a DP blur kernel that best de- scribes the spatial varying blur; •A reblurring regularization module. It follows the DP model, reuses the blur kernel from the deblur module, and regularizes the deblur module; •A sharp region preservation module. It mines in-focus regions of DP pairs, improves feature quality from in- focus regions, resulting in better restoration perfor- mance. Experimentally, we validate the proposed K3DN on four standard benchmark datasets [ 1,2,15,16], showing state-of- the-art performance in image restoration and model size.

Zhang_Text-Visual_Prompting_for_Efficient_2D_Temporal_Video_Grounding_CVPR_2023

Abstract In this paper, we study the problem of temporal video grounding ( TVG ), which aims to predict the starting/ending time points of moments described by a text sentence within a long untrimmed video. Benefiting from fine-grained 3D visual features, the TVG techniques have achieved remark- able progress in recent years. However, the high complexity of 3D convolutional neural networks (CNNs) makes extract- ing dense 3D visual features time-consuming, which calls for intensive memory and computing resources. Towards efficient TVG, we propose a novel text-visual prompting (TVP ) framework, which incorporates optimized perturba- tion patterns (that we call ‘prompts’) into both visual in- puts and textual features of a TVG model. In sharp con- trast to 3D CNNs, we show that TVP allows us to effec- tively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of cross- modal feature fusion using only low-complexity sparse 2D visual features. Further, we propose a Temporal- Distance IoU (TDIoU ) loss for efficient learning of TVG. Experi- ments on two benchmark datasets, Charades-STA and Ac- tivityNet Captions datasets, empirically show that the pro- posed TVP significantly boosts the performance of 2D TVG (e.g., 9.79% improvement on Charades-STA and 30.77% improvement on ActivityNet Captions) and achieves 5×in- ference acceleration over TVG using 3D visual features. Codes are available at Open.Intel .

1. Introduction In recent years, we have witnessed great progress on temporal video grounding ( TVG ) [30, 74]. One key to this success comes from the fine-grained dense 3D vi- sual features extracted by 3D convolutional neural networks (CNNs) ( e.g., C3D [56] and I3D [3]) since TVG tasks de- mand spatial-temporal context to locate the temporal inter- val of the moments described by the text query. However, due to the high cost of the dense 3D feature extraction, most existing TVG models only take these 3D visual features ex- Figure 1. The architecture and performance comparison among TVG methods: a)3D TVG methods [14, 16, 18, 34, 43, 60–62, 64, 67, 69, 71, 73], b)2D TVG methods [1, 7], and c)TVP-based 2D TVG (Ours), d)over- all performance comparison. Ours is the most efficient (least inference time) and achieves competitive performance compared to 3D TVG meth- ods. In contrast to existing TVG methods, which utilize dense video fea- tures extracted by non-trainable offline 3D CNNs and textual features, our proposed framework utilizes a trainable 2D CNN as the vision encoder to extract features from sparsely-sampled video frames with a universal set of frame-aware visual prompts and adds text prompts in textual feature space for end-to-end regression-based modeling. tracted by offline 3D CNNs as inputs instead of co-training during TVG model training. Although models using 3D visual features (that we call ‘ 3D methods or models ’) outperform these using the 2D features (that we call ‘ 2D methods or models ’), a unique advantage of 2D methods is that extracting 2D visual features can significantly reduce the cost in TVG tasks [14, 15, 30, 34, 61, 62, 69, 74, 75]. An efficient and lightweight solution with reasonable performance is also demanded in computer vision, NLP, and video-language tasks [19, 23, 38, 41, 68, 76–80]. As discussed above, the methods employing 3D video features are challenging to be This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14794 employed in practical applications. It thus has significant practical and economic value to develop compact 2D solu- tions for TVG tasks. In this paper, we ask: How to advance 2D TVG methods so as to achieve comparable results to 3D TVG methods? To address this problem, we propose a novel text-visual prompting ( TVP ) framework for training TVG models us- ing 2D visual features. As shown in Fig. 1 , for existing 2D TVG and 3D TVG methods, they all utilize offline pre- trained vision encoders and language encoders to perform feature extraction. In contrast, our proposed TVP frame- work is end-to-end trainable. Furthermore, benefiting from text-visual prompting and cross-modal pretraining on large- scale image-text datasets, our proposed framework could achieve comparable performance to 3D TVG methods with significant inference time acceleration. Conventionally, TVG methods consist of three stages: ①extracting feature from visual and text inputs; ②multi- modal feature fusion; ③cross-modal modelling. In contrast to conventional methods, TVP incorporates optimized input perturbation patterns (that we call ‘ prompts ’) into both vi- sual inputs and textual features of a TVG model. We apply trainable parameters in the textual features as text prompts and develop a universal set of frame-aware patterns as visual prompts. Specially, we sample a fixed number of frames from a video and optimize text prompts for the input query sentence and a set of visual prompts for frames with differ- ent temporal locations during training. During testing, the same set of optimized visual prompts and textual prompts are applied to all test-time videos. We refer readers to Fig. 2 for illustrations of visual prompts and text prompts intro- duced. To the best of our knowledge, our work makes the first attempt to utilize prompt learning to successfully im- prove the performance of regression-based TVG tasks using 2D visual features. Compared to 3D CNNs, 2D CNNs loses spatiotempo- ral information of the video during feature extraction. In- spired by the success of transformers on the vision-language tasks [9, 22, 35, 44, 47, 54, 55] and the recent application of prompt learning to transformers in both vision and language domains [2,25,27,32,37,40], we choose transformer as our base TVG model and propose to utilize prompts to compen- sate for the lack of spatiotemporal information in 2D visual features. Furthermore, we develop a Temporal- Distance IoU ( TDIoU ) loss for training our proposed framework. There are two aspects that distinguish our proposed frame- work from existing works. First , our proposed framework is designed to boost the performance of the regression-based TVG methods utilizing 2D CNNs as the vision encoder, not for transfer learning [2, 21, 26] Second , our proposed framework utilizes 2D CNN to extract visual features from (a) Text Prompts (b) Frame-aware Visual Prompts Figure 2. Text-visual prompting illustration. (a) Text prompts are directly applied in the feature space. (b) A set of visual prompts are applied to video frames in order. sparsely-sampled video frames, which requires less mem- ory and is easier to be applied in practical applications com- pared to 3D methods [34,60–62,69,75], especially for long videos. Furthermore, thanks to the compact 2D CNN as the vision encoder, our proposed framework could imple- ment the language encoder and visual encoder co-training for better multimodal feature fusion. In summary, the con- tributions of this work are unfolded below: • We propose an effective and efficient framework to train 2D TVG models, in which we leverage TVP (text-visual prompting) to improve the utility of sparse 2D visual features without resorting to costly 3D fea- tures. To the best of our knowledge, it is the first work to expand the application of prompt learning for re- solving TVG problems. Our method outperforms all of 2D methods and achieves competitive performance to 3D TVG methods. • Technology-wise, we integrate visual prompt with text prompt to co-improve the effectiveness of 2D visual features. On top of that, we propose TDIoU (temporal- distance IoU)-based prompt-model co-training method to obtain high-accuracy 2D TVG models. • Experiment-wise, we show the empirical success of our proposal to boost the performance of 2D TVG on Charades-STA and ActivityNet Captions datasets, e.g., 9.79% improvement in Charades-STA, and 30.77% in ActivityNet-Captions together with 5×inference time acceleration over 3D TVG methods.

Ye_DistilPose_Tokenized_Pose_Regression_With_Heatmap_Distillation_CVPR_2023

Abstract In the field of human pose estimation, regression-based methods have been dominated in terms of speed, while heatmap-based methods are far ahead in terms of per- formance. How to take advantage of both schemes re-mains a challenging problem. In this paper , we proposea novel human pose estimation framework termed Distil-Pose, which bridges the gaps between heatmap-based andregression-based methods. Specifically, DistilPose maxi-mizes the transfer of knowledge from the teacher model(heatmap-based) to the student model (regression-based)through Token-distilling Encoder (TDE) and SimulatedHeatmaps. TDE aligns the feature spaces of heatmap-based and regression-based models by introducing tok-enization, while Simulated Heatmaps transfer explicit guid-ance (distribution and confidence) from teacher heatmapsinto student models. Extensive experiments show that theproposed DistilPose can significantly improve the perfor-mance of the regression-based models while maintaining ef-ficiency. Specifically, on the MSCOCO validation dataset,DistilPose-S obtains 71.6% mAP with 5.36M parameters,2.38 GFLOPs, and 40.2 FPS, which saves 12.95 ×, 7.16× computational cost and is 4.9 ×faster than its teacher model with only 0.9 points performance drop. Furthermore,DistilPose-L obtains 74.4% mAP on MSCOCO validationdataset, achieving a new state-of-the-art among predomi-nant regression-based models. Code will be available athttps://github.com/yshMars/DistilPose .

1. Introduction 2D Human Pose Estimation (HPE) aims to detect the anatomical joints of a human in a given image to estimate the poses. HPE is typically used as a preprocessing mod-ule that participates in many downstream tasks, such as ac-tivity recognition [ 28,31], human motion analysis [ 1], mo- ∗Equal contribution. This work was done when Suhang Ye was an intern at Tencent Y outu Lab. †Corresponding author: [email protected] 0 1 02 03 04 05 06 07 08 0 90AP(%) ParameterDistilPose-S(ours)DistilPose-L(ours) RLE-Res50[ 7]RLE-HRNet-W48[ 7] PRTR-HRNet-W32[ 8] PRTR-Res50[ 8]SimpleBaseline-Res50[ 28] SimpleBaseline-Res152[ 28] 0 51 0 1 5 20 GFLOPsDiameter Figure 1. Comparisons between the SOTA methods and the proposed DistilPose on MSCOCO valdataset. Red circles at the upper left corner denote DistilPose. DistilPose outperforms SOTAmodels in terms of accuracy (AP), Parameter and computationalcost (GFLOPs). tion capture [ 16], etc. Previous studies on 2D HPE can be mainly divided into two mainstreams: heatmap-based and regression-based methods. Regression-based methods have significant advantages in speed and are well-suited formobile devices. However, the insufficient accuracy of re-gression models will affect the performance of downstreamtasks. In contrast, heatmap-based methods can explicitlylearn spatial information by estimating likelihood heatmaps,resulting in high accuracy on HPE tasks. But the estimationof likelihood heatmaps requires exceptionally high compu-tational cost, which leads to slow preprocessing operations.Thus, how to take advantages of both heatmap-based andregression-based methods remains a challenging problem. One possible way to solve the above problem is to trans- fer the knowledge from heatmap-based to regression-basedmodels [ 8,23]. However, due to the different output spaces of regression models and heatmap models (the former isa vector, and the latter is a heatmap), transferring knowl- edge between heatmaps and vectors faces the following twoproblems: (1) The regression head usually vectorizes the feature map output by the backbone. And much spatialinformation will be lost through Global Average Pooling This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2163 (GAP) or Flatten operation. Thus, previous work failed to transfer heatmap knowledge to regression models fully.(2) Compared to the coordinate regression, heatmaps natu-rally contain shape, position, and gradient information [ 3]. Due to the lack of explicit guidance for such information,regression-based methods are more difficult to learn theimplicit relationship between features and keypoints thanheatmap-based methods. In this paper, we propose a novel human pose esti- mation framework, DistilPose, which learns to transferthe heatmap-based knowledge from a teacher model to aregression-based student model. DistilPose mainly includesthe following two components: (1) A knowledge-transferring module called Token- distilling Encoder (TDE) is designed to align the featurespaces of heatmap-based and regression-based models byintroducing tokenization, which consists of a series of trans- former encoders. TDE can capture the relationship between keypoints and feature maps/other keypoints [ 10,32]. (2) We propose to simulate heatmaps to obtain the heatmap infor-mation for regression-based students explicitly. The result-ing Simulated Heatmaps provide two explicit guidelines,including each keypoint’s 2D distribution and confidence. Note that the proposed Simulated Heatmaps can be insertedbetween any heatmap-based and regression-based methodsto transfer heatmap knowledge to regression models. DistilPose achieves comparable performance to heatmap-based models with less computational cost andsurpasses the state-of-the-art (SOTA) regression-basedmethods. Specifically, on the MSCOCO validation dataset,DistilPose-S achieves 71.6% mAP with 5.36M parameters, 2.38GFLOPs and 40.3FPS. DistilPose-L achieves 74.4% mAP with 21.27M parameters and 10.33GFLOPs, which outperforms its heatmap-based teacher model in perfor-mance, parameters and computational cost. In summary,DistilPose significantly reduces the computation whileachieving competitive accuracy, bringing advantages fromboth heatmap-based and regression-based schemes. Asshown in Figure 1, DistilPose outperforms previous SOTA regression-based methods, such as RLE [ 8] and PRTR [ 9] with fewer parameters and GFLOPs. Our contributions are summarized as follows: • We propose a novel human pose estimation frame- work, DistilPose, which is the first work to trans-fer knowledge between heatmap-based and regression-based models losslessly. • We introduce a novel Token-distilling Encoder (TDE) to take advantage of both heatmap-based andregression-based models. With the proposed TDE, thegap between the output space of heatmaps and coordi-nate vectors can be facilitated in a tokenized manner. • We propose Simulated Heatmaps to model explicitheatmap information, including 2D keypoint distribu- tions and keypoint confidences. With the aid of Sim-ulated Heatmaps, we can transform the regression-based HPE task into a more straightforward learn-ing task that fully exploits local information. Simu-lated Heatmaps can be applied to any heatmap-basedand regression-based models for transferring heatmapknowledge to regression models.

Yang_NeRFVS_Neural_Radiance_Fields_for_Free_View_Synthesis_via_Geometry_CVPR_2023

Abstract We present NeRFVS, a novel neural radiance fields (NeRF) based method to enable free navigation in a room. NeRF achieves impressive performance in rendering im- ages for novel views similar to the input views while suffer- ing for novel views that are significantly different from the training views. To address this issue, we utilize the holis- tic priors, including pseudo depth maps and view coverage information, from neural reconstruction to guide the learn- ing of implicit neural representations of 3D indoor scenes. Concretely, an off-the-shelf neural reconstruction method is leveraged to generate a geometry scaffold. Then, two loss functions based on the holistic priors are proposed to im- prove the learning of NeRF: 1) A robust depth loss that can tolerate the error of the pseudo depth map to guide the ge- *Work done during an internship at Huawei Noah’s Ark Lab. †Corresponding Author.ometry learning of NeRF; 2) A variance loss to regularize the variance of implicit neural representations to reduce the geometry and color ambiguity in the learning procedure. These two loss functions are modulated during NeRF op- timization according to the view coverage information to reduce the negative influence brought by the view coverage imbalance. Extensive results demonstrate that our NeRFVS outperforms state-of-the-art view synthesis methods quanti- tatively and qualitatively on indoor scenes, achieving high- fidelity free navigation results.

1. Introduction Reconstructing an indoor scene from a collection of im- ages and enabling users to navigate inside it freely is a core component for many downstream applications. It is the most challenging novel-view-synthesis (NVS) task, since it requires high fidelity synthesis from any view , including not This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16549 only views similar to the training views (interpolation), but also views that are significantly different from input views (extrapolation), as shown in Fig. 1. To clarify its differ- ence to other NVS tasks, we term it as free-view-synthesis (FVS). The difficulties of FVS lie in not only the common obstacles in scene reconstruction, including low-texture ar- eas, complex scene geometry, and illumination change, but also view imbalance, e.g., casual photos usually cover the scene unevenly, with hundreds of frames for one table and a few for the floor and wall, as shown in Fig. 3. Recently, NeRF has emerged as a promising technique for 3D reconstruction and novel view synthesis. Although NeRF can achieve impressive interpolation performance, its extrapolation ability is relatively poor [35], especially for low-texture and few-shot regions. In contrast, some neural reconstruction methods can recover the holistic scene ge- ometry successfully with various priors [9,22,25,34], while the synthesized images from these methods contain plenty of artifacts and are over-smoothed. Inspired by the phe- nomena, we demonstrate that equipping the NeRF with the scene priors of the geometry captured from neural recon- struction is a potential solution for indoor FVS. Extending NeRF to enable FVS with geometry from neural reconstruction methods is a non-trivial task with two main challenges. 1) Depth error . The reconstructed ge- ometry might contain some failures, including holes, depth shifting, and floaters. The optimization of NeRF relies on the multi-view color consistency, while these failures may conflict with the multi-view color consistency, resulting in artifacts. 2) Distribution ambiguity . The depth from NeRF is a weighted sum of sampling distance. Merely supervis- ing the depth expectation leads to arbitrary radiance distri- bution, especially in low-texture and few-shot regions. This ambiguous distribution leads to floaters and blur among ren- dered images, as shown in Fig. 5. In this paper, we propose a novel method which exploits the holistic priors, including pseudo depth maps and view coverage information, outputted from a geometry scaffold to guide NeRF optimization, significantly improving qual- ity on low-texture and few-shot regions. Specifically, to ad- dress the depth error , we propose a robust depth loss that can tolerate the error from the pseudo depth maps, reduc- ing the negative impact of inaccura

Xu_NeuralLift-360_Lifting_an_In-the-Wild_2D_Photo_to_a_3D_Object_CVPR_2023

Abstract Virtual reality and augmented reality (XR) bring increas- ing demand for 3D content generation. However, creating high-quality 3D content requires tedious work from a hu- man expert. In this work, we study the challenging task of lifting a single image to a 3D object and, for the first time, demonstrate the ability to generate a plausible 3D object with 360◦views that corresponds well with the given ref- erence image. By conditioning on the reference image, our model can fulfill the everlasting curiosity for synthesizing novel views of objects from images. Our technique sheds light on a promising direction of easing the workflows for 3D artists and XR designers. We propose a novel frame- work, dubbed NeuralLift-360 , that utilizes a depth-aware neural radiance representation (NeRF) and learns to craft the scene guided by denoising diffusion models. By intro- ducing a ranking loss, our NeuralLift-360 can be guided with rough depth estimation in the wild. We also adopt a CLIP-guided sampling strategy for the diffusion prior to provide coherent guidance. Extensive experiments demon- strate that our NeuralLift-360 significantly outperforms ex- isting state-of-the-art baselines. Project page: https: //vita-group.github.io/NeuralLift-360/

1. Introduction Creating 3D content has been a long-standing problem in computer vision. This problem enables various applica- tions in game studios, home decoration, virtual reality, and augmented reality. Over the past few decades, the manual task has dominated real scenarios, which requires tedious professional expert modeling. Modern artists rely on spe- cial software tools (e.g., Blender, Maya3D, 3DS Max, etc.) and time-demanding manual adjustments to realize imag- inations and transform them into virtual objects. Mean- while, automatic 3D content creation pipelines serve as ef- fective tools to facilitate human efforts. These pipelines typ- ically capture hundreds of images and leverage multi-view stereo [71] to quickly model fine-grained virtual landscapes. More recently, researchers have started aiming at a moreambitious goal, to create a 3D object from a single im- age [9, 10, 12, 21, 27, 51, 66, 68]. This enables broad ap- plications since it greatly reduces the prerequisite to a min- imal request. Existing approaches can be mainly catego- rized into two major directions. One line of work utilizes learning priors from large-scale datasets with multi-view images [10,21,51,66,68]. These approaches usually learn a conditional neural network to predict 3D information based on input images. However, due to their poor generalization ability, drastic performance degradation is observed when the testing image is out-of-distribution. Another direction constructs the pipeline on top of the depth estimation techniques [65]. Under the guidance of monocular depth estimation networks [43, 44], the 2D im- age is firstly back-projected to a 3D data format (e.g., point cloud or multi-plane image) and then re-projected into novel views. After that, advanced image inpainting techniques are then adopted to fill missing holes [52] produced during the projection. However, most of them can be highly af- fected by the quality of the estimated depth maps. Though LeRes [73] attempted to rectify the predicted depth by refin- ing the projected 3D point cloud, their results do not gener- alize well to an arbitrary image in the wild. Overall, the aforementioned approaches are either adopted in limited scenarios (e.g. face or toy examples [4, 75]), or only pro- duce limited viewing directions [8, 70] when being applied to scenes in the wild. Different from all these approaches, we focus on a more challenging task and for the first time, show promising results by lifting a single in-the-wild im- age into a 3D object with 360◦novel views . Attracted by the dramatic progress of neural volumet- ric rendering on 3D reconstruction tasks, we consider building our framework based on Neural Radiance Fields (NeRFs) [35]. The original NeRF takes hundreds of training views and their camera poses as inputs to learn an implicit representation. Subsequent models dedicate tremendous ef- forts [8, 19, 70, 75] to apply NeRF to sparse training views. The most similar work [70] to our method proposes to opti- mize a NeRF using only a single image and its correspond- ing depth map. However, it renders limited views from a small range of angles, and the prerequisite of a high-quality This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4479 depth map largely constrains its practical usage. To address the above issues, we propose a novel frame- work, coined as NeuralLift-360 , which aims to ease the creation of 3D assets by building a bridge to convert diverse in-the-wild 2D photos to sophisticated 3D contents in 360◦ views and enable its automation. The major challenge in our work is that the content on the back side is hidden and hard to hallucinate. To tackle these hurdles, we consider the diffusion priors together with the monocular depth esti- mator as the cues for hallucination. Modern diffusion mod- els [41,48,50] are trained on a massive dataset (e.g., 5B text- to-image pairs [24]). During inference time, they can gen- erate impressive photorealistic photos based on simple text inputs. By adopting these learning priors with CLIP [40] guidance, NeuralLift-360 can generate plausible 3D consis- tent instances that correspond well to the given reference image while only requiring minimal additional input, the correct user prompts. Moreover, rather than simply taking the depth map from the pre-trained depth estimator as ge- ometry supervision, we propose to use the relative ranking information from the rough depth during the training pro- cess. This simple strategy is observed to robustly mitigate geometry errors for depth estimations in the wild. Our contributions can be summarized as follows, • Given a single image in the wild, we demonstrate promising results of them being lifted to 3D. We use NeRF as an effective scene representation and inte- grate prior knowledge from the diffusion model. • We propose a CLIP-guided sampling strategy that ef- fectively marries the prior knowledge from the diffu- sion model with the reference image. • When the reference image is hard to describe exactly, we finetune the diffusion model on the single image while maintaining its ability to generate diverse con- tents to guide the NeRF training. • We introduce scale-invariant depth supervision that uses the ranking information. This design alleviates the need for accurate multi-view consistent depth esti- mation and broadens the application of our algorithms.

Zhang_Implicit_Surface_Contrastive_Clustering_for_LiDAR_Point_Clouds_CVPR_2023

Abstract Self-supervised pretraining on large unlabeled datasets has shown tremendous success in improving the task per- formance of many 2D and small scale 3D computer vi- sion tasks. However, the popular pretraining approaches have not been impactfully applied to outdoor LiDAR point cloud perception due to the latter’s scene complexity and wide range. We propose a new self-supervised pretrain- ing method ISCC with two novel pretext tasks for LiDAR point clouds. The first task uncovers semantic information by sorting local groups of points in the scene into a glob- ally consistent set of semantically meaningful clusters using contrastive learning, complemented by a second task which reasons about precise surfaces of various parts of the scene through implicit surface reconstruction to learn geomet- ric structures. We demonstrate their effectiveness through transfer learning on 3D object detection and semantic seg- mentation in real world LiDAR scenes. We further design an unsupervised semantic grouping task to show that our approach learns highly semantically meaningful features.

1. Introduction Robust and reliable 3D LiDAR perception is critical for autonomous driving systems. Unlike images, LiDAR pro- vides unambiguous measurements of the vehicle’s 3D en- vironment. A plethora of perception models have arisen in recent years to enable a variety of scene understanding tasks using LiDAR input, such as object detection and semantic segmentation. However, training these models generally re- lies on a large quantity of human annotated data, which is tedious and expensive to produce. Recently, self-supervised learning has attracted signifi- cant research attention [5, 7,16–18], as it has the potential to increase performance on downstream tasks with limited quantities of annotated data in the image domain. However, self-supervised learning has shown less impact for outdoor *Work done at AWS AI. Figure 1. With frame tas input, we 1) apply contrastive cluster learning to reason about unsupervised semantic clustering and also enforce feature consistency across different views, and 2) conduct implicit surface reconstruction in randomly cropped out regions. 3D point clouds. A core difficulty stems from the rela- tive difficulty in designing appropriate pretext tasks used for self-supervision. In the image domain, the ImageNet dataset [35] provides millions of canonical images of ev- eryday objects, allowing for straightforward manipulations to generate pretext tasks that lead to strong object-centric or semantic group-centric feature learning. While large-scale unlabeled outdoor LiDAR datasets are relatively easy to col- lect, the data samples exhibit a high level of scene complex- ity, sparsity of measurements, and heavy dependency on the observer positioning and sensor type. These factors pose great challenges for creating useful pretext tasks. Recent works have proposed 3D-specific self-supervised learning, starting with scene-level contrastive learning [44, 49], followed by the work of [30] and [47], which use finer, region-level granularity to better encode individual compo- nents of large scale LiDAR point clouds. However, these techniques do not explicitly make use of regularities in 3D shapes and surfaces. In our work, we propose ISCC (Implicit Surface Con- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21716 trastive Clustering) which consists of two new pretext tasks to automatically learn semantically meaningful feature ex- traction without annotations for LiDAR point clouds. The first task focuses on learning semantic information by sort- ing local groups of points in the scene into a globally consis- tent set of semantically meaningful clusters using the con- trastive learning setup [5]. This is augmented with a second task which reasons about precise surfaces of various parts of the scene through implicit surface reconstruction to learn geometric regularities. A high level overview is found in Figure 1. Furthermore, we showcase a novel procedure to generate training signals for implicit surface reasoning in the absence of dense 3D surface meshes which are difficult to obtain for large scale LiDAR point clouds. Using the large real world KITTI [13] and Waymo [37] datasets, we show that our approach is superior to re- lated state-of-the-art self-supervised learning techniques in downstream finetuning performance in semantic segmenta- tion and object detection. For example, we see a 72% gain in segmentation performance on SemanticKITTI versus the state-of-the-art [48] when fine-tuned with 1% of the anno- tations, and exceeds the accuracy achieved by using twice the annotations with random initialized weights. As well, we analyze the semantic consistency of the learned features through a new unsupervised semantic grouping task, and show that our learned features are able to form semantic groups even in the absence of supervised fine-tuning.

Xu_UniDexGrasp_Universal_Robotic_Dexterous_Grasping_via_Learning_Diverse_Proposal_Generation_CVPR_2023

Abstract In this work, we tackle the problem of learning universal robotic dexterous grasping from a point cloud observation under a table-top setting. The goal is to grasp and lift up ob- jects in high-quality and diverse ways and generalize across hundreds of categories and even the unseen. Inspired by successful pipelines used in parallel gripper grasping, we split the task into two stages: 1) grasp proposal (pose) gen- eration and 2) goal-conditioned grasp execution. For the first stage, we propose a novel probabilistic model of grasp pose conditioned on the point cloud observation that fac- torizes rotation from translation and articulation. Trained *Equal contribution. †Corresponding author.on our synthesized large-scale dexterous grasp dataset, this model enables us to sample diverse and high-quality dex- terous grasp poses for the object point cloud. For the sec- ond stage, we propose to replace the motion planning used in parallel gripper grasping with a goal-conditioned grasp policy, due to the complexity involved in dexterous grasp- ing execution. Note that it is very challenging to learn this highly generalizable grasp policy that only takes re- alistic inputs without oracle states. We thus propose sev- eral important innovations, including state canonicaliza- tion, object curriculum, and teacher-student distillation. In- tegrating the two stages, our final pipeline becomes the first to achieve universal generalization for dexterous grasping, demonstrating an average success rate of more than 60% on thousands of object instances, which significantly out- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4737 performs all baselines, meanwhile showing only a minimal generalization gap.

1. Introduction Robotic grasping is a fundamental capability for an agent to interact with the environment and serves as a prerequi- site to manipulation, which has been extensively studied for decades. Recent years have witnessed great progress in developing grasping algorithms for parallel grippers [8, 16, 17, 20, 49, 50] that carry high success rate on univer- sally grasping unknown objects. However, one fundamental limitation of parallel grasping is its low dexterity which lim- its its usage to complex and functional object manipulation. Dexterous grasping provides a more diverse way to grasp objects and thus is of vital importance to robotics for func- tional and fine-grained object manipulation [2, 29, 38, 40, 52]. However, the high dimensionality of the actuation space of a dexterous hand is both the advantage that en- dows it with such versatility and the major cause of the diffi- culty in executing a successful grasp. As a widely used five- finger robotic dexterous hand, ShadowHand [1] amounts to 26 degrees of freedom (DoF), in contrast with 7 DoF for a typical parallel gripper. Such high dimensionality magni- fies the difficulty in both generating valid grasp poses and planning the execution trajectories, and thus distinguishes the dexterous grasping task from its counterpart for paral- lel grippers. Several works have tackled the grasping pose synthesis problem [6, 28, 33, 49], however, they all assume oracle inputs (full object geometry and states). Very few works [9,38] tackle dexterous grasping in a realistic robotic setting, but so far no work yet can demonstrate universal and diverse dexterous grasping that can well generalize to unseen objects. In this work, we tackle this very challenging task: learn- ing universal dexterous grasping skills that can generalize well across hundreds of seen and unseen object categories in a realistic robotic setting and only allow us to access depth observations and robot proprioception information. Our dataset contains more than one million grasps for 5519 object instances from 133 object categories, which is the largest robotic dexterous grasping benchmark to evaluate universal dexterous grasping. Inspired by the successful pipelines from parallel grip- pers, we propose to decompose this challenging task into two stages: 1) dexterous grasp proposal generation , in which we predict diverse grasp poses given the point cloud observations; and 2) goal-conditioned grasp execution , in which we take one grasp goal pose predicted by stage 1 as a condition and generates physically correct motion trajecto- ries that comply with the goal pose. Note that both of these two stages are indeed very challenging, for each of which we contribute several innovations, as explained below.For dexterous grasp proposal generation, we devise a novel conditional grasp pose generative model that takes point cloud observations and is trained on our synthesized large-scale table-top dataset. Here our approach empha- sizes the diversity in grasp pose generation, since the way we humans manipulate objects can vary in many different ways and thus correspond to different grasping poses. With- out diversity, it is impossible for the grasping pose gen- eration to comply with the demand of later dexterous ma- nipulation. Previous works [22] leverages CV AE to jointly model hand rotation, translation, and articulations and we observe that such CV AE suffers from severe mode collapse and can’t generate diverse grasp poses, owing to its lim- ited expressivity when compared to conditional normaliz- ing flows [12, 13, 15, 23, 35] and conditional diffusion mod- els [5, 43, 46]. However, no works have developed normal- izing flows and diffusion models that work for the grasp pose space, which is a Cartesian product of SO(3) of hand rotation and a Euclidean space of the translation and joint angles. We thus propose to decompose this conditional generative model into two conditional generative models: a conditional rotation generative model, namely GraspIPDF, leveraging ImplicitPDF [34] (in short, IPDF) and a con- ditional normalizing flow, namely GraspGlow, leveraging Glow [23]. Combining these two modules, we can sample diverse grasping poses and even select what we need ac- cording to language descriptions. The sampled grasps can be further refined to be more physically plausible via Con- tactNet, as done in [22]. For our grasp execution stage, we learn a goal- conditioned policy that can grasp any object in the way specified by the grasp goal pose and only takes realistic in- puts: the point cloud observation and robot proprioception information, as required by real robot experiments. Note that reinforcement learning (RL) algorithms usually have difficulties with learning such a highly generalizable pol- icy, especially when the inputs are visual signals without ground truth states. To tackle this challenge, we leverage a teacher-student learning framework that first learns an or- acle teacher model that can access the oracle state inputs and then distill it to a student model that only takes realis- tic inputs. Even though the teacher policy gains access to oracle information, making it successful in grasping thou- sands of different objects paired with diverse grasp goals is still formidable for RL. We thus introduce two critical inno- vations: a canonicalization step that ensures SO(2) equiv- ariance to ease the policy learning; and an object curricu- lum that first learns to grasp one object with different goals, then one category, then many categories, and finally all cat- egories. Extensive experiments demonstrate the remarkable per- formance of our pipelines. In the grasp proposal generation stage, our pipeline is the only method that exhibits high di- 4738 versity while maintaining the highest grasping quality. The whole dexterous grasping pipeline, from the vision to pol- icy, again achieves impressive performance in our simula- tion environment and, for the first time, demonstrates a uni- versal grasping policy with more than 60% success rate and remarkably outperforms all the baselines. We will make the dataset and the code publicly available to facilitate future research.

Yang_ReCo_Region-Controlled_Text-to-Image_Generation_CVPR_2023

Abstract Recently, large-scale text-to-image (T2I) models have shown impressive performance in generating high-fidelity images, but with limited controllability, e.g., precisely spec- ifying the content in a specific region with a free-form text description. In this paper, we propose an effective technique for such regional control in T2I generation. We augment T2I models’ inputs with an extra set of position tokens, which represent the quantized spatial coordinates. Each region is specified by four position tokens to represent the top-left and bottom-right corners, followed by an open-ended nat- ural language regional description. Then, we fine-tune a pre-trained T2I model with such new input interface. Our model, dubbed as ReCo (Region-Controlled T2I), enables the region control for arbitrary objects described by open- ended regional texts rather than by object labels from a constrained category set. Empirically, ReCo achieves bet- ter image quality than the T2I model strengthened by posi-tional words (FID: 8.82→7.36, SceneFID: 15.54→6.51 on COCO), together with objects being more accurately placed, amounting to a 20.40% region classification accu- racy improvement on COCO. Furthermore, we demonstrate that ReCo can better control the object count, spatial re- lationship, and region attributes such as color/size, with the free-form regional description. Human evaluation on PaintSkill shows that ReCo is +19.28% and+17.21% more accurate in generating images with correct object count and spatial relationship than the T2I model. Code is available athttps://github.com/microsoft/ReCo .

1. Introduction Text-to-image (T2I) generation aims to generate faith- ful images based on an input text query that describes the image content. By scaling up the training data and model size, large T2I models [31, 34, 36, 45] have recently shown 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14246 remarkable capabilities in generating high-fidelity images. However, the text-only query allows limited controllability, e.g., precisely specifying the content in a specific region. The naive way of using position-related text words, such as “top left” and “bottom right,” often results in ambigu- ous and verbose input queries, as shown in Figure 2 (a). Even worse, when the text query becomes long and com- plicated, or describes an unusual scene, T2I models [31,45] might overlook certain details and rather follow the visual or linguistic training prior. These two factors together make region control difficult. To get the desired image, users usu- ally need to try a large number of paraphrased queries and pick an image that best fits the desired scene. The process known as “prompt engineering” is time-consuming and of- ten fails to produce the desired image. The desired region-controlled T2I generation is closely related to the layout-to-image generation [7,9,22,23,34,38, 44, 49]. As shown in Figure 2 (b), layout-to-image mod- els take all object bounding boxes with labels from a close set of object vocabulary [24] as inputs. Despite showing promise in region control, they can hardly understand free- form text inputs, nor the region-level combination of open- ended text descriptions and spatial positions. The two in- put conditions of text and box provide complementary re- ferring capabilities. Instead of separately modeling them as in text-to-image and layout-to-image generations, we study “region-controlled T2I generation” that seamlessly com- bines these two input conditions. As shown in Figure 2 (c), the new input interface allows users to provide open-ended descriptions for arbitrary image regions, such as precisely placing a “brown glazed chocolate donut” in a specific area. To this end, we propose ReCo (Region-Controlled T2I) that extends pre-trained T2I models to understand spatial coordinate inputs. The core idea is to introduce an ex- tra set of input position tokens to indicate the spatial po- sitions. The image width/height is quantized uniformly into Nbinsbins. Then, any float-valued coordinate can be ap- proximated and tokenized by the nearest bin. With an extra embedding matrix ( EP), the position token can be mapped onto the same space as the text token. Instead of designing a text-only query with positional words “in the topred donut” as in Figure 2 (a), ReCo takes region-controlled text inputs “<x 1>, <y 1>, <x 2>, <y 2>red donut,” where <x> ,<y> are the position tokens followed by the corresponding free- form text description. We then fine-tune a pre-trained T2I model with EPto generate the image from the extended input query. To best preserve the pre-trained T2I capabil- ity, ReCo training is designed to be similar to the T2I pre- training, i.e., introducing minimal extra model parameters (EP), jointly encoding position and text tokens with the text encoder, and prefixing the image description before the ex- tended regional descriptions in the input query. Figure 1 visualizes ReCo’s use cases and capabilities. As …… …… Figure 2. (a) With positional words ( e.g., bottom/top/left/right and large/small/tall/long), the T2I model (Stable Diffusion [34]) does not manage to create objects with desired properties. (b)Layout- to-image generation [22, 34, 38, 49] takes all object boxes and la- bels as the input condition, but only works well with constrained object labels. (c)Our ReCo model synergetically combines the text and box referring, allowing users to specify an open-ended re- gional text description precisely at any image region. shown in Figure 1 (a), ReCo could reliably follow the input spatial constraints and generate the most plausible images by automatically adjusting object statues, such as the view (front/side) and type (single-/double-deck) of the “bus.” Po- sition tokens also allow the user to provide free-form re- gional descriptions, such as “an orange cat wearing a red hat” at a specific location. Furthermore, we empirically ob- serve that position tokens are less likely to get overlooked or misunderstood than text words. As shown in Figure 1 (b), ReCo has better control over object count, spatial relation- ship, and size properties, especially when the query is long and complicated, or describes a scene that is less common in real life. In contrast, T2I models [34] may struggle with generating scenes with correct object counts (“ten”), rela- tionships (“boat below traffic light”), relative sizes (“chair larger than airplane”), and camera views (“zoomed out”). To evaluate the region control, we design a comprehen- sive experiment benchmark based on a pre-trained regional object classifier and an object detector. The object classi- fier is applied on the generated image regions, while the detector is applied on the whole image. A higher accuracy means a better alignment between the generated object lay- out and the region positions in user queries. On the COCO dataset [24], ReCo shows a better object classification ac- curacy ( 42.02%→62.42%) and detector averaged preci- sion ( 2.3→32.0), compared with the T2I model with care- fully designed positional words. For image generation qual- ity, ReCo improves the FID from 8.82to7.36, and Scene- FID from 15.54to6.51. Furthermore, human evaluations on PaintSkill [5] show +19.28% and+17.21% accuracy gain in more correctly generating the query-described ob- ject count and spatial relationship, indicating ReCo’s capa- bility in helping T2I models to generate challenging scenes. Our contributions are summarized as follows. 2 14247 • We propose ReCo that extends pre-trained T2I mod- els to understand coordinate inputs. Thanks to the in- troduced position tokens in the region-controlled input query, users can easily specify free-form regional de- scriptions in arbitrary image regions. • We instantiate ReCo based on Stable Diffusion. Ex- tensive experiments show that ReCo strictly follows the regional instructions from the input query, and also generates higher-fidelity images. • We design a comprehensive evaluation benchmark to validate ReCo’s region-controlled T2I generation ca- pability. ReCo significantly improves both the region control accuracy and the image generation quality over a wide range of datasets and designed prompts.

Yu_Turning_a_CLIP_Model_Into_a_Scene_Text_Detector_CVPR_2023

Abstract The recent large-scale Contrastive Language-Image Pre- training (CLIP) model has shown great potential in various downstream tasks via leveraging the pretrained vision and language knowledge. Scene text, which contains rich textual and visual information, has an inherent connection with a model like CLIP . Recently, pretraining approaches based on vision language models have made effective progresses in the field of text detection. In contrast to these works, this paper proposes a new method, termed TCM, focusing on Turning the CLIP Model directly for text detection without pretraining process. We demonstrate the advantages of the proposed TCM as follows: (1) The underlying principle of our framework can be applied to improve existing scene text detector. (2) It facilitates the few-shot training capability of existing methods, e.g., by using 10% of labeled data, we sig- nificantly improve the performance of the baseline method with an average of 22% in terms of the F-measure on 4 benchmarks. (3) By turning the CLIP model into existing scene text detection methods, we further achieve promising domain adaptation ability. The code will be publicly released at https://github.com/wenwenyu/TCM.

1. Introduction Scene text detection is a long-standing research topic aiming to localize the bounding box or polygon of each text instance from natural images, as it has wide practical applications scenarios, such as office automation, instant translation, automatic driving, and online education. With the rapid development of fully-supervised deep learning technologies, scene text detection has achieved remarkable progresses. Although supervised approaches have made remarkable progress in the field of text detection, they require extensive and elaborate annotations, e.g., character-level, word-level, and text-line level bounding boxes, especially polygonal boxes for arbitrarily-shaped scene text. Therefore, it is very important to investigate text detection methods *Equal contribution.†Corresponding author. (a) Enhancing backbone with cross -modal interactions (b) Enhancing backbone with text (c) OursImage Encoder Image -level TextImage Recognition DecoderImage Encoder Bbox/MaskDetection HeadImage AdaptPre-training Fine -tuningImage EncoderImage Cross -Modal InteractionsImage Encoder Bbox/MaskDetection HeadImage Text EncoderAdaptPre-training Fine -tuning Image EncoderImage MatchingText EncoderText Prompt Bbox/MaskDetection HeadFine -tuning(Partial) Text Designed Pretext Tasks MaskFigure 1. Comparisons of different paradigms of using text knowl- edge for scene text detection. under small amount of labeled data, i.e., few-shot training. Recently, through leveraging the pretrained vision and language knowledge, the large-scale Contrastive Language- Image Pretraining (CLIP) model [26] has demonstrated its significance in various downstream tasks. e.g., image classi- fication [53], object detection [5], and semantic segmenta- tion [12, 27, 43]. Compared to general object detection, scene text in natu- ral images usually presents with both visual and rich char- acter information, which has a natural connection with the CLIP model. Therefore, how to make full use of cross-modal information from visual, semantic, and text knowledge to improve the performance of the text detection models re- ceives increasing attentions in recent studies. For exam- ples, Song et al. [29], inspired by CLIP, adopts fine-grained cross-modality interaction to align unimodal embeddings for learning better representations of backbone via carefully designed pretraining tasks. Xue et al. [46] presents a weakly supervised pretraining method to jointly learn and align vi- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6978 sual and partial textual information for learning effective visual text representations for scene text detection. Wan et al.[35] proposes self-attention based text knowledge mining to enhance backbone via an image-level text recognition pretraining tasks. Different from these works, as shown in Figure 1, this paper focuses on turning the CLIP model for text detection without pretraining process. However, it is not trivial to incorporate the CLIP model into a scene text detector. The key is seeking a proper method to exploit the visual and semantic prior information conditioned on each image. In this paper, we develop a new method for scene text detec- tion, termed as TCM, short for Turning a CLIPModel into a scene text detector, which can be easily plugged to im- prove the scene text detection frameworks. We design a cross-modal interaction mechanism through visual prompt learning, which is implemented by cross-attention to recover the locality feature from the image encoder of CLIP to cap- ture fine-grained information to respond to the coarse text region for the subsequent matching between text instance and language. Besides, to steer the pretrained knowledge from the text encoder conditioned independently on different input images, we employ the predefined language prompt, learnable prompt, and a language prompt generator using simple linear layer to get global image information. In ad- dition, we design an instance-language matching method to align the image embedding and text embedding, which en- courages the image encoder to explicitly refine text regions from cross-modal visual-language priors. Compared to pre- vious pretraining approaches, our method can be directly finetuned for the text detection task without pretraining pro- cess, as elaborated in Fig. 1. In this way, the text detector can absorb the rich visual or semantic information of text from CLIP. We summarize the advantages of our method as follows: •We construct a new text detection framework, termed as TCM, which can be easily plugged to enhance the existing detectors. •Our framework can enable effective few-shot training capability. Such advantage is more obvious when using less training samples compared to the baseline detectors. Specifically, by using 10% of labeled data, we improve the performance of the baseline detector by an average of 22% in terms of the F-measure on 4 benchmarks. •TCM introduces promising domain adaptation ability, i.e., when using training data that is out-of-distribution of the testing data, the performance can be significantly improved. Such phenomenon is further demonstrated by a NightTime-ArT text dataset1, which we collected from the ArT dataset. 1NightTime-ArT Download Link•Without pretraining process using specific pretext tasks, TCM can still leverage the prior knowledge from the CLIP model, outperforming previous scene text pre- training methods [29, 35, 46].

Xu_Probabilistic_Knowledge_Distillation_of_Face_Ensembles_CVPR_2023

Abstract Mean ensemble (i.e. averaging predictions from multiple models) is a commonly-used technique in machine learning that improves the performance of each individual model. We formalize it as feature alignment for ensemble in open-set face recognition and generalize it into Bayesian Ensemble Averaging (BEA) through the lens of probabilistic modeling. This generalization brings up two practical benefits that ex- isting methods could not provide: (1) the uncertainty of a face image can be evaluated and further decomposed into aleatoric uncertainty and epistemic uncertainty, the latter of which can be used as a measure for out-of-distribution detection of faceness; (2) a BEA statistic provably reflects the aleatoric uncertainty of a face image, acting as a mea- sure for face image quality to improve recognition perfor- mance. To inherit the uncertainty estimation capability from BEA without the loss of inference efficiency, we propose BEA-KD, a student model to distill knowledge from BEA. BEA-KD mimics the overall behavior of ensemble members and consistently outperforms SOTA knowledge distillation methods on various challenging benchmarks.

1. Introduction Knowledge Distillation (KD) is an active research area that has profound benefits for model compression, wherein competitive recognition performance can be achieved by smaller models (student models) via a distillation process from teacher models. As such, smaller models can be de- ployed into space-constrained environments such as mobile and embedded devices. There has been abundant literature in KD for face recog- nition [22,33,34]. However, all the existing approaches fall into the “one-teacher-versus-one-student” paradigm. This learning paradigm has several limitations. Firstly, a single teacher can be biased, which further results in biased esti- *Equal contribution. Figure 1. A conceptual illustration of BEA and BEA-KD. Given a face image x, we have n= 5 probabilistic ensemble members {r-vMF (µi, κi)}n i=1(marked by light blue). Bayesian ensemble averaging (marked by dark blue) returns a single r-vMF (˜µx,¯κ(n) x) that accounts for the expected positions and confidence by all the ensemble members. To emulate the ensemble’s probabilistic be- havior, we employ a parametrized distribution qϕ,φ(z|x)to ap- proximate BEA. mates of face feature embeddings given by a student after knowledge distillation from the biased teacher. Secondly, it only yields point estimates of face feature embeddings, un- able to provide uncertainty measure for face recognition in a safety-sensitive scenario. Compared with single-teacher KD, KD from multiple teachers (a.k.a. ensemble KD) is beneficial and has been extensively explored in literature [9, 11, 29, 32, 36, 37]. However, these approaches are designed solely for closed- setclassification tasks, distilling logits in a fixed simplex space via KL divergence (as the label set remains the same throughout training and test). In contrast, face recognition is inherently an open-set problem where classes cannot be known a priori. More specifically, face identities appear- ing during the inference stage scarcely overlap with those in the training phase. Consequently, without a fixed sim- plex space for logit distillation, existing approaches cannot be readily applied to face recognition. As will be shown in our empirical studies, existing closed-set KD approaches This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3489 exhibit inferior performance in face recognition tasks with million-scale label sets. How we treat teachers highly affects KD performance. Unlike prior art [20, 26] that takes average of predictions (termed ‘mean ensemble’ throughout this paper), we treat teachers as draws in a probabilistic manner. We find that this treatment leads to a generalization of mean ensemble, namely Bayesian Ensemble Averaging (BEA), which fur- ther brings up two practical benefits that existing methods could not provide: (1) the uncert

Yang_MIANet_Aggregating_Unbiased_Instance_and_General_Information_for_Few-Shot_Semantic_CVPR_2023

Abstract Existing few-shot segmentation methods are based on the meta-learning strategy and extract instance knowledge from a support set and then apply the knowledge to seg- ment target objects in a query set. However, the extracted knowledge is insufficient to cope with the variable intra- class differences since the knowledge is obtained from a few samples in the support set. To address the problem, we propose a multi-information aggregation network (MI- ANet) that effectively leverages the general knowledge, i.e., semantic word embeddings, and instance information for accurate segmentation. Specifically, in MIANet, a general information module (GIM) is proposed to extract a general class prototype from word embeddings as a supplement to instance information. To this end, we design a triplet loss that treats the general class prototype as an anchor and samples positive-negative pairs from local features in the support set. The calculated triplet loss can transfer seman- tic similarities among language identities from a word em- bedding space to a visual representation space. To alle- viate the model biasing towards the seen training classes and to obtain multi-scale information, we then introduce a non-parametric hierarchical prior module (HPM) to gener- ate unbiased instance-level information via calculating the pixel-level similarity between the support and query image features. Finally, an information fusion module (IFM) com- bines the general and instance information to make pre- dictions for the query image. Extensive experiments on PASCAL-5iand COCO-20ishow that MIANet yields supe- rior performance and set a new state-of-the-art. Code is available at github.com/Aldrich2y/MIANet.

1. Introduction The challenge of few-shot semantic segmentation (FSS) is how to effectively use one or five labeled samples to seg- ment a novel class. Existing few-shot segmentation meth- *Corresponding author ([email protected]). Support imageExtract featuresExtract features CNN Category Labels bottle cat dog person boattv bird sheepCategory Labels bottle cat dog person boattv bird sheep get the vectorInstance Info Generation General Info GenerationInfo Fusion Prediction Instance Info Path Main Path General Info Path (b) Our semantic segmentation network Support image Extract features Extract featuresCNNSupport mask Guidance information ClassifierPrediction (a) Existing few -shot segmentation methods Query image Query image w2vbottle Query image Figure 1. Comparison between (a) existing FSS methods and (b) proposed MIANet. (a) Existing methods extract instance-level knowledge from the support images, which is not able to cope with large intra-class variation. (b) our MIANet extracts instance-level knowledge from the support images and obtains general class in- formation from word embeddings. These two types of information benefit the final segmentation. ods [28, 30, 33, 37] adopt the metric-based meta-learning strategy [26, 29]. The strategy is typically composed of two stages: meta-training and meta-testing. In the meta- training stage, models are trained by plenty of independent few-shot segmentation tasks. In meta-testing, models can thus quickly adapt and extrapolate to new few-shot tasks of unseen classes and segment the novel categories since each training task involves a different seen class. As shown in Figure 2, natural images of same categories have semantic differences and perspective distortion, which leads to intra-class differences. Current FSS approaches segment a query image by matching the guidance informa- tion from the support set with the query features (Figure 1 (a)). Unfortunately, the correlation between the support im- age and the query image is not enough to support the match- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7131 Chair Bird Bird Aeroplane (a) (b) Perspective distortion Semantic differencesFigure 2. We define two types of intra-class variation. (a) The object in each column has the same semantic label but belongs to different fine-grained categories. (b) The object belonging to the same category differs greatly in appearance due to the existence of perspective distortion. ing strategy in some support-query pairs due to the diversity of intra-class differences, which affects the generalization performance of the models. On the other hand, modules with numerous learnable parameters are devised by FSS methods to better use the limited instance information. And lots of few-shot segmentation tasks of seen classes are used to train the models in the meta-training stage. Although cur- rent methods freeze the backbone, the rest parameters will inevitably fit the feature distribution of the training data and make the trained models misclassify the seen training class to the unseen testing class. To address the above issues, a multi-information ag- gregation network is proposed for accurate segmentation. Specifically, we first design a general information module (GIM) to produce a general class prototype by leveraging class-based word embeddings. This prototype represents general information for the class, which is beyond the sup- port information and can supplement some missing class information due to intra-class differences. As shown in Fig- ure 1 (b), the semantic word vectors for each class can be obtained by a pre-trained language model, i.e., word2vec . Then, GIM takes the word vector and a support prototype as input to get the general prototype. Next, a well-designed triplet loss [25] is applied to achieve the alignment be- tween the semantic prototype and the visual features. The triplet loss extracts positive-negative pairs from local fea- tures which distinguishes our method from other improved triplets [3,4,11]. The semantic similarity between the word embeddings in a word embedding space can therefore be transferred to a visual embedding space. Finally, the pro- jected prototype is supplemented into the main branch as the general information of the category for information fu- sion to alleviate the intra-class variance problem. Moreover, to capture the instance-level details and allevi-ate the model biasing towards the seen classes, we propose a non-parametric hierarchical prior module (HPM). HPM works in two aspects. (1) HPM is class-agnostic since it does not require training. (2) HPM can generate hierarchi- cal activation maps for the query image by digging out the relationship between high-level features for accurate seg- mentation of unseen classes. In addition, we build informa- tion channels between different scales to preserve discrim- inative information in query features. Finally, the unbiased instance-level information and the general information are aggregated by an information fusion module (IFM) to seg- ment the query image. Our main contributions are summa- rized as follows: (1) We propose a multi-information aggregation network (MIANet) to aggregate general information and unbi- ased instance-level information for accurate segmenta- tion. (2) To the best of our knowledge, this is the first time to use word embeddings in FSS, and we design a general information module (GIM) to obtain the general class information from word embeddings for each class. The module is optimized through a well-designed triplet loss and can provide general class information to alleviate intra-class differences. (3) A non-parametric hierarchical prior module (HPM) is proposed to supply MIANet with unbiased instance- level segmentation knowledge, which provides the prior information of the query image on multi-scales and alleviates the bias problem in testing. (4) Our MIANet achieves state-of-the-art results on two few-shot segmentation benchmarks, i.e., PASCAL-5i and COCO-20i. Extensive experiments validate the ef- fectiveness of each component in our MIANet.

Yan_Towards_Trustable_Skin_Cancer_Diagnosis_via_Rewriting_Models_Decision_CVPR_2023

Abstract Deep neural networks have demonstrated promising per- formance on image recognition tasks. However, they may heavily rely on confounding factors, using irrelevant arti- facts or bias within the dataset as the cue to improve perfor- mance. When a model performs decision-making based on these spurious correlations, it can become untrustable and lead to catastrophic outcomes when deployed in the real- world scene. In this paper, we explore and try to solve this problem in the context of skin cancer diagnosis. We intro- duce a human-in-the-loop framework in the model training process such that users can observe and correct the model’s decision logic when confounding behaviors happen. Specif- ically, our method can automatically discover confounding factors by analyzing the co-occurrence behavior of the sam- ples. It is capable of learning confounding concepts using easily obtained concept exemplars. By mapping the black- box model’s feature representation onto an explainable con- cept space, human users can interpret the concept and in- tervene via first order-logic instruction. We systematically evaluate our method on our newly crafted, well-controlled skin lesion dataset and several public skin lesion datasets. Experiments show that our method can effectively detect and remove confounding factors from datasets without any prior knowledge about the category distribution and does not require fully annotated concept labels. We also show that our method enables the model to focus on clinical- related concepts, improving the model’s performance and trustworthiness during model inference.

1. Introduction Deep neural networks have achieved excellent perfor- mance on many visual recognition tasks [11,14,35]. Mean- while, there are growing concerns about the trustworthiness of the model’s black-box decision-making process. In med- ical imaging applications, one of the major issues is deep learning models’ confounding behaviors using irrelevant ar- tifacts ( i.e. rulers, dark corners) or bias ( i.e. image back- Figure 1. Our method allows people to correct the model’s con- founding behavior within the skin lesion training set via rewriting the model’s logic. grounds, skin tone) as the cue to make the final predictions. These spurious correlations in the training distribution can make models fragile when novel testing samples are pre- sented. Therefore, transparency of decision-making and human-guided bias correction will significantly increase the reliability and trustworthiness of model deployments in a life-critical application scenario like cancer diagnosis. For instance, due to the nature of dermatoscopic images, the skin cancer diagnosis often involves confounding fac- tors [4, 22, 24, 40], i.e., dark corners, rulers, and air pock- ets. Bissoto et al . [40] shows that deep learning models trained on common skin lesion datasets can be biased. Sur- prisingly, they found the model can reach an AUC of 73%, even though lesions in images are totally occluded. To bet- ter understand the issue of confounding behaviors, we il- lustrate a motivating example. Fig. 1 shows a represen- tative confounding factor “dark corners” in the ISIC2019- 2020 [36, 38] dataset, where the presence of dark corners is much higher in the melanoma images than in benign im- ages. A model trained on this dataset tends to predict a le- sion as melanoma when dark corners appear. This model is undoubtedly untrustable and is catastrophic for deployment in real clinical practice. To deal with this nuisance and im- prove the model’s trustworthiness, an ideal method would be: i ) the model itself is explainable so that humans can un- derstand the prediction logic behind it. and ii) the model has an intervening mechanism that allows humans to correct the model once confounding behavior happens. Through this This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11568 way, the model can avoid dataset biasing issues and rely on expected clinical rules when needed to diagnose. In this work, our overall goal is to improve model trans- parency as well as develop a mechanism that human-user can intervene the model training process when confound- ing factors are observed. The major difference between our proposed solution and existing works [19,28,29,33] are: (1) Formodel visualization , we focus on generating human- understandable textual concepts rather than pixel-wise at- tribution explanations like Class Activation Maps [42]. (2) Forconcept learning , we do not hold any prior knowl- edge about the confounding concepts within the dataset. This makes our problem set-up more realistic and practi- cal than the existing concept-based explanation or inter- action [19, 33] framework where fully-supervised concept labels are required. (3) For method generalization , our method can be applied on top of any deep models. Therefore, we propose a method that is capable of learning confounding concepts with easily obtained con- cept examplars. This is realized via clustering model’s co-occurrence behavior based on spectral relevance anal- ysis [20] and concept learning based on concept activation vector (CA Vs) [18]. Also, a concept space learned by CA Vs is more explainable than the feature-based counterparts. To the end, we propose a human-in-the-loop framework that can effectively discover and remove the confounding behaviors of the classifier by transforming its feature rep- resentation into explainable concept scores. Then, humans are allowed to intervene based on first-order logic. Through human interaction (see Fig. 1) on the model’s concept space, people can directly provide feedback to the model training (feedback turned into gradients) and remove unwanted bias and confounding behaviors. Moreover, we notice that no suitable dermatology datasets are available for confounding behavior analysis. To increase the methods’ reproducibil- ity and data accessibility in this field, we curated a well- controlled dataset called ConfDerm containing 3576 im- ages based on well-known ISIC2019 andISIC2020 datasets. Within the training set, all images in one of the classes are confounded by one of five confounding factors, including dark corners, borders, rulers, air pockets, and hairs. Still, in the testing set, all images are random. Some confounding factors within the dataset are illustrated in Fig. 2. We summarize our main contributions as: (1) We crafted a novel dataset called ConfDerm, which is the first skin lesion dataset used for systematically evaluating the model’s trustworthiness under different confounding behav- iors within the training set. (2) Our new spectral rele- vance analysis algorithm on the popular skin cancer dataset ISIC2019-2020 has revealed insights that artifacts such as dark corners, rulers, and hairs can significantly confound modern deep neural networks. (3) We developed a human- in-the-loop framework using concept mapping and logic (a) (b) (c) (d) (e) Figure 2. Observed confounding concepts in ISIC2019-2020 datasets, the top row shows sample images, and the bottom row is the corresponding heatmap from GradCAM: (a) dark corners. (b) rulers. (c) dark borders. (e) dense hairs. (f) air pockets. layer rewriting to make the model self-explainable and enable users effectively remove the model’s confounding behaviors. (4) Experimental results on different datasets demonstrate the superiority of our method on performance improvement, artifact removal, and skin tone debiasing.

Zhang_Ingredient-Oriented_Multi-Degradation_Learning_for_Image_Restoration_CVPR_2023

Abstract Learning to leverage the relationship among diverse im- age restoration tasks is quite beneficial for unraveling the intrinsic ingredients behind the degradation. Recent years have witnessed the flourish of various All-in-one methods, which handle multiple image degradations within a single model. In practice, however, few attempts have been made to excavate task correlations in that exploring the underly- ing fundamental ingredients of various image degradations, resulting in poor scalability as more tasks are involved. In this paper, we propose a novel perspective to delve into the degradation via an ingredients-oriented rather than pre- vious task-oriented manner for scalable learning. Specif- ically, our method, named Ingredients-oriented Degrada- tion Reformulation framework (IDR), consists of two stages, namely task-oriented knowledge collection and ingredients- oriented knowledge integration. In the first stage, we con- duct ad hoc operations on different degradations according to the underlying physics principles, and establish the cor- responding prior hubs for each type of degradation. While the second stage progressively reformulates the preceding task-oriented hubs into single ingredients-oriented hub via learnable Principal Component Analysis (PCA), and em- ploys a dynamic routing mechanism for probabilistic un- known degradation removal. Extensive experiments on var- ious image restoration tasks demonstrate the effectiveness and scalability of our method. More importantly, our IDR exhibits the favorable generalization ability to unknown downstream tasks.

1. Introduction Image restoration aims to recover the high-quality im- ages from their degraded observations, which is a general term of a series of low-level vision tasks. In addition to achieving satisfactory visual effects in photography, image restoration is also widely used in many other real world scenarios, such as autopilot and surveillance. Complex *Both authors contributed equally to this research. †Corresponding author. Figure 1. An illustration of our proposed ingredients-oriented degradation reformulation principle. Instead of previous task- oriented paradigm where each tasks are learned exclusively, we perform an ingredients-oriented paradigm to explore the correla- tion among diverse restoration tasks for scaleable degradation re- formulated learning, where the Conv. ,Add. andMul. means the convolution, addition and multiplication. environments put forward higher requirements for image restoration algorithms, when considering the variability and unknowability of the corruption types. Since most exist- ing methods have been dedicated into single degradation re- moval, such as denoising [15,24,61], deraining [20,52,55], debluring [8, 40, 42], dehazing [26, 44, 45], low-light en- hancement [14, 34, 50], etc., which do not satisfy the ap- plications in real world scenarios. Recently, all-in-one fashion methods have been com- ing to the fore, which handle multiple image degradations within a single model. These methods can be roughly categorized into two families, i.e., corruption-specific and corruption-agnostic. Representative studies of the former [2, 28] deal with different degradations via separate sub- networks, which demands pre-specification of corruption types, limiting the scope of further application. While the efforts in latter [25, 47] release the model from the prior of the corruption types, improving the flexibility in prac- tice. However, both of them suffer from poor scalability as more tasks are involved, indicating that the diverse degrada- tions are learned exclusively under the potential capability bottleneck, without touching the intrinsic correlation among them, which we referred as task-oriented paradigm. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5825 To solve the above problem, we ask two questions: i) ’whether there are commonalities between different degra- dations?’ During past decades, few of works have been devoted to this field, [10] presented the interrelationship between image dehazing and low-light image enhancement. Going a step further, we envision that such association are widespread in various degradations, such as directionality in deblurring and deraining, unnatural image layering in deraining and denoising. Therefore, it is of great interest to consider the correlation among various restoration tasks for learning the intrinsic ingredients behind the degrada- tion, which we referred as ingredient-oriented paradigm. ii) ’Whether an corrupted image definitely ascribed to only one type of degradation?’ In real world scenarios, it is hard to determine as multiple degradations may occur simul- taneously, such as heavy rain typically accumulated with mist, or low-light combined with blur in night-time surveil- lance [63]. Therefore, it is inappropriate to learn each restoration task exclusively. In this paper, we propose Ingredients-oriented Degrada- tion Reformulation framework (IDR) for all-in-one image restoration, which provides a novel perspective via delving into the degradation and concentrating on the underlying fundamental ingredients. Specifically, the learning proce- dure of IDR consists of two stages, namely task-oriented knowledge collection and ingredients-oriented knowledge integration. We perform the above reformulation in the meta prior learning module (MPL) with the collaboration of both degradation representation and degradation opera- tion, while the backbone network can be any transformer- based architecture. In the first stage, we conduct ad hoc operations for different degradations depending on the un- derlying physics principles, which pre-embedding the pri- ors of disparate physics characteristics respectively. Mean- while, separate task-oriented prior hubs are established for each type of degradation, responsible for excavating the specific degradation ingredients for compositional represen- tation. While the second stage progressively reformulates the proceeding task-oriented hubs into single ingredients- oriented hub via learnable Principal Component Analysis (PCA), striving for commonalities among multiple degrada- tions in terms of the ingredient-level, while preserving re- spective variance information as much as possible. Besides, a dynamic soft routing mechanism is employed in MPL for probabilistic unknown1degradation removal, according to the operation priors embedded in the first stage. The contributions of this work are summarized as below: • We rethink the current paradigm of all-in-one fashion methods, and propose to delve into the degradation for intrinsic ingredients excavation, in that improving the scalability of the model. 1Namely, the degradation types are not available in the second stage.• We propose the Ingredients-oriented Degradation Re- formulation framework (IDR) for image restoration, which consists of two stages, i.e., task-oriented knowl- edge collection and ingredients-oriented knowledge integration, collaborating on both degradation repre- sentation and degradation operation. • Extensive experiments are conducted to verify the ef- fectiveness of our method. As far as we know, IDR is the first work to perform up to five image restoration tasks in an all-in-one fashion.

Yang_POEM_Reconstructing_Hand_in_a_Point_Embedded_Multi-View_Stereo_CVPR_2023

Abstract Enable neural networks to capture 3D geometrical- aware features is essential in multi-view based vision tasks. Previous methods usually encode the 3D information of multi-view stereo into the 2D features. In contrast, we present a novel method, named POEM, that directly oper- ates on the 3D POintsEmbedded in the Multi-view stereo for reconstructing hand mesh in it. Point is a natural form of 3D information and an ideal medium for fusing fea- tures across views, as it has different projections on dif- ferent views. Our method is thus in light of a simple yet effective idea, that a complex 3D hand mesh can be rep- resented by a set of 3D points that 1) are embedded in the multi-view stereo, 2) carry features from the multi-view images, and 3) encircle the hand. To leverage the power of points, we design two operations: point-based feature fusion and cross-set point attention mechanism. Evalua- tion on three challenging multi-view datasets shows that POEM outperforms the state-of-the-art in hand mesh re- construction. Code and models are available for research atgithub.com/lixiny/POEM

1. Introduction Hand mesh reconstruction plays a central role in the field of augmented and mixed reality, as it can not only deliver realistic experiences for the users in gaming but also sup- port applications involving teleoperation, communication, education, and fitness outside of gaming. Many significant efforts have been made for the monocular 3D hand mesh reconstruction [1, 5, 7, 9, 31, 32]. However, it still strug- gles to produce applicable results, mainly for these three reasons. (1)Depth ambiguity. Recovery of the absolute position in a monocular camera system is an ill-posed prob- lem. Hence, previous methods [9, 31, 54] only recovered the hand vertices relative to the wrist ( i.e. root-relative). (2)Unknown perspectives. The shape of the hand’s 2D †Cewu Lu is the corresponding author, the member of Qing Yuan Re- search Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China and Shanghai Qi Zhi institute. Figure 1. Intersection area of Ncameras’ frustum spaces. The gray dots represent the point cloud Paggregated from Nfrustums. Our method: POEM, standing for the point embedded multi-view stereo, focuses on the dark area scatted with gray dots. projection is highly dependent on the camera’s perspec- tive model ( i.e. camera intrinsic matrix). However, the monocular-based methods usually suggest a weak perspec- tive projection [1, 27], which is not accurate enough to re- cover the hand’s 3D structure. (3)Occlusion. The occlu- sion between the hand and its interacting objects also chal- lenges the accuracy of the reconstruction [32]. These issues limit monocular-based methods from practical application, in which the absolute and accurate position of the hand sur- face is required for interacting with our surroundings. Our paper is thus focusing on reconstructing hands from multi-view images. Motivation comes from two aspects. First, the issues mentioned above can be alleviated by lever- aging the geometrical consistency among multi-view im- ages. Second, the prospered multi-view hand-object track- ing setups [2, 4, 49, 55] and VR headsets bring us an urgent demand and direct application of multi-view hand recon- struction in real-time. A common practice of multi-view 3D pose estimation follows a two-stage design. It first estimates the 2D key points of the skeleton in each view and then back-project them to 3D space through several 2D-to-3D lifting methods, e.g. algebraic triangulation [17,18,39], Pic- torial Structures Model (PSM) [33, 38], 3D CNN [18, 43], This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21108 plane sweep [26], etc. However, these two-stage methods are not capable of reconstructing an animatable hand mesh that contains both skeleton and surface. It was not until re- cently that a one-stage multi-view mesh regression model was proposed [45]. How to effectively fuse the features from different im- ages is a key component in the multi-view setting. Ac- cordingly, previous methods can be categorized into three types. (1) Fusing in 2D. The features are directly fused in the 2D domain using explicit epipolar transform [17, 38] or implicit representations that encode the camera transfor- mation ( i.e. camera intrinsic and extrinsic matrix) into 2D features, e.g. feature transform layer (FTL) [14, 39] and 3D position embedding (RayConv) [45]; (2) Fusing in 3D. The features are fused in a 3D voxel space via PSM [33, 38] or 3D CNNs [18, 43]; (3) Fusing via 3D-2D projection. The features are fused by first projecting the 3D keypoints’ ini- tial guess into each 2D plane and then fusing multi-view features near those 2D locations [45]; The fusion mode in type 1 is considered as holistic , since it indiscriminately fuses all the features from differ- ent views. Consequentially, it ignores the structure of the underlying hand model that we are interested in. On the contrary, the fusion mode in type 3 is considered as lo- cal. However, only the features around the 2D keypoints are hard to capture the consistent geometrical features from a global view. Besides, the 3D keypoints initial guess may not be accurate enough, resulting in the fusion being unsta- ble. The fusion mode in type 2 is not in our consideration as it tends to be computationally expensive and suffers from quantization error. Based on the above discussion, we aim to seek a fea- ture representation and a fusion mode between type 1 and type 3 for both holistically and locally fusing the features in multi-views, and to explore a framework for robust and accurate hand mesh reconstruction. Our method is called POEM, standing for POintEmbedded Multi-view Stereo . We draw inspiration from the Basis Point Set (BPS) [34], which bases on a simple yet effective idea that a complex 3D shape can be represented by a fixed set of points (BPS) that wraps the shape in it. If we consider the intersection of different cameras’ frustum spaces as a point cloud, and the hand’s vertices as another point cloud, then the intersected space is the basis point set for hand vertices (see Fig. 1). Once we assign the multi-view image features to the point cloud in the intersected space, fusing image features across different views becomes fusing the point features from dif- ferent camera frustums. The advantages of this representa- tion are two-fold: (i)The hand is wrapped in a dense point cloud (set) that carries dense image features collected from different views, which is more holistic and robust than the local fusion mode in type 3. (ii)For each vertex on the hand surface, it interacts with basis points in its local neighbor-hood ( i.e.knearest neighbor), which is more selective than the holistic fusion mode in type 1. Fig. 2 shows our model’s architecture. POEM consists of two stages. In the first stage (Sec. 3.2), POEM takes the multi-view images as input and predicts the 2D keypoints of the hand skeleton in each view. Then, the 3D keypoints are recovered by an algebraic triangulation module. In the second stage, POEM fuses the features from different views in a space embedded by points and predicts the hand mesh in this space (Sec. 3.3). The point feature on hand vertices will iteratively interact with the features of the embedded points through a cross-set attention mechanism, and the up- dated vertex features are further used by POEM to predict the vertex’s refined position (Sec. 3.3.3). We conduct extensive experiments on three multi-view datasets for hand mesh reconstruction under the object’s oc- clusion, namely HO3D [12], DexYCB [4], and OakInk [49]. With the proposed fusion mode and attention mechanism, POEM achieves state-of-the-art on all three datasets. Our contributions are in three-fold: • We investigate the multi-view pose and shape recon- struction problem from a new perspective, that is, the interaction between a target point set ( i.e. mesh vertices) and a basis point set ( i.e. point cloud in the camera frus- tum spaces). • According to that, we propose an end-to-end learning framework: POEM for reconstructing hand mesh from multi-view images through a point embedded multi-view stereo. To encourage interaction between two point sets, POEM introduces two new operations: a point-based feature fusion strategy and a cross-set point attention. • We conduct extensive experiments to demonstrate the efficacy of the architecture in POEM. As a regression model targeting mesh reconstruction, POEM achieves significant improvement compared to the previous state- of-the-art.

Yuan_Bi3D_Bi-Domain_Active_Learning_for_Cross-Domain_3D_Object_Detection_CVPR_2023

Abstract Unsupervised Domain Adaptation (UDA) technique has been explored in 3D cross-domain tasks recently. Though preliminary progress has been made, the performance gap between the UDA-based 3D model and the supervised one trained with fully annotated target domain is stilllarge. This motivates us to consider selecting partial-yet- important target data and labeling them at a minimum cost,to achieve a good trade-off between high performance andlow annotation cost. To this end, we propose a Bi-domainactive learning approach, namely Bi3D, to solve the cross-domain 3D object detection task. The Bi3D first develops adomainness-aware source sampling strategy, which identi-fies target-domain-like samples from the source domain toavoid the model being interfered by irrelevant source data.Then a diversity-based target sampling strategy is devel- oped, which selects the most informative subset of target do-main to improve the model adaptability to the target domain using as little annotation budget as possible. Experimentsare conducted on typical cross-domain adaptation scenar- ios including cross-LiDAR-beam, cross-country, and cross- sensor , where Bi3D achieves a promising target-domain de- tection accuracy ( 89.63% on KITTI) compared with UDA- based work ( 84.29% ), even surpassing the detector trained on the full set of the labeled target domain ( 88.98% ). Our code is available at: https://github.com/PJLab- ADG/3DTrans .

1. Introduction LiDAR-based 3D Object Detection (3DOD) [ 5,13,26, 28,40] has advanced a lot recently. However, the gen- eralization of a well-trained 3DOD model from a sourcepoint cloud dataset (domain) to another one, namely cross- ∗This work was done when Jiakang Y uan was an intern at Shanghai AI Laboratory. †Corresponding to: Tao Chen ([email protected]), Bo Zhang ([email protected])(a) General 3D Object Detection2D BackboneRPNDetector Head Map to BEVTrain (b) Self-training based UDA for 3D Object Detection Detector Pseudo LabelPseudo LabelPredicted Bounding BoxesPredicted Bounding Boxes Filtering low confidence bounding boxesNMSData Augmentation Train3D backbone BEV features (c) ADA for 3D Object Detection Detector Selected representative unlabel target framesInstance-level features Scene-level features AnnotateTrainSource Target OracleInstance categories Bounding boxes Active sampling criteria(a) General 3D Object Detection2D Ba ckboneRPNDetector HeadMap to BE V Train 3D backbone BEV features(a) General 3D Object Detection2D BackboneRPNDetector Head Map to BEV Train (b) Self -fftraining based UDA for 3D Object Detection Detector Pseudo LabelPredicted Bounding Boxe s Filtering low confidence bounding boxesNMS Data Augmentatio n Train Instance categorie s Bounding boxe s(b) Self-training based UDA for 3D Object Detection Detector Pseudo LabelPredicted Bounding Boxes Filtering low confidence bounding boxesNMS Data Augmentation Train3D backbone BEV features (c) ADA for 3D Object Detection Detector Selected representativ e unlabel target frame s Instance-level features Scene-level features AnnotateTrain Source Target OracleActive sampling criteria(c) ADA for 3D Object Detection Detector Selected representative unlabel target frames Instance-level features Scene-level features AnnotateTrain Source Target OracleInstance categories Bounding boxes Active sampling criteria Figure 1. Comparisons among (a) The general 3DOD pipeline, (b) Self-training based Unsupervised Domain Adaptation 3DODpipeline, and (c) Active Domain Adaptation 3DOD pipeline that selects representative target data, and then annotates them by an oracle (human expert) for subsequent model refinement. domain 3DOD, is still under-explored. Such a task in fact is important in many real-world applications. For example,in the autonomous driving scenario, the target scene distri-bution frequently changes due to unforeseen differences indynamically changing environments, making cross-domain3DOD an urgent problem to be resolved. Benefiting from the success of Unsupervised Domain Adaptation (UDA) technique in 2D cross-domain tasks [ 3, 7,10,14,32,45,48], several attempts are made to apply UDA for tackling 3D cross-domain tasks [ 15,20,22,36,39,42,46]. ST3D [ 42] designs a self-training-based framework to adapt a pre-trained detector from the source domain to a new tar-get domain. LiDAR distillation [ 36] exploits transferable knowledge learned from high-beam LiDAR data to the low one. Although these UDA 3D models have achieved sig-nificant performance gains for the cross-domain task, there This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15599 is still a large performance gap between these UDA models and the supervised ones trained using a fully-annotated tar- get domain. For example, ST3D [ 42] only achieves 72.94% AP 3Din nuScenes [ 1]-to-KITTI [ 8] cross-domain setting, yet the fully-supervised result using the same baseline de- tector can reach to 82.50% AP 3Don KITTI. To further reduce the detection performance gap between UDA-based 3D models and the fully-supervised ones, aninitial attempt is to leverage Active Domain Adaptation(ADA) technique [ 6,17,29,37,38], whose goal is to se- lect a subset quota of all unlabeled samples from the targetdomain to perform the manual annotation for model train-ing. Actually, the ADA task has been explored in 2D vi- sion fields such as AADA [ 29], TQS [ 6], and CLUE [ 17], but its research on 3D point cloud data still remains blank. In order to verify the versatility of 2D image-based ADAmethods towards 3D point cloud, we conduct extensive at- tempts by integrating the recently proposed ADA methods,e.g., TQS [ 6] and CLUE [ 17], into several typical 3D base- line detectors, e.g., PV-RCNN [ 26] and V oxel R-CNN [ 5]. Results show that these 2D ADA methods cannot obtain sat- isfactory detection accuracy under the 3D scene’s domaindiscrepancies. For example, PV-RCNN coupled with TQS only achieves 75.40% AP 3D, which largely falls behind the fully-supervised result 82.50% AP 3D. As a result, directly selecting a subset of given 3D frames using 2D ADA methods to tackle 3D scene’s do- main discrepancies is challenging, which can be attributedto the following reasons. (1) The sparsity of the 3D point clouds leads to huge inter-domain discrepancies that harm the discriminability of domain-related features. (2) Theintra-domain feature variations are widespread within the source domain, which enlarges the differentiation betweenthe selected target domain samples and the entire source do-main samples, bringing negative transfer to the model adap-tation on the target domain. To this end, we propose a Bi-domain active learning (Bi3D) framework to conduct the active learning for the 3D point clouds. To tackle the problem of sparsity ,w e design a foreground region-aware discriminator, which ex-ploits an RPN-based attention enhancement to derive aforeground-related domainness metric, that can be regarded as an important proxy for active sampling strategy. To address the problem of intra-domain feature variations within the source domain, we conceive a Bi-domain sam- pling approach, where Bi-domain means that data from bothsource and target domains are picked up for safe and robust model adaptation. Specifically, the Bi3D is composed of adomainness-aware source sampling strategy and a diversity- based target sampling strategy. The source sampling strat- egy aims to select target-domain-like samples from the source domain, by judging the corresponding domainnessscore of each given source sample. Then, the target sam-pling strategy is utilized to select diverse but representative data from the target domain by dynamically maintaining a similarity bank. Finally, we employ the sampled data fromboth domains to adapt the source pre-trained detector on anew target domain at a low annotation cost. The main contributions can be summarized as follows: 1. From a new perspective of chasing high performance at a low cost, we explore the possibilities of leverag- ing active learning to achieve effective 3D scene-levelcross-domain object detection. 2. A Bi-domain active sampling approach is proposed, consisting of a domainness-aware source samplingstrategy and a diversity-based target sampling strat- egy to identify the most informative samples from both source and target domains, boosting the model’s adap-tation performance. 3. Experiments show that Bi3D outperforms state-of-the- art UDA works with only 1% target annotation bud- get for cross-domain 3DOD. Moreover, Bi3D achieves89.63% AP BEV in the nuScenes-to-KITTI scenario, surpassing the fully supervised result ( 88.98% AP BEV) on the KITTI dataset.

Yong_A_General_Regret_Bound_of_Preconditioned_Gradient_Method_for_DNN_CVPR_2023

Abstract While adaptive learning rate methods, such as Adam, have achieved remarkable improvement in optimizing Deep Neural Networks (DNNs), they consider only the diago- nal elements of the full preconditioned matrix. Though the full-matrix preconditioned gradient methods theoretically have a lower regret bound, they are impractical for use to train DNNs because of the high complexity. In this paper, we present a general regret bound with a constrained full- matrix preconditioned gradient, and show that the updat- ing formula of the preconditioner can be derived by solving a cone-constrained optimization problem. With the block- diagonal and Kronecker-factorized constraints, a specific guide function can be obtained. By minimizing the up- per bound of the guide function, we develop a new DNN optimizer, termed AdaBK. A series of techniques, includ- ing statistics updating, dampening, efficient matrix inverse root computation, and gradient amplitude preservation, are developed to make AdaBK effective and efficient to imple- ment. The proposed AdaBK can be readily embedded into many existing DNN optimizers, e.g., SGDM and AdamW, and the corresponding SGDM BK and AdamW BK algo- rithms demonstrate significant improvements over existing DNN optimizers on benchmark vision tasks, including im- age classification, object detection and segmentation. The code is publicly available at https://github.com/ Yonghongwei/AdaBK .

1. Introduction Stochastic gradient descent (SGD) [26] and its vari- ants [21, 23], which update the parameters along the oppo- site of their gradient directions, have achieved great success in optimizing deep neural networks (DNNs) [14, 24]. In- stead of using a uniform learning rate for different parame- ters, Duchi et al. [5] proposed the AdaGrad method, which adopts an adaptive learning rate for each parameter, and proved that AdaGrad can achieve lower regret bound than SGD. Following AdaGrad, a class of adaptive learning rate gradient descent methods has been proposed. For example,RMSProp [30] and AdaDelta [35] introduce the exponential moving average to replace the sum of second-order statis- tics of the gradient for computing the adaptive learning rate. Adam [15] further adopts the momentum into the gradient, and AdamW [22] employs a weight-decoupled strategy to improve the generalization performance. RAdam [18], Ad- abelief [38] and Ranger [19,32,37] are proposed to acceler- ate training and improve the generalization capability over Adam. The adaptive learning rate methods have become the mainstream DNN optimizers. In addition to AdaGrad, Duchi et al. [5] provided a full- matrix preconditioned gradient descent (PGD) method that adopts the matrix HT= (PT t=1gtg⊤ t)1 2to adjust the gra- dientgT, where tdenotes the iteration number and Tis the number of the current iteration. It has been proved [5] that the preconditioned gradient H−1 TgThas a lower regret bound than the adaptive learning rate methods that only consider the diagonal elements of HT. However, the full-matrix preconditioned gradient is impractical to use due to its high dimension, which limits its applica- tion to DNN optimization. Various works have been re- ported to solve this problem in parameter space by adding some structural constraints on the full-matrix HT. For in- stances, GGT [1] stores only the gradients of recent itera- tions so that the matrix inverse root can be computed effi- ciently by fast low-rank computation tricks. Yun et al. [34] proposed a mini-block diagonal matrix framework to re- duce the cost through coordinate partitioning and grouping strategies. Gupta et al . [9] proposed to extend AdaGrad with Kronecker products of full-matrix preconditioners to make it more efficient in DNN training. Besides, natural gradient approaches [6, 7], which adopt the approximations of the Fisher matrix to correct the descent direction, can also be regarded as full-matrix preconditioners. The existing constrained PGD (CPGD) methods, how- ever, are heuristic since manually designed approximations to the full matrix HTare employed in them, while their in- fluence on the regret bound is unknown. By far, they lack a general regret-bound theory that can guide us to design the full-matrix preconditioned gradient methods. On the other hand, the practicality and effectiveness of these precondi- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7866 tioner methods are also an issue, which prevents them from being widely used in training DNNs. To address the above-mentioned issues, in this paper we present a theorem to connect the regret bound of the con- strained full-matrix preconditioner with a guide function. By minimizing the guide function under the constraints, an updating formula of the preconditioned gradient can be de- rived. That is, optimizing the guide function of the precon- ditioner will minimize its regret bound at the same time, while different constraints can yield different updating for- mulas. With the commonly-used constraints on DNN pre- conditioners, such as the block-diagonal and Kronecker- factorized constraints [7, 9], specific guide functions can be obtained. By minimizing the upper bound of the guide func- tion, a new optimizer, namely AdaBK, is derived. We further propose a series of techniques, including statistics updating, dampening, efficient matrix inverse root computation and gradient norm recovery, to make AdaBK more practical to use for DNN optimization. By embedding AdaBK into SGDM and AdamW (or Adam), we develop two new DNN optimizers, SGDM BK and AdamW BK. With acceptable extra computation and memory cost, they achieve significant performance gain in convergence speed and generalization capability over state-of-the-art DNN op- timizers, as demonstrated in our experiments in image clas- sification, object detection and segmentation. For a better understanding of our proposed regret bound and the developed DNN optimizer, in Fig. 1, we illustrate the existing major DNN optimizers and their relationships. SGD and its momentum version (SGDM) apply the same learning rate to all parameters based on their gradient de- scent directions. The adaptive learning rate methods as- sign different learning rates to different parameters by using second-order information of the gradients, achieving bet- ter convergence performance. The adaptive learning rate methods can be viewed as special cases of PGD methods by considering only the diagonal elements of the full pre- conditioned matrix of gradients. Our method belongs to the class of PGD methods, while our proposed general regret bound of constrained PGD methods can be applied to the PGD optimizers under different constraints, including Ada- Grad, Full-Matrix AdaGrad and our AdaBK. Notation system. We denote by wtandgtthe weight vector and its gradient of a DNN model in the t-th iteration. Denote by gt,ithe gradient of the i-th sample in a batch in the t-th iteration, we have gt=1 nPn i=1gt,i, where n is the batch size. The notations A⪰0andA≻0for a matrix Adenote that Ais symmetric positive semidef- inite (PSD) and symmetric positive definite, respectively. A⪰BorA−B⪰0means that A−Bis PSD. Tr (A) represents the trace of the matrix A. For a PSD matrix A, Aα=UΣαU⊤, where UΣU⊤is the Singular Value De- composition (SVD) of A.||x||A=√ x⊤Axis the Maha- SGD SGDM AdaGrad RMSProp Adam Adabelief Shampoo FM-AdaGrad KFAC AdaBKStochastic Gradient Descent Adaptive Learning Rate Method Preconditioned Gradient Descent Regret Bound Theorem of CPGD෥𝒈𝒈 𝑯−1𝒈 𝑯−1𝒈Diagonal Constraint General ConstraintsFigure 1. Illustration of the main DNN optimizers. lanobis norm of xinduced by PSD matrix A, and its dual norm is ||x||∗ A=√ x⊤A−1x.A⊗Bmeans the Kro- necker product of AandB, while A⊙BandA⊙αare the element-wise matrix product and element-wise power oper- ation, respectively. Diag (x)is a diagonal matrix with diag- onal vector x, and vec (·)denotes the vectorization function.

Zhang_PromptCAL_Contrastive_Affinity_Learning_via_Auxiliary_Prompts_for_Generalized_Novel_CVPR_2023

Abstract Although existing semi-supervised learning models achieve remarkable success in learning with unannotated in-distribution data, they mostly fail to learn on unlabeled data sampled from novel semantic classes due to their closed-set assumption. In this work, we target a prag- matic but under-explored Generalized Novel Category Dis- covery (GNCD) setting. The GNCD setting aims to cat- egorize unlabeled training data coming from known and novel classes by leveraging the information of partially la- beled known classes. We propose a two-stage Contrastive Affinity Learning method with auxiliary visual Prompts, dubbed PromptCAL , to address this challenging problem. Our approach discovers reliable pairwise sample affini- ties to learn better semantic clustering of both known and novel classes for the class token and visual prompts. First, we propose a discriminative prompt regularization loss to reinforce semantic discriminativeness of prompt-adapted pre-trained vision transformer for refined affinity relation- ships. Besides, we propose contrastive affinity learning to calibrate semantic representations based on our itera- tive semi-supervised affinity graph generation method for semantically-enhanced supervision. Extensive experimen- tal evaluation demonstrates that our PromptCAL method is more effective in discovering novel classes even with lim- ited annotations and surpasses the current state-of-the-art on generic and fine-grained benchmarks ( e.g., with nearly 11% gain on CUB-200, and 9%on ImageNet-100) on over- all accuracy. Our code is available at https : / / github . c om/ sheng- e atamath/ Pr omptCAL .

1. Introduction The deep neural networks have demonstrated favorable performance in the Semi-Supervised Learning (SSL) set- ting [15,29,40,46,49]. Some recent works can even achieve comparable performance to their fully-supervised counter- Figure 1. PromptCAL Overview. In contrast to previous method based on semi-supervised contrastive learning, PromptCAL con- structs affinity graph on-the-fly to guide representation learning of the class token and prompts. Meanwhile, our prompt-adapted backbone can be tuned to enhance semantic discriminativeness. PromptCAL can discover reliable affinities from a memory bank, especially for novel classes. Therefore, our PromptCAL is better task-aligned and discriminative to novel semantic information. parts using few annotations for image recognition [3,38,46]. However, these approaches heavily rely on the closed-world assumption that unlabeled data share the same underlying class label space as the labeled data [10, 47]. In many real- istic scenarios, this assumption does not hold true because of the inherent dynamism of real-world tasks where novel classes can emerge in addition to known classes. In contrast to SSL, the Novel Category Discovery (NCD) problem was introduced by [12] to relax the closed-world assumption of SSL, which assumes the unlabeled data con- tain novel classes. Recently, the nascent Generalized Novel Category Discovery (GNCD) problem, first proposed in [4, 41], extends NCD and assumes the unlabeled data can contain both known and novel classes, which is more prag- matic and challenging. To be more specific, GNCD intends to categorize images sampled from predefined categories in the training set comprising labeled-knowns, unlabeled- knowns, and unlabeled-novels. 1 This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3479 Our work focuses on GNCD problem. The key challenge of GNCD is to discriminate among novel classes when only the ground truths of known classes are accessible in the training set. Recent studies show that self-supervised pre- trained representations are conducive to discovering novel semantics [4, 5, 11, 41, 53]. A typical work on GNCD [41] takes advantage of the large-scale pre-trained visual trans- former (ViT) [37], and learns robust clusters for known and novel classes through semi-supervised contrastive learning on downstream datasets. However, we discover that the re- markable potential of pre-trained ViT is actually suppressed by this practice, due to the class collision [52] issue induced by abundant false negatives in contrastive loss, i.e., consid- ering different unlabeled images from the same or similar semantic class as false negatives. As supported by empiri- cal studies, abundant false negatives in contrastive training can deteriorate the compactness and purity of semantic clus- tering [5, 16, 21, 52]. Based on empirical investigation, we show that this issue is particularly severe in category discov- ery. Furthermore, although the existing commonly adopted practice [4, 41] of freezing most parts of the pre-trained backbone can alleviate overfitting on known classes, it con- strains the flexibility and adaptability of backbones [18]. Lack of adaptability inhibits models from learning discrim- inative semantic information on downstream datasets. To address above limitations and learn better semanti- cally discriminative representations, we propose Prompt - based Contrastive Affinity Learning ( PromptCAL ) frame- work to tackle GNCD problem. To be specific, our ap- proach aims to discover semantic clusters in unlabeled data by simultaneous semantic prompt learning based on our Discriminative Prompt Regularization (DPR) loss and rep- resentation calibration based on our Contrastive Affinity Learning (CAL) process. Firstly , CAL discovers abundant reliable pseudo positives for DPR loss and contrastive loss based on generated affinity graphs. These semantic-aware pseudo labels further enhance the semantic discriminative- ness of DPR supervision. Secondly , DPR regularizes se- mantic representations of ensembled prompts, which facil- itates the discovery of more accurate pseudo labels at the next-step of CAL. Therefore, as model and prompt repre- sentations are iteratively enhanced, we can obtain higher quality pseudo positives for further self-training as well as acquire better semantic clustering. Our PromptCAL achieves State-Of-The-Art (SOTA) performance in extensive experimental evaluation on six benchmarks. Specifically, PromptCAL remarkably sur- passes previous SOTA by more than 10% clustering ac- curacy on the fine-grained CUB-200 and StandfordCars datasets; it also significantly outperforms previous SoTAs by nearly 4%on ImageNet-100 and 7%on CIFAR-100. In- terestingly, we identify that both DPR supervised prompts and unsupervised prompts of PromptCAL can learn se-mantic discriminativeness, which advances the flexibility of the pre-trained backbone. Furthermore, PromptCAL still achieves the best performance in challenging low-labeling and few-class setups. Our contributions include: (1)We propose a two-stage framework for the generalized novel category discovery problem, in which semantic prompt tuning and contrastive affinity learning mutually reinforce and benefit each other during the learning process. (2)We propose two syn- ergistic learning objectives, contrastive affinity loss and discriminative prompt regularization loss, based on our semi-supervised adapted affinity graphs to enhance seman- tic discriminativeness. (3)We comprehensively evaluate our method on three generic ( i.e., CIFAR-10, CIFAR-100, and ImageNet-100) and three fine-grained benchmarks ( i.e., CUB-200, Aircraft, and StandfordCars), achieving state-of- the-art performance, thereby showing its effectiveness. (4) We further showcase generalization ability of PromptCAL and its effectiveness in more challenging low-labeling and few-class setups.

Zhang_Semi-DETR_Semi-Supervised_Object_Detection_With_Detection_Transformers_CVPR_2023

Abstract We analyze the DETR-based framework on semi- supervised object detection (SSOD) and observe that (1) the one-to-one assignment strategy generates incorrect match- ing when the pseudo ground-truth bounding box is inaccu- rate, leading to training inefficiency; (2) DETR-based de- tectors lack deterministic correspondence between the in- put query and its prediction output, which hinders the ap- plicability of the consistency-based regularization widely used in current SSOD methods. We present Semi-DETR, the first transformer-based end-to-end semi-supervised ob- ject detector, to tackle these problems. Specifically, we propose a Stage-wise Hybrid Matching strategy that com- bines the one-to-many assignment and one-to-one assign- ment strategies to improve the training efficiency of the first stage and thus provide high-quality pseudo labels for the training of the second stage. Besides, we introduce a Cross- view Query Consistency method to learn the semantic fea- ture invariance of object queries from different views while avoiding the need to find deterministic query correspon- dence. Furthermore, we propose a Cost-based Pseudo La- bel Mining module to dynamically mine more pseudo boxes based on the matching cost of pseudo ground truth bound- ing boxes for consistency training. Extensive experiments on all SSOD settings of both COCO and Pascal VOC bench- mark datasets show that our Semi-DETR method outper- forms all state-of-the-art methods by clear margins.

1. Introduction Semi-supervised object detection (SSOD) aims to boost the performance of a fully-supervised object detector by ex- *Equally-contributed authors. Work done during an internship at Baidu. †Corresponding author. SupervisionWeakAug StrongAugBipartiteMatching Stage-wiseHybridMatchingEMA…………DETR(a)DETR-SSODvanilla(b)Semi-DETRDETR……TeacherDETR ……TeacherDETR StudentCross-viewqueryConsistencyWeakAug StrongAugEMA SupervisionStudentFigure 1. Comparisons between the vanilla DETR-SSOD frame- work based on the Teacher-Student architecture and our proposed Semi-DETR framework. Semi-DETR consists of Stage-wise Hy- brid Matching, Cross-view Query Consistency powered by a cost- based pseudo label mining strategy. ploiting a large amount of unlabeled data. Current state-of- the-art SSOD methods are primarily based on object detec- tors with many hand-crafted components, e.g., rule-based label assigner [9,26,27,31] and non-maximum suppression (NMS) [1] post-processing. We term this type of object de- tector as a traditional object detector. Recently, DETR [2], a simple transformer-based end-to-end object detector, has received growing attention. Generally, the DETR-based framework builds upon transformer [32] encoder-decoder architecture and generates unique predictions by enforcing a set-based global loss via bipartite matching during training. It eliminates the need for various hand-crafted components, achieving state-of-the-art performance in fully-supervised object detection. Although the performance is desirable, how to design a feasible DETR-based SSOD framework re- mains under-explored. There are still no systematic ways to fulfill this research gap. Designing an SSOD framework for DETR-based de- tectors is non-trivial. Concretely, DETR-based detectors take a one-to-one assignment strategy where the bipartite- matching algorithm forces each ground-truth (GT) bound- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23809 1% 5% 10% 100% Proportion of used labeled data152025303540455055COCO mAP 22.432.536.046.1 22.333.037.146.1 30.540.143.551.8 PseCo(Faster RCNN) Dense Teacher(FCOS) Semi-DETR(DINO)Figure 2. Performance comparisons between the proposed Semi- DETR and other SSOD methods, including PseCo [16] and Dense- Teacher [42]. ing box to match a candidate proposal as positive, treating remains as negatives. It goes well when the ground-truth bounding boxes are accurate. However, directly integrat- ing DETR-based framework with SSOD is problematic, as illustrated in Fig. 1 (a) where a DETR-SSOD vanilla frame- work utilizes DETR-based detectors to perform pseudo la- beling on unlabeled images. In the Teacher-Student archi- tecture, the teacher model usually generates noisy pseudo bounding boxes on the unlabeled images. When the pseudo bounding box is inaccurate, the one-to-one assignment strategy is doomed to match a single inaccurate proposal as positive, leaving all other potential correct proposals as negative, thus causing learning inefficiency. As a compar- ison, the one-to-many assignment strategy adopted in the traditional object detectors maintains a set of positive pro- posals, having a higher chance of containing the correct positive proposal. On the one hand, the one-to-one assign- ment strategy enjoys the merits of NMS-free end-to-end detection but suffers the training inefficiency under semi- supervised scenarios; on the other hand, the one-to-many assignment strategy obtains candidate proposal set with bet- ter quality making the detector optimized more efficiently but inevitably resulted in duplicate predictions. Designing a DETR-based SSOD framework that embraces these two merits could bring the performance to the next level. Additionally, the consistency-based regularization com- monly used in current SSOD methods becomes infeasible in DETR-based SSOD. Specifically, current SSOD meth- ods [3, 10, 13, 16] utilize consistency-based regulariza- tion to help object detectors learn potential feature invari- ance by imposing consistency constraints on the outputs of pairs-wise inputs (such as scale consistency [3, 10, 16], weak-strong consistency [13], etc.). Since the input fea- tures are deterministic in traditional object detectors, there is a one-to-one correspondence between the inputs and out- puts, which makes the consistency constraint convenientto implement. However, this is not the case in DETR- based detectors. DETR-based detectors [2, 15, 20, 40, 44] use randomly initialized learnable object queries as inputs and constantly update the query features through the atten- tion mechanism. As the query features update, the corre- sponding prediction results constantly change, which has been verified in [15]. In other words, there is no determinis- tic correspondence between the input object queries and its output prediction results, which prevents consistency regu- larization from being applied to DETR-based detectors. According to the above analysis, we propose a new DETR-based SSOD framework based on the Teacher- Student architecture, which we term Semi-DETR presented in Fig. 1 (b). Concretely, we propose a Stage-wise Hybrid Matching module that imposes two stages of training us- ing the one-to-many assignment and the one-to-one assign- ment, respectively. The first stage aims to improve the train- ing efficiency via the one-to-many assignment strategy and thus provide high-quality pseudo labels for the second stage of one-to-one assignment training. Besides, we introduce aCross-view Query Consistency module that constructs cross-view object queries to eliminate the requirement of finding deterministic correspondence of object queries and aids the detector in learning semantically invariant charac- teristics of object queries between two augmented views. Furthermore, we devise a Cost-based Pseudo Label Min- ingmodule based on the Gaussian Mixture Model (GMM) that dynamically mines reliable pseudo boxes for consis- tency learning according to their matching cost distribution. Differently, Semi-DETR is tailored for DETR-based frame- work, which achieves new SOTA performance compared to the previous best SSOD methods as illustrated in Fig. 2. To sum up, this paper has the following contributions: • We present a new DETR-based SSOD method based on the Teacher-Student architecture, called Semi- DETR. To our best knowledge, we are the first to ex- amine the DETR-based detectors on SSOD, and we identify core issues in integrating DETR-based detec- tors with the SSOD framework. • We propose a stage-wise hybrid matching method that combines the one-to-many assignment and one-to-one assignment strategies to address the training ineffi- ciency caused by the inherent one-to-one assignment within DETR-based detectors when applied to SSOD. • We introduce a consistency-based regularization scheme and a cost-based pseudo-label mining algo- rithm for DETR-based detectors to help learn semantic feature invariance of object queries from different aug- mented views. • Extensive experiments show that our Semi-DETR method outperforms all previous state-of-the-art meth- ods by clear margins under various SSOD settings on both MS COCO and Pascal VOC benchmark datasets. 23810

Yu_Efficient_Loss_Function_by_Minimizing_the_Detrimental_Effect_of_Floating-Point_CVPR_2023

Abstract Attackers can deceive neural networks by adding hu- man imperceptive perturbations to their input data; this reveals the vulnerability and weak robustness of current deep-learning networks. Many attack techniques have been proposed to evaluate the model’s robustness. Gradient- based attacks suffer from severely overestimating the ro- bustness. This paper identifies that the relative error in cal- culated gradients caused by floating-point errors, including floating-point underflow and rounding errors, is a funda- mental reason why gradient-based attacks fail to accurately assess the model’s robustness. Although it is hard to elim- inate the relative error in the gradients, we can control its effect on the gradient-based attacks. Correspondingly, we propose an efficient loss function by minimizing the detri- mental impact of the floating-point errors on the attacks. Experimental results show that it is more efficient and reli- able than other loss functions when examined across a wide range of defence mechanisms.

1. Introduction AI with deep neural networks (DNNs) as the core [24] has achieved great success in different research directions and has been widely used in many safety-critical systems, such as aviation [1, 8, 35], medical diagnosis [27, 49], self- driving [4,26,47] , etc. However, a severe problem with cur- rent DNNs is that they are vulnerable to adversarial attacks. Adding human imperceptive perturbations to input data can mislead the model to output incorrect results [16, 45]. Such vulnerability poses a significant security risk to all DNN- based systems. Therefore, increasing the models’ robust- ness and providing efficient and reliable evaluation methods are becoming increasingly urgent and vital. *Corresponding author. 1https://robustbench.github.io/ 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 Iterations (k)53.053.554.054.555.0Robust Accuracy (%)MIFPE MIFPECE CE target C&W C&W target DLRDLR target MIFPE MIFPE target BestFigure 1. Comparison of the effectiveness of non-targeted and multi-targeted PGD 100 iterations attacks using various loss func- tions on the CIFAR-10 dataset. The best results were obtained from RobustBench1using an ensemble attack with a minimum of 4900 iterations. The defence model is from [59]. Numerous defence strategies [6, 10, 13, 16, 29, 31, 55] to improve the model’s robustness and many attacks [2, 29, 33, 41, 50, 52, 54, 59] to evaluate these strategies have been proposed. Currently, the most efficient attack method is known as white-box attack, where the attacker has com- plete knowledge of the target defending strategy, includ- ing its model architecture, parameters, detail of the train- ing algorithm, dataset, etc. A typical example is Projec- tive gradient descent (PGD) [29] for evaluating the model’s robustness. However, studies have shown that PGD with cross-entropy (CE) loss fails to provide accurate evaluation results [6, 41] and significantly overestimates the model’s robustness(Figure 1). To enhance the evaluation accuracy, recent powerful evaluation methods were proposed to en- semble diverse attack algorithms [10, 30]. However, their performance is at the cost of high computational overhead. The research community attributes the failure of gradient-based attacks to gradient masking [15]. The core of gradient masking is that the calculated gradients are not This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4056 necessarily efficient in guiding the generation of adversarial examples. The causes of inefficient gradients may be related to floating-point errors or other factors. This paper reveals that the relative error in the gradients caused by floating- point errors is one of the fundamental reasons for the failure of gradient-based attacks, and the relative error is directly affected by the value of the difference of the f

Zhang_Class_Relationship_Embedded_Learning_for_Source-Free_Unsupervised_Domain_Adaptation_CVPR_2023

Abstract This work focuses on a practical knowledge transfer task defined as Source-Free Unsupervised Domain Adaptation (SFUDA), where only a well-trained source model and un- labeled target data are available. To fully utilize source knowledge, we propose to transfer the class relationship, which is domain-invariant but still under-explored in previ- ous works. To this end, we first regard the classifier weights of the source model as class prototypes to compute class re- lationship, and then propose a novel probability-based sim- ilarity between target-domain samples by embedding the source-domain class relationship, resulting in Class Rela- tionship embedded Similarity (CRS). Here the inter-class term is particularly considered in order to more accurately represent the similarity between two samples, in which the source prior of class relationship is utilized by weighting. Finally, we propose to embed CRS into contrastive learn- ing in a unified form. Here both class-aware and instance discrimination contrastive losses are employed, which are complementary to each other. We combine the proposed method with existing representative methods to evaluate its efficacy in multiple SFUDA settings. Extensive experimen- tal results reveal that our method can achieve state-of-the- art performance due to the transfer of domain-invariant class relationship.1

1. Introduction Benefiting from the large amount of labeled training data, deep neural networks have achieved promising results in many computer vision tasks [7, 15, 19, 95]. To reduce the annotation cost, Unsupervised Domain Adaptation (UDA) has been devised by transferring knowledge from a label- rich source domain to a label-scarce target domain. Cur- rently, many UDA methods have been proposed that jointly *Corresponding author 1Code is available at https://github.com/zhyx12/CRCo Sample feature Classifier weight Class 1 Class 2 Class 3 Class- aware contrastive loss Instance discrimination contrastive loss Class relationship embedded contrastive learning 𝒘𝒘𝑖𝑖(𝑖𝑖∈1,2,3)∈ℝD×1: normalized source classifier weight of i-thclass 𝑆𝑆cr𝒑𝒑,𝒑𝒑′=� 𝑖𝑖=13 𝑝𝑝𝑖𝑖×𝑝𝑝𝑖𝑖′+� 𝑖𝑖,𝑗𝑗=1;𝑖𝑖≠𝑗𝑗3 𝐴𝐴𝑖𝑖𝑗𝑗𝑠𝑠×𝑝𝑝𝑖𝑖×𝑝𝑝𝑗𝑗′ Intra -class similarity Inter-class similarity Pull together Push away Source pre -trained model Our method 𝑨𝑨𝑠𝑠= 𝒘𝒘1𝑻𝑻𝒘𝒘1𝒘𝒘1𝑻𝑻𝒘𝒘2𝒘𝒘1𝑻𝑻𝒘𝒘3 𝒘𝒘2𝑻𝑻𝒘𝒘1𝒘𝒘2𝑻𝑻𝒘𝒘2𝒘𝒘2𝑻𝑻𝒘𝒘3 𝒘𝒘3𝑻𝑻𝒘𝒘1𝒘𝒘3𝑻𝑻𝒘𝒘𝟐𝟐𝒘𝒘3𝑻𝑻𝒘𝒘3=1 0.10.02 0.1 1 0.02 0.020.02 1Figure 1. Illustration of our proposed method. The upper left rep- resents the target-domain featrue distribution by source pre-trained model. The upper right is the feature distribution obtained by our method. The bottom is the process of class relationship embedded contrastive learning. Best viewed in color. learn on the source and target data. But they would be unapplicable for some real-world scenarios involving pri- vacy ( e.g., medical images, surveillance videos) because the source-domain data cannot be accessed. Thus, more recent methods [36, 42, 71, 79, 87, 88] focus on Source-Free Un- supervised Domain Adaptation (SFUDA). Under this set- ting, the labeled source data are not accessible any more when training the target model, but the pre-trained model in the source domain is provided. Then a natural question arises, i.e., what knowledge should we transfer to facilitate the learning of unlabeled target-domain data? Some methods [21, 36, 42] assume that the source hy- pothesis ( i.e., classifier [42] or whole model [21, 36]) con- tains sufficient knowledge for target domain. Then they transfer source knowledge by directly aligning features with the fixed source classifier [42], resorting to historical mod- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7619 els [21], or weight regularization [36]. Another line of works [87, 89, 90] assume the source model already forms some semantic structure, and then utilize the local infor- mation ( i.e., neighborhood) of feature space to enforce the similarity in the output space. Despite these progress, what knowledge to be transferred remains an open question. In this work, we propose to transfer the class relation- ship represented by the similarity of classifier weights. Ac- tually, the class relationship is domain-invariant [14], since the same class in the source and target domains essentially represents the same semantic in spite of domain discrep- ancy. For example, the computers are always more simi- lar with the TV than the scissors. Thus, it is reasonable to guide the target domain learning using class relationship prior. Unlike previous methods that learn class distribu- tion by pseudo labeling [21, 42] and local aggregation of neighborhood predictions [79, 87, 89, 90], here we explic- itly model the source-domain class relationship. To be spe- cific, we regard each weight of classifier as the class proto- type [90], and then compute a class relationship matrix As by cosine similarity, as shown in Figure 1. Before explaining how to use this matrix, we illustrate the purpose of representation learning in the target domain using Figure 1, where three classes are adopted for clarity. The top left shows the target-domain feature distribution to- gether with the class weights learned from the source do- main. It can be seen that such a situation makes it difficult to perform correct classification. In fact, it is expected that the learned features are discriminative and compact around the corresponding class weight, as shown on the top right. To this end, an intuitive way is to make the relationship of learned class prototypes in the target domain consistent with that in the source domain. However, it has a very lim- ited effect on training the target-domain model as such a prototype-level constraint is too weak for optimization. In this work, we propose to embed the source-domain class relationship in contrastive learning, which has been shown to be outstanding in representation learning [9, 18]. Here we design a novel sample similarity by taking into ac- countAs. Specifically, we compute the similarity between two target domain samples in the output space, represented by the prediction probabilities ( i.e.,pandp′), as shown in Figure 1. In particular, we consider the inter-class term ( i.e.P i̸=jpip′ j) in addition to the traditional intra-class term (i.e.P ipip′ i). Note that the sum of these two terms equals one, in which the intra-class term measures the similarity that two samples come from the same class, while the inter- class term measures that from different classes. Consider- ing the relationship between classes, we weight the inter- class term by the non-diagonal elements in As. In partic- ular, if the class iis closer to jthan k(i.e.,As ij> As ik), pip′ jwould be more important in calculating the similarity ofpandp′. By the above design, our proposed similaritycan more accurately express the relationship of two samples based on output space. For example, among three classes 1,2,3in Figure 1, the classes 1and2are closer. Given three samples x1, x2, x3belonging to them with the probabilities [0.9,0.05,0.05],[0.05,0.9,0.05], and [0.05,0.05,0.9]. If we only use the intra-class term, all three samples have the same similarity. But for our designed similarity, x1is closer tox2thanx3, which is more reasonable. On the basis of the proposed similarity, we further pro- pose to perform contrastive learning in a unified form. In- spired by recent success in semi-supervised learning [67, 85, 91] and unsupervised learning [9, 18], we propose two types of contrastive losses. As shown in Figure 1, the first one is ClassRelationship embedded Class-Aware Contrastive (CR-CACo) loss where the high-confident sam- ples are enforced to be close to the corresponding proto- type and away from other prototypes. Due to embedding the prior class relationship, our CR-CACo loss is more ro- bust to label noise caused by domain shift. The second one isClass Relationship embedded Instance Discrimination Contrastive (CR-IDCo) loss where two views of the same sample are encouraged to be close and away from all other samples. Benefited from our designed accurate similarity, the CR-IDCo loss would more effectively learn discrimina- tive features [9, 18, 75]. Actually, these two losses are com- plementary to each other, and their combination can achieve better performance. Our contributions are summarized as follows: • We propose to explicitly transfer class relationship for SFUDA which is more domain-invariant. And we pro- pose to embed the class relationship into contrastive learning in order to effectively perform representation learning in the target domain. • We propose a novel class relationship embedded sim- ilarity which can more accurately express the sample relationship in the output space. Furthermore, we pro- pose two contrastive losses ( i.e., CR-CACo and CR- IDCo) exploiting our designed similarity. • We conduct extensive experiments to evaluate the pro- posed method, and the results validate the effective- ness of our method, which achieves the state-of-the-art performance on multiple SFUDA benchmarks.

Yu_Distribution_Shift_Inversion_for_Out-of-Distribution_Prediction_CVPR_2023

Abstract Machine learning society has witnessed the emergence of a myriad of Out-of-Distribution (OoD) algorithms, which address the distribution shift between the training and the testing distribution by searching for a unified predictor or invariant feature representation. However, the task of di- rectly mitigating the distribution shift in the unseen testing set is rarely investigated, due to the unavailability of the testing distribution during the training phase and thus the impossibility of training a distribution translator mapping between the training and testing distribution. In this pa- per, we explore how to bypass the requirement of testing distribution for distribution translator training and make the distribution translation useful for OoD prediction. We propose a portable Distribution Shift Inversion (DSI) algo- rithm, in which, before being fed into the prediction model, the OoD testing samples are first linearly combined with ad- ditional Gaussian noise and then transferred back towards the training distribution using a diffusion model trained only on the source distribution. Theoretical analysis reveals the feasibility of our method. Experimental results, on both multiple-domain generalization datasets and single-domain generalization datasets, show that our method provides a general performance gain when plugged into a wide range of commonly used OoD algorithms. Our code is available at https://github.com/yu-rp/Distribution-Shift-Iverson.

1. Introduction The ubiquity of the distribution shift between the training and testing data in the real-world application of machine †Corresponding author.learning systems induces the study of Out-of-Distribution (OoD) generalization (or domain generalization). [18, 64, 71, 79] Within the scope of OoD generalization, machine learning algorithms are required to generalize from the seen training domain to the unseen testing domain without the independent and identically distributed assumption. The bulk of the OoD algorithms in previous literature focuses on promoting the generalization capability of the machine learning models themselves by utilizing the domain invariant feature [2, 13, 36], context-based data augmentation [47, 74], distributionally robust optimization [59], subnetwork searching [81], neural network calibration [70], etc. In this work, orthogonal to enhancing the generalization capability of the model, we consider a novel pathway to OoD prediction. On the way, the testing(target) distribution is explicitly transformed towards the training(source) dis- tribution to straightforwardly mitigate the distribution shift between the testing and the training distribution. Therein, the OoD prediction can be regarded as a two-step procedure, (1) transferring testing samples back towards training dis- tribution, and (2) drawing prediction. The second step can be implemented by any OoD prediction algorithm. In this paper, we concentrate on the exploration of the first step, the distribution transformation. Unlike previous works on distribution translation and do- main transformation, in which certain target distribution is accessible during the training phase, here the target distri- bution is arbitrary and unavailable during the training. We term this new task as Unseen Distribution Transformation (UDT), in which a domain translator is trained on the source distribution and works to transform unseen target distribu- tion towards the source distribution. The uniqueness, as well as the superiority, of UDT is listed as followings. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3592 /uni00000057/uni00000003/uni00000020/uni00000003/uni00000013 /uni00000014/uni00000013/uni00000013 /uni00000015/uni00000013/uni00000013 /uni00000057/uni00000003/uni00000020/uni00000003/uni00000016/uni0000001a/uni00000013 /uni00000027/uni0000004c/uni00000049/uni00000049/uni00000058/uni00000056/uni0000004c/uni00000052/uni00000051/uni00000003/uni00000030/uni00000052/uni00000047/uni00000048/uni0000004f/uni00000003/uni00000037/uni00000055/uni00000044/uni00000051/uni00000056/uni00000049/uni00000055/uni00000052/uni00000050/uni00000044/uni00000057/uni0000004c/uni00000052/uni00000051/uni00000026/uni0000004f/uni00000044/uni00000056/uni00000056/uni00000003/uni000010de /uni00000026/uni0000004f/uni00000044/uni00000056/uni00000056/uni00000003/uni000010df /uni00000013/uni00000011/uni00000014 /uni00000013 /uni00000037 /uni00000048/uni00000056/uni00000057/uni0000004c/uni00000051/uni0000004a /uni00000027/uni0000004c/uni00000056/uni00000057/uni00000055/uni0000004c/uni00000045/uni00000058/uni00000057/uni0000004c/uni00000052/uni00000051/uni00000010/uni00000014/uni00000018/uni00000010/uni00000014/uni00000013/uni00000010/uni00000018/uni00000013/uni00000018/uni00000014/uni00000013/uni00000014/uni00000018/uni00000032/uni00000052/uni00000027/uni00000003/uni00000036/uni00000044/uni00000050/uni00000053/uni0000004f/uni00000048/uni00000056 /uni00000013 /uni00000013/uni00000011/uni00000014 /uni00000013/uni00000011/uni00000015 /uni00000013/uni00000011/uni00000016 /uni00000036/uni00000052/uni00000058/uni00000055/uni00000046/uni00000048/uni00000003/uni00000027/uni0000004c/uni00000056/uni00000057/uni00000055/uni0000004c/uni00000045/uni00000058/uni00000057/uni0000004c/uni00000052/uni00000051/uni00000031/uni00000052/uni0000004c/uni00000056/uni00000048/uni00000003/uni00000036/uni00000053/uni00000044/uni00000046/uni00000048 /uni00000024/uni0000004f/uni0000004c/uni0000004a/uni00000051/uni00000050/uni00000048/uni00000051/uni00000057/uni00000037/uni00000055/uni00000044/uni00000051/uni00000056/uni00000049/uni00000052/uni00000055/uni00000050/uni00000048/uni00000047 /uni00000036/uni00000044/uni00000050/uni00000053/uni0000004f/uni00000048/uni00000056Figure 2. Using a Diffusion Model to solve the one-dimensional UDT. OoD samples are transformed to the source distribution with limited failure of label preservation. •UDT puts no requirements on the data from both source and testing distribution like previous works do. This is practically valuable, because the real-world testing dis- tributions are uncountable and dynamically changing. •UDT is able to transform various distributions using only one model. However, the previous distribution translator works for the translation between certain source and target distributions. With a different source- target pair, a new translator is required. •Considering the application of UDT in OoD prediction, it is free from the extra assumptions commonly used by the OoD generalization algorithms, such as the multi- training domain assumption and the various forms of the domain invariant assumption. Despite the advantages, the unavailability of the testing distribution poses new difficulty. Releasing this constraint, the idea of distribution alignment is well established in do- main adaptation (DA). Wherein, a distribution translator is trained with the (pixel, feature, and semantic level) cycle consistency loss [23, 38, 82]. However, the training of such distribution transfer modules necessitates the testing distribu- tion, which is unsuitable under the setting of OoD prediction and makes the transplant of the methods in DA to OoD even impossible. To circumvent the requirement of testing distribution dur- ing training time, we propose a novel method, named Distri- bution Shift Inversion (DSI). Instead of using a model trans- ferring from testing distribution to training distribution, an unconditional generative model, trained only on the source distribution, is used, which transfers data from a reference noise distribution to the source distribution. The method operates in two successive parts. First, the OoD target dis- tribution is transferred to the neighborhood of the noise dis- tribution and aligned with the input of the generative model, thereafter we refer to this process as the forward transforma- tion. The crux of this step is designing to what degree the target distribution is aligned to the noise distribution. In our implementation, the forward transformation is conducted by linearly combining the OoD samples and random noise withthe weights controlling the alignment.Then, in the second step, the outcome of the first step is transferred towards the source distribution by a generative model, thereafter we refer to this process as the backward transformation. In this paper, the generative model is chosen to be the diffusion model [21]. The superiority of the diffusion model is that its input is the linear combination of the source sample and the noise with varying magnitude, which is in accord with our design of the forward transformation and naturally allows strength control. Comparatively, V AE [28] and GAN [16] have a fixed level of noise in their input, which makes the forward transfor- mation strength control indirect. Our theoretical analyses of the diffusion model also show the feasibility of using the diffusion model for UDT. Illustrative Example. A one-dimensional example is shown in Fig. 2.The example considers a binary classifica- tion problem, in which, given label, the conditional distribu- tions of the samples are Gaussian in both the source and the testing domain. The testing distribution is constructed to be OoD and located in the region where the source distribution has a low density. The diffusion model is trained only on the source distribution. Passing through the noise space align- ment and diffusion model transformation, the OoD samples are transformed to the source distribution with limited failure of label preservation. Transformed Images. Fig. 1 shows some transformation results of OoD images towards the source distribution. The observation is twofold. (1) The distribution (here is the style) of the images is successfully transformed. All of the transferred images can be correctly classified by the ERM model trained on the source domain. (2) The transformed images are correlated to the original images. Some structural and color characteristics are mutually shared between them. This indicates that the diffusion model has extracted some low-level information and is capable to preserve it during the transformation. We would like to highlight again that, during the training, the diffusion model is isolated from the testing domain. Our contributions are therefore summarized as: •We put forward the unseen distribution transformation (UDT) and study its application to OoD prediction. •We offer theoretical analyses of the feasibility of UDT. •we propose DSI, a sample adaptive distribution trans- formation algorithm for efficient distribution adaptation and semantic information preservation. •We perform extensive experiments to demonstrate that our method is suitable for various OoD algorithms to achieve performance gain on diversified OoD bench- marks. On average, adding in our method produces 2.26% accuracy gain on multi-training domain general- ization datasets and 2.28% on single-training domain generalization datasets. 3593

Zhang_Multi-View_Stereo_Representation_Revist_Region-Aware_MVSNet_CVPR_2023

Abstract Deep learning-based multi-view stereo has emerged as a powerful paradigm for reconstructing the complete geometrically-detailed objects from multi-views. Most of the existing approaches only estimate the pixel-wise depth value by minimizing the gap between the predicted point and the intersection of ray and surface, which usually ig- nore the surface topology. It is essential to the textureless regions and surface boundary that cannot be properly re- constructed. To address this issue, we suggest to take ad- vantage of point-to-surface distance so that the model is able to perceive a wider range of surfaces. To this end, we predict the distance volume from cost volume to esti- mate the signed distance of points around the surface. Our proposed RA-MVSNet is patch-awared, since the percep- tion range is enhanced by associating hypothetical planes with a patch of surface. Therefore, it could increase the completion of textureless regions and reduce the outliers at the boundary. Moreover, the mesh topologies with fine de- tails can be generated by the introduced distance volume. Comparing to the conventional deep learning-based multi- view stereo methods, our proposed RA-MVSNet approach obtains more complete reconstruction results by taking ad- vantage of signed distance supervision. The experiments on both the DTU and Tanks & Temples datasets demonstrate that our proposed approach achieves the state-of-the-art re- sults.

1. Introduction Multi-view stereo (MVS) is able to efficiently recover geometry from multiple images, which makes use of the matching relationship and stereo correspondences of over- lapping images. To achieve the promising reconstruction results, the con- ventional patch-based and PatchMatch-based methods [2, 11, 27] require rich textures and restricted lighting condi- *Corresponding author is Jianke Zhu. Hypothetical plane point Baseline RA-MVSNetAssociated surface point Figure 1. Comparison on reconstruction results between base- line and RA-MVSNet. Our RA-MVSNet enables the model to perceive a wider range of surfaces so as to achieve the promising performance in complementing textureless regions and removing outliers at boundaries. Furthermore, our model is able to gener- ate correct mesh topologies with fine details based on estimated point-to-surface distances of spatially sampled points. tions. Alternatively, the deep learning-based approaches [4, 14,15,40] try to take advantage of global scene semantic in- formation, including environmental illumination and object materials, to maintain high performance in complex light- ing. The key of these methods is to warp deep image fea- tures into the reference camera frustum so that the 3D cost volume can be built via differentiable homographies. Then, the depth map is predicted by regularizing cost volume with 3D CNNs. Despite the encouraging results, the pixel-wise depth es- timation suffers from two intractable flaws. One is the low estimation confidence in the textureless area. The other is many outliers near the boundary of the object. This is mainly because the surface is usually treated as a set of un- correlated sample points rather than the one with topology. As each ray is only associated with a single surface sam- pling point, it is impossible to pay attention to the adjacent area of the surface. As shown in Fig. 1, the estimation of each depth value is constrained by only one surface sample point, which makes it unable to use the surrounding surface for inference. Unfortunately, it is difficult to infer without This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17376 broader surface information in textureless regions and ob- ject boundaries. Therefore, too small perception range lim- its the existing learning-based MVS methods. To tackle this issue, we present a novel RA-MVSNet framework that is able to make each hypothetical plane as- sociated with a wider surface area through point-to-surface distance. Thus, our presented method is capable of inferring the surrounding surface information at textureless areas and object boundaries. To this end, our network not only es- timates the probability volume but also predicts the point- to-surface distance of each hypothetical plane. Specifically, RA-MVSNet makes use of the cost volume to generate the probability and distance volumes, which are further com- bined to estimate the final depth map. The introduction of point-to-surface distance supervision uses the model patch- aware in estimating the depth value corresponding to a par- ticular pixel. This leads to the improved performance in textureless or boundary areas. Since the distance volume estimates the length of the sample points near the surface, we are able to predict a SDF-based implicit representation with the correct topology and fine details. In summary, our contribution is three-fold: • We introduce point-to-surface distance supervision of sampled points to expand the perception range pre- dicted by the model, which achieves complete estima- tion in textureless areas and reduce outliers in object boundary regions. • To tackle the challenge of lacking the ground-truth mesh, we compute the signed distance between point sets based on the triangulated mesh, which trades off between accuracy and speed. • Experimental results on the challenging MVS datasets show that our proposed approach performs the best both on indoor dataset DTU [1] and large-scale out- door dataset Tanks and Temples [17].

Yan_Linking_Garment_With_Person_via_Semantically_Associated_Landmarks_for_Virtual_CVPR_2023

Abstract In this paper, a novel virtual try-on algorithm, dubbed SAL-VTON, is proposed, which links the garment with the person via semantically associated landmarks to alleviate misalignment. The semantically associated landmarks are a series of landmark pairs with the same local seman- tics on the in-shop garment image and the try-on image. Based on the semantically associated landmarks, SAL- VTON effectively models the local semantic association between garment and person, making up for the misalign- ment in the overall deformation of the garment. The outcome is achieved with a three-stage framework: 1) the semantically associated landmarks are estimated using the landmark localization model; 2) taking the landmarks as input, the warping model explicitly associates the corre- sponding parts of the garment and person for obtaining the local flow, thus refining the alignment in the global flow; 3) finally, a generator consumes the landmarks to better capture local semantics and control the try-on results. Moreover, we propose a new landmark dataset with a unified labelling rule of landmarks for diverse styles of garments. Extensive experimental results on *Co-first authors contributed equally,†Corresponding author.popular datasets demonstrate that SAL-VTON can handle misalignment and outperform state-of-the-art methods both qualitatively and quantitatively. The dataset is available on https://modelscope.cn/datasets/damo/SAL-HG/summary.

1. Introduction In recent years, with the rapid popularization of online shopping, virtual try-on [6, 9, 16, 32, 52] has attracted extensive attention for its potential applications. Image- based virtual try-on [4, 34, 49] aims to synthesize a photo- realistic try-on image by transferring a garment image onto the corresponding region of a person. Commonly, there are significant spatial geometric gaps between the in-shop garment image and the person image, leading to garments failing to align the corresponding body parts of person. To address the above issue, prior arts take geometric deformation models to align the garment with the person’s body. Early works [16, 22, 43] widely use the Thin- Plate Spline (TPS) deformation model [39], whereas the smoothness constraint of TPS transformation limits the warping capacity. Recently, the flow operation is applied, with a high dimension of freedom to warp garments [5, 11, 15, 18]. Nonetheless, the flow operation falls short on gar- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17194 ment regions with large deformation. The aforementioned methods focus on modeling the overall deformation of the garment, but ignore the local semantic association between garment and person. Therefore, when there are large local deformations of garments, the try-on results usually occur misalignment, such as missing or mixing garments (see left part of Fig. 1). To address the local misalignment problem, Xieet al. [46] introduce the patch-routed disentanglement module to splice different parts of the garment. However, this method may result in significant blank spaces between spliced parts of the garment. Fortunately, the landmarks in the garment image and the person image naturally have local semantic associations. As can be observed from Fig. 2, the pixels around landmark A on the try-on result should come from the landmark A′area on the garment. Such a pair of landmarks with the same local semantics are referred to as semantically associated landmarks. Based on this observation, this paper presents a novel virtual try-on algorithm named SAL-VTON, which links the garment with the person via semantically as- sociated landmarks to help align the garment with the person. Notably, the proposed approach varies differently from the previous landmark-guided try-on methods [28,37]. LM-VTON [28] and LG-VTON [37] utilize landmarks to supervise the TPS transformation. However, the potential of local semantic association has not been fully explored, and the limited degrees of freedom of TPS transformation further hinder performance improvements. SAL-VTON, for the first time, introduces the local flow estimated via semantically associated landmarks to effectively model the local semantic association. In addition, a generator with Landmark-Aware Semantic Normalization Layer (LASNL) is carried out to better capture local semantics. Specifically, the proposed SAL-VTON consists of three stages. Firstly, the semantically associated landmarks are estimated using the landmark localization model. Sub- sequently, the semantically associated landmarks are em- ployed as a new representation for virtual try-on, and fed into the warping model. Based on the semantically associated landmarks and learnable deformable patches, the warping model explicitly associates the corresponding parts of the garment and person to obtain the local flow, which contributes significantly to refine the poor alignment in the global flow. Finally, conditioned on the landmarks, the LASNL generator can achieve improved alignment in virtual try-on images. The estimated landmarks on the try- on result assist the generator in determining if a specific region needs to generate corresponding garment parts. In this way, SAL-VTON can effectively model the local semantic association between the garment and the person, making up for the misalignment in the overall deformation of the garment. Moreover, the try-on results of SAL-VTON can be precisely controlled by manually manipulating the Person Garment Try-on result Figure 2. An example for the semantically associated landmarks on the in-shop garment image and the try-on image. landmarks (see right part of Fig. 1). To this end, we re-annotate images on the popular virtual try-on benchmarks including VITON [16] and VITON-HD [4] datasets. Existing popular clothing landmark datasets [12, 54] adopt different landmark definitions for different categories of garments. In contrast to other datasets, we adopt a unified labelling rule of landmarks for diverse styles of garments, including both standard and non-standard va- rieties. In the proposed dataset1, every image is annotated with 32 landmarks, each of which possesses three kinds of attributes: visible, occluded and absent. The landmarks with the same serial number have the same semantics, which enhances the universality of the dataset. This work makes the following main contributions : (1) A novel virtual try-on algorithm, SAL-VTON, is proposed, which links the garment with the person via semantically associated landmarks. SAL-VTON, for the first time, intro- duces the local flow that can alleviate the misalignment and the LASNL generator for virtual try-on. (2) A new land- mark dataset is proposed, providing a new representation for virtual try-on, with a unified labelling rule of landmarks for diverse styles of garments. (3) Extensive experiments over two popular datasets demonstrate that SAL-VTON is capable of handling misalignment and significantly out- performs other state-of-the-art methods. Furthermore, the extended experiments show that the virtual try-on results can be edited via the landmarks.

Zhang_Weakly_Supervised_Video_Emotion_Detection_and_Prediction_via_Cross-Modal_Temporal_CVPR_2023

Abstract Automatically predicting the emotions of user-generated videos (UGVs) receives increasing interest recently. How- ever, existing methods mainly focus on a few key visual frames, which may limit their capacity to encode the con- text that depicts the intended emotions. To tackle that, in this paper, we propose a cross-modal temporal eras- ing network that locates not only keyframes but also con- text and audio-related information in a weakly-supervised manner. In specific, we first leverage the intra- and inter-modal relationship among different segments to ac- curately select keyframes. Then, we iteratively erase keyframes to encourage the model to concentrate on the contexts that include complementary information. Exten- sive experiments on three challenging video emotion bench- marks demonstrate that our method performs favorably against state-of-the-art approaches. The code is released on https://github.com/nku-zhichengzhang/WECL.

1. Introduction Emotion analysis in user-generated videos (UGVs) has attracted much attention since a growing number of people tend to express their views on social networks [20, 32, 36]. Automatic predictions of video emotions [52, 54] can po- tentially be applied in various areas like online content fil- tering [1], attitude recognition [32], and customer behavior analysis [39]. Emotions evoked in UGVs usually depend on multiple perspectives, such as actions, events, and objects, where different frames in a UGV may contribute unequally to conveying emotions. Existing methods in this field mainly focus on extracting keyframes from visual content, assuming that these frames hold the dominant information for the intended emotions in videos. For example, Tu et al. [12] introduce an attribu- tion network to locate keyframes with temporal annotations, which are more precise than video-level labeling and lead † Corresponding author. Figure 1. Illustration of the keyframes with larger boxes that are detected by off-the-shelf method [65] on the Ekman-6 dataset [54]. Note that the deeper color represents the higher impact on overall video emotion, and “GT” represents the category label from the ground truth. The texts with orange color are the categorical pre- dicting results. to better performance for video emotion recognition. Yet, annotating the emotional labels frame-by-frame is labor- sensitive and time-consuming [63]. Zhao et al . [65] fur- ther present the visual-audio network (V AANet) conduct- ing three types of attention to automatically discover the discriminative keyframes, which makes it state-of-the-art. However, the selected “keyframes” may fail to represent the intended emotions exactly due to the inherent charac- teristics of human emotions, i.e., subjectivity and ambigu- ity [49, 57, 66]. As illustrated in Figure 1 (a), a woman receives a gift and moves to cry. The video-level emo- tion category is labeled as ‘surprise’. We could observe that V AANet [65] gives the most attention to the keyframes (i.e., “crying” frames in the larger boxes) while ignoring the context and leading to the wrong prediction. Further- more, in Figure 1 (b), a man sees his beloved girl talking with another man happily, which makes him feel sad. How- ever, V AANet only focuses on frames about chatting and categorizes the emotion of this video as ‘joy’. Therefore, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18888 keyframes may lead to limited prediction results. Although the detected keyframes directly convey emotions in most videos, some other information that contains the necessary context should not be ignored. This is because the contex- tual frames could not only provide complementary infor- mation for understanding emotions in UGVs (especially for the cases where keyframes are hard to be distinguished), but also make the model more robust since more cues are considered to recognize emotions rather than the dominant information. To address these problems, we propose a novel cross- modal temporal erasing network, which enables the model to be aware of context for recognizing emotions. Our pro- posed method mainly contains two crucial modules: tempo- ral correlation learning module, and temporal erasing mod- ule. First, we extract the visual feature together with the audio one from each equal-length segment derived from the video. Second, the temporal correlation learning module is introduced to fully discover comprehensive implicit cor- respondences among different segments across audio and visual modal. Then, keyframes are selected by consider- ing the correspondences with other frames in a weakly- supervised manner, where only the video-level class label is used. Finally, the temporal erasing module iteratively erases the most dominant visual and audio information to train with difficult samples online, that encouraging the model to detect more complementary information from context in- stead of the most dominant ones. Our contributions can be summarized as follows: 1) We introduce a weakly-supervised network to exploit keyframes and the necessary context in a unified CNN framework, which encourages model to extract features from multiple discriminative parts and learn better represen- tation for video emotion analysis. 2) We exploit intra- and inter-modal relationships to provide frame-level localized information only with video-level annotation, with which the model consolidates both holistic and local representa- tions for affective computing. We demonstrate the advan- tages of the above contributions with extensive experiments. Our method achieves state-of-the-art on three video emo- tion datasets.

Zhang_CloSET_Modeling_Clothed_Humans_on_Continuous_Surface_With_Explicit_Template_CVPR_2023

Abstract Creating animatable avatars from static scans re- quires the modeling of clothing deformations in different poses. Existing learning-based methods typically add pose- dependent deformations upon a minimally-clothed mesh template or a learned implicit template, which have limita- tions in capturing details or hinder end-to-end learning. In this paper, we revisit point-based solutions and propose to decompose explicit garment-related templates and then add pose-dependent wrinkles to them. In this way, the clothing deformations are disentangled such that the pose-dependent wrinkles can be better learned and applied to unseen poses. Additionally, to tackle the seam artifact issues in recent state-of-the-art point-based methods, we propose to learn point features on a body surface, which establishes a contin- uous and compact feature space to capture the fine-grained and pose-dependent clothing geometry. To facilitate the re- search in this field, we also introduce a high-quality scan dataset of humans in real-world clothing. Our approach is validated on two existing datasets and our newly introduced dataset, showing better clothing deformation results in un- seen poses. The project page with code and dataset can be found at https://www.liuyebin.com/closet.

1. Introduction Animating 3D clothed humans requires the modeling of pose-dependent deformations in various poses. The diver- sity of clothing styles and body poses makes this task ex- tremely challenging. Traditional methods are based on ei- ther simple rigging and skinning [4,20,33] or physics-based simulation [14, 23, 24, 49], which heavily rely on artist ef- forts or computational resources. Recent learning-based methods [10, 36, 39, 59] resort to modeling the clothing de- formation directly from raw scans of clothed humans. De- spite the promising progress, this task is still far from be- ing solved due to the challenges in clothing representations, generalization to unseen poses, and data acquisition, etc. Figure 1. Our method learns to decompose garment templates (top row) and add pose-dependent wrinkles upon them (bottom row). For the modeling of pose-dependent garment geome- try, the representation of clothing plays a vital role in a learning-based scheme. As the relationship between body poses and clothing deformations is complex, an effec- tive representation is desirable for neural networks to cap- ture pose-dependent deformations. In the research of this line, meshes [2, 12, 38], implicit fields [10, 59], and point clouds [36, 39] have been adopted to represent clothing. In accordance with the chosen representation, the clothing de- formation and geometry features are learned on top of a fixed-resolution template mesh [8, 38], a 3D implicit sam- pling space [10,59], or an unfolded UV plane [2,12,36,39]. Among these representations, the mesh is the most efficient one but is limited to a fixed topology due to its discretization scheme. The implicit fields naturally enable continuous fea- ture learning in a resolution-free manner but are too flexible to satisfy the body structure prior, leading to geometry arti- facts in unseen poses. The point clouds enjoy the compact nature and topology flexibility and have shown promising results in the recent state-of-the-art solutions [36,39] to rep- resent clothing, but the feature learning on UV planes still leads to discontinuity artifacts between body parts. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 501 To model the pose-dependent deformation of clothing, body templates such as SMPL [34] are typically leveraged to account for articulated motions. However, a body tem- plate alone is not ideal, since the body template only models the minimally-clothed humans and may hinder the learning of actual pose-dependent deformations, especially in cases of loose clothing. To overcome this issue, recent implicit approaches [59] make attempts to learn skinning weights in the 3D space to complement the imperfect body templates. However, their pose-dependent deformations are typically coarse due to the difficulty in learning implicit fields. For explicit solutions, the recent approach [32] suggests learn- ing coarse templates implicitly at first and then the pose- dependent deformations explicitly. Despite its effective- ness, such a workaround requires a two-step modeling pro- cedure and hinders end-to-end learning. In this work, we propose CloSET, an end-to-end method to tackle the above issues by modeling Clothed humans on a continuous Surface with Explicit Template decomposition. We follow the spirit of recent state-of-the-art point-based approaches [32, 36, 39] as they show the efficiency and po- tential in modeling real-world garments. We take steps for- ward in the following aspects for better point-based model- ing of clothed humans. First, we propose to decompose the clothing deformations into explicit garment templates and pose-dependent wrinkles. Specifically, our method learns a garment-related template and adds the pose-dependent dis- placement upon them, as shown in Fig. 1. Such a garment- related template preserves a shared topology for various poses and enables better learning of pose-dependent wrin- kles. Different from the recent solution [32] that needs two-step procedures, our method can decompose the ex- plicit templates in an end-to-end manner with more gar- ment details. Second, we tackle the seam artifact issues that occurred in recent point-based methods [36, 39]. In- stead of using unfolded UV planes, we propose to learn point features on a body surface, which supports a continu- ous and compact feature space. We achieve this by learn- ing hierarchical point-based features on top of the body surface and then using barycentric interpolation to sample features continuously. Compared to feature learning in the UV space [39], on template meshes [8, 38], or in the 3D implicit space [57, 59], our body surface enables the net- work to capture not only fine-grained details but also long- range part correlations for pose-dependent geometry mod- eling. Third, we introduce a new scan dataset of humans in real-world clothing, which contains more than 2,000 high- quality scans of humans in diverse outfits, hoping to facili- tate the research in this field. The main contributions of this work are summarized below: • We propose a point-based clothed human modeling method by decomposing clothing deformations into explicit garment templates and pose-dependent wrin-kles in an end-to-end manner. These learnable tem- plates provide a garment-aware canonical space so that pose-dependent deformations can be better learned and applied to unseen poses. • We propose to learn point-based clothing features on a continuous body surface, which allows a continu- ous feature space for fine-grained detail modeling and helps to capture long-range part correlations for pose- dependent geometry modeling. • We introduce a new high-quality scan dataset of clothed humans in real-world clothing to facilitate the research of clothed human modeling and animation from real-world scans.

Xu_Multi-View_Adversarial_Discriminator_Mine_the_Non-Causal_Factors_for_Object_Detection_CVPR_2023

Abstract Domain shift degrades the performance of object de- tection models in practical applications. To alleviate the influence of domain shift, plenty of previous work try to decouple and learn the domain-invariant (common) fea- tures from source domains via domain adversarial learn- ing (DAL). However, inspired by causal mechanisms, we find that previous methods ignore the implicit insignificant non-causal factors hidden in the common features. This is mainly due to the single-view nature of DAL. In this work, we present an idea to remove non-causal factors from com- mon features by multi-view adversarial training on source domains, because we observe that such insignificant non- causal factors may still be significant in other latent spaces (views) due to the multi-mode structure of data. To sum- marize, we propose a Multi-view Adversarial Discriminator (MAD) based domain generalization model, consisting of a Spurious Correlations Generator (SCG) that increases the diversity of source domain by random augmentation and a Multi-View Domain Classifier (MVDC) that maps features to multiple latent spaces, such that the non-causal factors are removed and the domain-invariant features are purified. Extensive experiments on six benchmarks show our MAD obtains state-of-the-art performance.

1. Introduction The problem of how to adapt object detectors to un- known target domains in real world has drawn increasing attention. Traditional object detection methods [11, 12, 25, 29,30] are based on independent and identically distributed (i.i.d.) hypothesis, which assume that the training and test- ing datasets have the same distribution. However, the target distribution can hardly be estimated in real world and dif- fers from the source domains, which is coined as domain *Corresponding author (Lei Zhang) Domain classifier FeatureExtractorDomain classifierResnet blockSunnyFoggy Figure 1. Illustration of the biased learning of conventional DAL. The domain classifier easily encounters early stop and fails. shift [38]. And the performance of object detection models will sharply drop when facing the domain shift problem. Domain adaptation (DA) [3, 6, 17, 34, 40, 44, 52] is pro- posed to deal with the domain shift problem, which enables the model to be adapted to the target distribution by aligning features extracted from the source and unlabeled target do- mains. However, the requirement of target domain datasets still limits the applicability of DA methods in reality. Do- main generalization (DG) [49] goes one step further, aiming to train a model from single or multiple source domains that can generalize to unknown target domains. Although lots of DG methods have been proposed in the image classification field, there are still some unresolved problems. In our opinion, the common features extracted by previous DG methods are still not pure enough. The main reason is that through a single-view domain discriminator in DAL, only the significant domain style information can be removed, while some implicit and insignificant non-causal factors in source domains may be absorbed by the feature extractor as a part of common features. This has never been noticed. This implies the multi-mode structure of data and single-view domain discriminator cannot fully interpret the data. There is a piece of evidence to support our claim. To confirm our suspicions on the domain discriminator, we designed a validation experiment. As is shown in Fig. 1, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8103 Causal Domain common But Non-causalDomain specific and Non -causal Sunny FoggyLight color Background Resolution IlluminationCamera angle Size Significant domain differences Potential domain differencesFigure 2. Relationships among causal factors, noncausal factors, domain specific feature and domain common feature. we use DANN model [9] with DAL strategy to train a com- mon feature extractor. When domain classifier converges, we freeze feature extractor and re-train domain classifier with a newly added residual block [14]. We observe an interesting phenomenon: when re-trained with the newly added residual block, the domain classifier loss continues to decline . That is, some domain-specific information still exists. This phenomenon confirms our claim that in existing DG, DAL cannot explore and remove all domain specific features. This is because domain classifier only observes significant domain-specific feature in a single-view, while insignificant domain specific features in one view (space) can be significant in other views (latent spaces). Based on the former experiment, we propose that min- ing common features through DAL in single-view on a lim- ited number of domains is insufficient. By using traditional DAL, only the primary style information w.r.t. domain la- bels can be removed. Here we analyse this problem from the perspective of causality. As shown in Fig. 2, in a lim- ited number of domains, the common features still contain non-causal factors such as light color, illumination, back- ground, etc., which is expressed as the orange arrows in the figure. And such insignificant non-causal factors observed from one view may still be significant uninformative fea- tures in other latent spaces (views). So a natural idea is to explore and remove the implicit non-causal information from multiple views and purify the common features for generalizing to unseen domains. In order to remove the potential non-causal information, we rethink the domain discriminator in DAL and propose a multi-view adversarial domain discriminator (MAD) that can observe the implicit insignificant non-causal factors. In our life, in order to get the whole architecture of an object, we often need to observe it from multiple views/profiles. A toy example is shown in Fig. 3 (left part). When we observe the Penrose triangle from one specific view, we might mis- classify it as a triangle , ignoring that it might also appear to be Lfrom another perspective. Following this intuition, we construct a Multi-View Domain Classifier (MVDC) that can discriminate features in multiple views. Specifically, photo art cartoon sketchview 1 view 2 view 3 Figure 3. An illustration of the multi-view idea and effect of MAD. Left: a toy example. Right: attention heatmaps of different views. we simulate multi-view observations by mapping the fea- tures to different latent spaces with auto-encoders [16], and discriminate these transformed features via multi-view do- main classifiers. By mining and removing as many non- causal factors as possible, MVDC encourages the feature extractor to learn more domain-invariant but causal factors. We conduct an experiment based on MVDC and show the heatmaps from different views in Fig. 3 (right part), which verifies our idea that different noncausal factors can be un- veiled in different views. Although the Multi-View Domain Classifier can remove the implicit non-causal features in principle, it still implies a sufficient diversity of source domains during training. So we further design a Spurious Correlation Generator (SCG) to increase the diversity of source domains. Our SCG gen- erates non-causal spurious connections by randomly trans- forming the low-frequency and extremely high-frequency components, as [19] points out that in the spectrum of im- ages, the extremely high and low frequency parts contain the majority of domain-specific components. Combining MVDC and SCG, the Multi-view Adversar- ial Discriminator (MAD) is formalized. Cross-domain ex- periments on six standard datasets show our MAD achieves the SOTA performance compared to other mainstream DGOD methods. The contributions are three-fold: 1. We point out that existing DGOD work focuses on ex- tracting common features but fails to mine and remove the potential spurious correlations from a causal perspective. 2. We propose a Multi-view Adversarial Discriminator (MAD) to eliminate implicit non-causal factors by discrim- inating non-causal factors from multiple views and extract- ing domain-invariant but causal features. 3. We test and analyze our method on standard datasets, verifying the effectiveness and superiority of our method.

Zhang_Nerflets_Local_Radiance_Fields_for_Efficient_Structure-Aware_3D_Scene_Representation_CVPR_2023

Abstract We address efficient and structure-aware 3D scene rep- resentation from images. Nerflets are our key contribution– a set of local neural radiance fields that together represent a scene. Each nerflet maintains its own spatial position, ori- entation, and extent, within which it contributes to panop- tic, density, and radiance reconstructions. By leveraging only photometric and inferred panoptic image supervision, we can directly and jointly optimize the parameters of a set of nerflets so as to form a decomposed representation of the scene, where each object instance is represented by a group of nerflets. During experiments with indoor and outdoor en- vironments, we find that nerflets: (1) fit and approximate the scene more efficiently than traditional global NeRFs, (2) al- low the extraction of panoptic and photometric renderings from arbitrary views, and (3) enable tasks rare for NeRFs, such as 3D panoptic segmentation and interactive editing. Our project page.

1. Introduction This paper aims to produce a compact, efficient, and comprehensive 3D scene representation from only 2D im- ages. Ideally, the representation should reconstruct appear-ances, infer semantics, and separate object instances, so that it can be used in a variety of computer vision and robotics tasks, including 2D and 3D panoptic segmentation, interac- tive scene editing, and novel view synthesis. Many previous approaches have attempted to generate rich 3D scene representations from images. PanopticFu- sion [34] produces 3D panoptic labels from images, though it requires input depth measurements from specialized sen- sors. NeRF [32] and its descendants [3, 4,33,39] produce 3D density and radiance fields that are useful for novel view synthesis, surface reconstruction, semantic segmenta- tion [50, 60], and panoptic segmentation [5, 21]. However, existing approaches require 3D ground truth supervision, are inefficient, or do not handle object instances. We propose nerflets, a 3D scene representation with mul- tiple local neural fields that are optimized jointly to describe the appearance, density, semantics, and object instances in a scene (Figure 1). Nerflets constitute a structured andirreg- ular representation– each is parameterized by a 3D center, a 3D XYZ rotation, and 3 (per-axis) radii in a 9-DOF coordi- nate frame. The influence of every nerflet is modulated by a radial basis function (RBF) which falls off with increasing distance from the nerflet center according to its orientation and radii, ensuring that each nerflet contributes to a local This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8274 part of the scene. Within that region of influence, each ner- flet has a miniature MLP to estimate density and radiance. It also stores one semantic logit vector describing the category (e.g., “car”) of the nerflet, and one instance label indicating which real-world object it belongs to (e.g., “the third car”). In Figure 1, each ellipsoid is a single nerflet, and they are colored according to their semantics. A scene can contain any number of nerflets, they may be placed anywhere in space, and they may overlap, which provides the flexibility to model complex, sparse 3D scenes efficiently. Since multiple nerflets can have the same in- stance label, they can combine to represent the density and radiance distributions of complex object instances. Con- versely, since each nerflet has only one instance label, the nerflets provide a complete decomposition of the scene into real-world objects. Nerflets therefore provide a 3D panoptic decomposition of a scene that can be rendered and edited. Synthesizing images using nerflets proceeds with density-based volume rendering just as in NeRF [32]. How- ever, instead of evaluating one large MLP at each point sam- ple along a ray, we evaluate the small MLPs of only the ner- flets near a sample. We average the results, weighting by the influence each nerflet has over that sample. The ren- dering is fully-differentiable with respect to all continuous parameters of the nerflets. Fitting the nerflet representation is performed from a set of posed RGB images with a single training stage. After training, instance labels are assigned based on the scene structure, completing the representation. Experiments with indoor and outdoor datasets confirm the main benefits of nerflets. We find that: 1) Parsimony en- courages the optimizer to decompose the scene into nerflets with consistent projections into novel panoptic images (Sec- tion 4.1); 2) Semantic supervision can be beneficial to novel view synthesis (Section 4.2); 3) Structure encourages effi- ciency, compactness, and scalability (Section 3.4); and 4) the explicit decomposition of a scene improves human inter- pretability for easy interactive editing, including adding and removing objects (Section 4.1). These benefits enable state- of-the-art performance on the KITTI360 [24] novel seman- tic view synthesis benchmark, competitive performance on ScanNet 3D panoptic segmentation tasks with more limited supervision, and an interactive 3D editing tool that lever- ages the efficiency and 3D decomposition of nerflets. The following summarizes our main contributions: • We propose a novel 3D scene representation made of small, posed, local neural fields named nerflets. • The pose, shape, panoptic, and appearance information of nerflets are all fit jointly in a single training stage, re- sulting in a comprehensive learned 3D decomposition from real RGB images of indoor or outdoor scenes. • We test nerflets on 4 tasks- novel view synthesis, panoptic view synthesis, 3D panoptic segmentationand reconstruction, and interactive editing. • We achieve 1st place on the KITTI-360 semantic novel view synthesis leaderboard.

Yang_Complementary_Intrinsics_From_Neural_Radiance_Fields_and_CNNs_for_Outdoor_CVPR_2023

Abstract Relighting an outdoor scene is challenging due to the diverse illuminations and salient cast shadows. Intrinsic image decomposition on outdoor photo collections could partly solve this problem by weakly supervised labels with albedo and normal consistency from multi-view stereo. With neural radiance fields (NeRF), editing the appearance code could produce more realistic results without interpreting the outdoor scene image formation explicitly. This paper proposes to complement the intrinsic estimation from vol- ume rendering using NeRF and from inversing the photo- metric image formation model using convolutional neural networks (CNNs). The former produces richer and more reliable pseudo labels (cast shadows and sky appearances in addition to albedo and normal) for training the latter to predict interpretable and editable lighting parameters via a single-image prediction pipeline. We demonstrate the ad- vantages of our method for both intrinsic image decompo- sition and relighting for various real outdoor scenes.

1. Introduction The same landmark may appear with drastically vary- ing appearances in different photos, even if they are taken from the same viewpoint with the same camera parameters, e.g., the Taj Mahal may look golden or white at sunset or in the afternoon1. For a set of photos containing the same landmark captured in different seasons and times, their “dy- namic” lighting changes (compared with the relatively “sta- ble” geometry and reflectance) play a vital role in explain- ing such great appearance variations. If we can indepen- dently manipulate lighting in these photos, the relighted outdoor scenes could substantially improve experiences for taking digital photographs. Outdoor scene relighting could be realized by learning a style transfer procedure [1, 5]. Such a process only requires #Contributed equally to this work as first authors ∗Corresponding author 1Changing colours of Taj Mahal: agratajcitytour.coma single reference image for editing a target image, but it ap- parently retouches the image to “look like” each other, with- out explicitly modeling the lighting changes. Intrinsic im- age decomposition, which inversely decomposes the photo- metric image formation model, has been extended to work with outdoor photo collections [32–34]. The common ge- ometry/reflectance (for the whole collection) and distinctive lighting components (for each image) are estimated using deep convolutional neural networks (CNNs), so that relight- ing could be achieved by keeping the former while editing the latter in a physics-aware manner. These methods ex- plicitly conduct computationally expensive multi-view re- construction of the scene at the training stage. The weakly supervised constraints built upon albedo and normal con- sistency via multi-view correspondence cannot support the handling of cast shadows [33] or still struggle with strong cast shadows [32]. Recently, the emerging of neural radiance fields (NeRF) [23] has not only boosted the performance of novel view synthesis with significantly better quality for outdoor scene photo collections [22], but has also been demon- strated to be capable of transferring the lighting appearance across the image set using hallucination [4] or a parametric lighting model [28]. However, these existing NeRF meth- ods in the outdoor scene either miss the explanation to some important intrinsics for relighting such as cast shadows (ex- cept for [28]) or ignore distinctive characteristics between the non-sky and sky regions, in a physically interpretable manner. In this paper, we hope to conduct outdoor scene re- lighting by mutually complementing intrinsics estimated from NeRF and CNN and taking advantages of the com- prehensive representation power of NeRF and physics in- terpretability of CNN-based single-image decomposition, in asingle-image inference pipeline as shown in Figure 1. We formulate the color formation of a pixel as a combina- tion of objects and the sky by tracing rays from camera ori- gins to the furthest plane and accumulating the voxel intrin- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16600 𝛾!(𝐱)𝛾"(𝐝) Input image 𝐈 Ray sampling𝐱# Cameraparameters𝐫=𝐨+𝑡𝐝𝐱#$% IntrinsicCNN LightingCNN Reconstruction ,𝐂Reconstruction .𝐈SH lighting𝐋 Color ,𝐂&'Albedo 𝐀Shadow 𝐒Normal 𝐍Albedo 3𝐀Shadow 4𝐒Normal 3𝐍Sky mask 𝐌()*Rendered ,𝐂()*Relighting renderer IntrinsicsSky renderingPseudo-supervision 𝐹+,-,. 𝐹/-012, 𝐹(3/2,' 𝐹21&(45* 𝐹()* Figure 1. Illustration of our overall pipeline. For each image, the corresponding camera rays and sampled 3D points are positionally- encoded for MLPs (gray block). Our IntrinsicCNN and LightingCNN modules derive lighting-(in)dependent intrinsic components and second-order spherical harmonics (SH) lighting coefficients (green block), while Our NeRF module provides the pseudo labels of intrinsics and sky mask by volumetric rendering through MLPs (blue block). Our SkyMLP module renders the sky with viewing direction and extracted lighting (orange block). Given lighting extracted from the input/reference image, the reconstructed/relighted image is rendered by photometric image formation along with the rendered sky (yellow block). sics. We then propose a modified NeRF system to estimate diffuse albedo, surface normal, cast shadow, and illumina- tion parameters. Our NeRF rendering naturally shares the albedo and normal of one point across all images and inter- prets the geometry from voxel density, which provides more accurate pseudo labels for identifying shadows than purely CNN-based approaches [32–34]. We finally predict the in- trinsics and lighting parameters by designing two separate CNN modules based on the NeRF-produced pseudo labels with a clearer separation of lighting-dependent and indepen- dent intrinsic components to achieve high-quality relighting via single-image prediction. Hence, our contribution becomes clear in three folds: Based on 1) the newly proposed “object-sky” hybrid image formation, 2) the intrinsics estimated from NeRF provide more accurate pseudo labels to complement 3) the intrin- sics estimated from CNNs, for conducting outdoor scene relighting using a single image in a physically interpretable manner and with a visually pleasing appearance, which is demonstrated by our experimental results.

Yu_Zero-Shot_Referring_Image_Segmentation_With_Global-Local_Context_Features_CVPR_2023

Abstract Referring image segmentation (RIS) aims to find a seg- mentation mask given a referring expression grounded to a region of the input image. Collecting labelled datasets for this task, however, is notoriously costly and labor-intensive. To overcome this issue, we propose a simple yet effective zero-shot referring image segmentation method by leverag- ing the pre-trained cross-modal knowledge from CLIP . In order to obtain segmentation masks grounded to the input text, we propose a mask-guided visual encoder that cap- tures global and local contextual information of an input im- age. By utilizing instance masks obtained from off-the-shelf mask proposal techniques, our method is able to segment fine-detailed instance-level groundings. We also introduce a global-local text encoder where the global feature captures complex sentence-level semantics of the entire input expres- sion while the local feature focuses on the target noun phrase extracted by a dependency parser. In our experiments, the proposed method outperforms several zero-shot baselines of the task and even the weakly supervised referring expression segmentation method with substantial margins. Our code is available at https://github.com/Seonghoon-Yu/Zero-shot-RIS.

1. Introduction Recent advances of deep learning has revolutionised com- puter vision and natural language processing, and addressed various tasks in the field of vision-and-language [4, 19, 27, 28, 36, 43, 50]. A key element in the recent success of the multi-modal models such as CLIP [43] is the contrastive image-text pre-training on a large set of image and text pairs. It has shown a remarkable zero-shot transferability on a wide range of tasks, such as object detection [9, 10, 13], semantic segmentation [7, 12, 59, 63], image captioning [40], visual question answering (VQA) [47] and so on. Despite its good transferability of pre-trained multi-modal models, it is not straightforward to handle dense prediction tasks such as object detection and image segmentation. A pixel-level dense prediction task is challenging since there Figure 1. Illustrations of the task of referring image segmenta- tion and motivations of global-local context features. To find the grounded mask given an expression, we need to understand the relations between the objects as well as their semantics. is a substantial gap between the image-level contrastive pre- training task and the pixel-level downstream task such as se- mantic segmentation. There have been several attempts to re- duce gap between two tasks [44,54,63], but these works aim to fine-tune the model consequently requiring task-specific dense annotations, which is notoriously labor-intensive and costly. Referring image segmentation is a task to find the specific region in an image given a natural language text describing the region, and it is well-known as one of challenging vision- and-language tasks. Collecting annotations for this task is even more challenging as the task requires to collect precise referring expression of the target region as well as its dense mask annotation. Recently, a weakly-supervised referring image segmentation method [48] is proposed to overcome this issue. However, it still requires high-level text expres- sion annotations pairing with images for the target datasets and the performance of the method is far from that of the supervised methods. To tackle this issue, in this paper, we fo- cus on zero-shot transferring from the pre-trained knowledge This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19456 of CLIP to the task of referring image segmentation. Moreover, this task is challenging because it requires high-level understanding of language and comprehensive understanding of an image, as well as a dense instance-level prediction. There have been several works for zero-shot semantic segmentation [7, 12, 59, 63], but they cannot be directly extended to the zero-shot referring image segmenta- tion task because it has different characteristics. Specifically, the semantic segmentation task does not need to distinguish instances, but the referring image segmentation task should be able to predict an instance-level segmentation mask. In addition, among multiple instances of the same class, only one instance described by the expression must be selected. For example, in Figure 1, there are two cats in the input image. If the input text is given by “a cat is lying on the seat of the scooter” , the cat with the green mask is the proper output. To find this correct mask, we need to understand the relation between the objects ( i.e.“lying on the seat” ) as well as their semantics ( i.e.“cat”, “scooter” ). In this paper, we propose a new baseline of zero-shot refer- ring image segmentation task using a pre-trained model from CLIP, where global and local contexts of an image and an ex- pression are handled in a consistent way. In order to localize an object mask region in an image given a textual referring expression, we propose a mask-guided visual encoder that captures global and local context information of an image given a mask. We also present a global-local textual encoder where the local-context is captured by a target noun phrase and the global context is captured by a whole sentence of the expressions. By combining features in two different context levels, our method is able to understand a comprehensive knowledge as well as a specific trait of the target object. Note that, although our method does not require any additional training on CLIP model, it outperforms all baselines and the weakly supervised referring image segmentation method with a big margin. Our main contributions can be summarised as follows: •We propose a new task of zero-shot referring image segmentation based on CLIP without any additional training. To the best of our knowledge, this is the first work to study the zero-shot referring image segmenta- tion task. •We present a visual encoder and a textual encoder that integrates global and local contexts of images and sen- tences, respectively. Although the modalities of two encoders are different, our visual and textual features are dealt in a consistent way. •The proposed global-local context features take full advantage of CLIP to capture the target object seman- tics as well as the relations between the objects in both visual and textual modalities.•Our method consistently shows outstanding results com- pared to several baseline methods, and also outperforms the weakly supervised referring image segmentation method with substantial margins.

Yang_QPGesture_Quantization-Based_and_Phase-Guided_Motion_Matching_for_Natural_Speech-Driven_Gesture_CVPR_2023

Abstract Speech-driven gesture generation is highly challenging due to the random jitters of human motion. In addition, there is an inherent asynchronous relationship between hu- man speech and gestures. To tackle these challenges, we in- troduce a novel quantization-based and phase-guided mo- tion matching framework. Specifically, we first present a gesture VQ-VAE module to learn a codebook to summa- rize meaningful gesture units. With each code representing a unique gesture, random jittering problems are alleviated effectively. We then use Levenshtein distance to align di- verse gestures with different speech. Levenshtein distance based on audio quantization as a similarity metric of cor- responding speech of gestures helps match more appropri- ate gestures with speech, and solves the alignment prob- lem of speech and gestures well. Moreover, we introduce phase to guide the optimal gesture matching based on the semantics of context or rhythm of audio. Phase guides when text-based or speech-based gestures should be performed to make the generated gestures more natural. Extensive experiments show that our method outperforms recent ap- proaches on speech-driven gesture generation. Our code, database, pre-trained models and demos are available at https://github.com/YoungSeng/QPGesture .

1. Introduction Nonverbal behavior plays a key role in conveying mes- sages in human communication [26], including facial ex- pressions, hand gestures and body gestures. Co-speech ges- ture helps better self-expression [45]. However, producing human-like and speech-appropriate gestures is still very dif- ficult due to two main challenges: 1) Random jittering : People make many small jitters and movements when they speak, which can lead to a decrease in the quality of the gen- Most people just laughed face my at two men came up Suddenly to me Figure 1. Gesture examples generated by our proposed method on various types of speech. The character is from Mixamo [2]. erated gestures. 2) Inherent asynchronicity with speech : Unlike speech with face or lips, there is an inherent asyn- chronous relationship between human speech and gestures. Most existing gesture generation studies intend to solve the two challenges in a single ingeniously designed neural network that directly maps speech to 3D joint sequence in high-dimensional continuous space [18, 24, 27, 31] using a sliding window with a fixed step size [17, 46, 47]. How- ever, such methods are limited by the representation power of proposed neural networks, like the GENEA gesture- generation challenge results. No system in GENEA chal- lenge 2020 [26] rated above a bottom line that paired the in- put speech audio with mismatched excerpts of training data motion. In GENEA challenge 2022 [48], a motion match- ing based method [50] ranked first in the human-likeness evaluation and upper-body appropriateness evaluation, and outperformed all other neural network-based models. These results indicate that motion matching based models, if de- signed properly, are more effective than neural network based models. Inspired by this observation, in this work, we propose a novel quantization-based motion matching framework for audio-driven gesture generation. Our framework includes two main components aiming at solving the two above chal- lenges, respectively. First, to address the random jittering This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2321 AudioAudio Cand. TextćThe book is openĊĈ 34 96ĊQuan. Seed poseQuan. 72 511Ċ 784 37ĊMotion Matching Speech-gesture databaseText Cand.Phase GuidanceSeed Phase Text matching gesture Audio matching gesture Figure 2. Gesture generation pipeline of our proposed framework. ‘Quan.’ is short for ‘quantization’ and ‘Cand.’ is short for ‘candidate’. Given a piece of audio, text and seed pose, the audio and gesture are quantized. The candidate for the speech is calculated based on the Levenshtein distance, and the candidate for the text is calculated based on the cosine similarity. The optimal gesture is selected based on phase-guidance corresponding to the seed code and the phase corresponding to the two candidates. challenge, we compress human gestures into a space that is lower dimensional and discrete, to reduce input redundancy. Instead of manually indicating the gesture units [23], we use a vector quantized variational autoencoder (VQ-V AE) [42] to encode and quantize joint sequences to a codebook in an unsupervised manner, using a quantization bottleneck. Each learned code is shown to represent a unique gesture pose. By reconstructing the discrete gestures, some ran- dom jittering problems such as grabbing hands and push- ing glasses will be solved. Second, to address the inherent asynchronicity of speech and gestures, Levenshtein distance [28] is used based on audio quantization. Levenshtein dis- tance helps match more appropriate gestures with speech, and solves the alignment problem of speech and gestures well. Moreover, unlike the recent gesture matching mod- els [17, 50], we also consider the semantic information of the context. Third, since the body motion is composed of multiple periodic motions spatially, meanwhile the phase values are able to describe the nonlinear periodicity of the high-dimensional motion curves well [39], we use phase to guide how the gestures should be matched to speech and text. The inference procedure of our framework is shown in Figure 2. Given a piece of audio, text and seed pose, the audio and gesture are first quantized. The best candidate for the speech is calculated based on the Levenshtein dis- tance, and the best candidate for the text is calculated based on the cosine similarity. Then the most optimal gesture is selected based on the phase corresponding to the seed code and the phase corresponding to the two candidates. Our code, database, pre-trained models and demos will be pub- licly available soon. The main contributions of our work are: • We present a novel quantization-based motion match- ing framework for speech-driven gesture generation.• We propose to align diverse gestures with different speech using Levenshtein distance, based on audio quantization. • We design a phase guidance strategy to select optimal audio and text candidates for motion matching. • Extensive experiments show that jittering and asyn- chronicity issues can be effectively alleviated by our framework.

Ye_DeepSolo_Let_Transformer_Decoder_With_Explicit_Points_Solo_for_Text_CVPR_2023

Abstract End-to-end text spotting aims to integrate scene text de- tection and recognition into a unified framework. Deal- ing with the relationship between the two sub-tasks plays a pivotal role in designing effective spotters. Although Transformer-based methods eliminate the heuristic post- processing, they still suffer from the synergy issue between the sub-tasks and low training efficiency. In this paper, we present DeepSolo , a simple DETR-like baseline that lets a single Decoder with Explicit Points Solo for text detec- tion and recognition simultaneously. Technically, for each text instance, we represent the character sequence as or- dered points and model them with learnable explicit point queries. After passing a single decoder, the point queries have encoded requisite text semantics and locations, thus can be further decoded to the center line, boundary, script, and confidence of text via very simple prediction heads in parallel. Besides, we also introduce a text-matching cri- terion to deliver more accurate supervisory signals, thus enabling more efficient training. Quantitative experiments on public benchmarks demonstrate that DeepSolo outper- forms previous state-of-the-art methods and achieves better training efficiency. In addition, DeepSolo is also compati- ble with line annotations, which require much less annota- tion cost than polygons. The code is available at https: //github.com/ViTAE-Transformer/DeepSolo .

1. Introduction Detecting and recognizing text in natural scenes, a.k.a. text spotting, has drawn increasing attention due to its wide range of applications [5, 34, 56, 59], such as autonomous driving [57] and intelligent navigation [7]. How to deal with the relationship between detection and recognition is a long- *Equal contribution. †Corresponding author. This work was done dur- ing Maoyuan Ye’s internship at JD Explore Academy. Feature Extraction Spotting (a) RoI -based (b) Seg -based TrDec. (c) Ours TrDec. Generic Object QueryLinear, RNN, MLP, DeConv TrDec. Point Query+Point Prior+Linear#1 & #2, MLP# 1 & #2 (d) TTS (e) TESTR (f) OursCNN +FPNIMAGE IMAGECNN +FPN CNN +TrEnc.Detector Seg. Seg. Char ...Connector Recognizer Grouping separately trained Char Query+1D EncodingTrDec.#2 +Linear#2 TrDec.#2 +Linear#2 Char Query+1D EncodingTrDec.#2 +Linear#2 TrDec.#1 Point Query+Box Prior+Linear#1, MLP TrDec.#1 Point Query+Box Prior+Linear#1, MLPIMAGEFigure 1. Comparison of pipelines and query designs. TrEnc. (TrDec.): Transformer encoder (decoder). Char: character. standing problem in designing the text spotting pipeline and has a significant impact on structure design, spotting perfor- mance, training efficiency, and annotation cost, etc. Most pioneering end-to-end spotting methods [16,25,28, 29,31,37,43,44,51] follow a detect-then-recognize pipeline, which first detects text instances and then exploits Region- of-Interest (RoI) based connectors to extract features within the detected area, finally feeds them into the following rec- ognizer (Fig. 1a). Although these methods have achieved great progress, there are two main limitations. 1) An extra connector for feature alignment is indispensable. Moreover, some connectors require polygon annotations, which are not applicable when only weak annotations are available. 2) Additional efforts are desired to address the synergy is- sue [17,62] between the detection and recognition modules. In contrast, the segmentation-based methods [49, 53] try to isolate the two sub-tasks and complete spotting in a paral- lel multi-task framework with a shared backbone (Fig. 1b). Nevertheless, they are sensitive to noise and require group- ing post-processing to gather unstructured components. Recently, Transformer [47] has improved the perfor- mance remarkably for various computer vision tasks [9, 10, 23, 32, 33, 40, 46, 50, 54, 60], including text spotting [17, 21, 39, 61]. Although the spotters [21, 61] based on DETR [3] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19348 can get rid of the connectors and heuristic post-processing, they lack efficient joint representation to deal with scene text detection and recognition, e.g., requiring an extra RNN module in TTS [21] (Fig. 1d) or exploiting individual Trans- former decoder for each sub-task in TESTR [61] (Fig. 1e). The generic object query exploited in TTS fails to consider the unique characteristics of scene text, e.g., location and shape. While TESTR uses point queries with box positional prior that is coarse for point predicting, and the queries are different for detection and recognition, introducing unex- pected heterogeneity. Consequently, these designs have a side effect on the performance and training efficiency [55]. In this paper, we propose a novel query form based on explicit point representations of text lines. Built upon it, we present a succinct DETR-like baseline that lets a sin- gleDecoder with Explicit Points Solo (dubbed DeepSolo ) for detection and recognition simultaneously (Fig. 1c and Fig. 1f). Technically, for each instance, we first represent the character sequence as ordered points, where each point has explicit attributes of position, offsets to the top and bot- tom boundary, and category. Specifically, we devise top- K Bezier center curves to fit scene text instances with arbitrary shape and sample a fixed number of on-curve points cov- ering characters in each text instance. Then, we leverage the sampled points to generate positional queries and guide the learnable content queries with explicit positional prior. Next, we feed the image features from the Transformer en- coder and the point queries into a single Transformer de- coder, where the output queries are expected to have en- coded requisite text semantics and locations. Finally, we adopt several very simple prediction heads (a linear layer or MLP) in parallel to decode the queries into the center line, boundary, script, and confidence of text, thereby solving de- tection and recognition simultaneously. In summary, the main contributions are three-fold: 1) We propose DeepSolo, i.e., a succinct DETR-like baseline with a single Transformer decoder and several simple pre- diction heads, to solve text spotting efficiently. 2)We pro- pose a novel query form based on explicit points sampled from the Bezier center curve representation of text instance lines, which can efficiently encode the position, shape, and semantics of text, thus helping simplify the text spotting pipeline. 3)Experimental results on public datasets demon- strate that DeepSolo is superior to previous representative methods in terms of spotting accuracy, training efficiency, and annotation flexibility.

Yu_MonoHuman_Animatable_Human_Neural_Field_From_Monocular_Video_CVPR_2023

Abstract Animating virtual avatars with free-view control is cru- cial for various applications like virtual reality and digi- tal entertainment. Previous studies have attempted to uti- lize the representation power of the neural radiance field (NeRF) to reconstruct the human body from monocular videos. Recent works propose to graft a deformation net- work into the NeRF to further model the dynamics of the hu- man neural field for animating vivid human motions. How- ever, such pipelines either rely on pose-dependent repre- sentations or fall short of motion coherency due to frame- independent optimization, making it difficult to generalize to unseen pose sequences realistically. In this paper, we propose a novel framework MonoHuman , which robustly renders view-consistent and high-fidelity avatars under ar- bitrary novel poses. Our key insight is to model the de- formation field with bi-directional constraints and explic- itly leverage the off-the-peg keyframe information to reason the feature correlations for coherent results. Specifically, we first propose a Shared Bidirectional Deformation mod- ule, which creates a pose-independent generalizable defor- mation field by disentangling backward and forward defor- mation correspondences into shared skeletal motion weight and separate non-rigid motions. Then, we devise a Forward Correspondence Search module, which queries the corre- spondence feature of keyframes to guide the rendering net-work. The rendered results are thus multi-view consistent with high fidelity, even under challenging novel pose set- tings. Extensive experiments demonstrate the superiority of our proposed MonoHuman over state-of-the-art methods.

1. Introduction Rendering a free-viewpoint photo-realistic view synthe- sis of a digital avatar with explicit pose control is an im- portant task that will bring benefits to AR/VR applica- tions, virtual try-on, movie production, telepresence, etc. However, previous methods [32, 33, 58] usually require carefully-collected multi-view videos with complicated sys- tems and controlled studios, which limits the usage in gen- eral and personalized scenarios applications. Therefore, though challenging, it has a significant application value todirectly recover and animate the digital avatar from a monocular video. Previous rendering methods [33] can synthesize realis- tic novel view images of the human body, but hard to ani- mate the avatar in unseen poses. To address this, some re- cent methods deform the neural radiance fields (NeRF [27]) [32, 49] to learn a backward skinning weight of paramet- ric models depending on the pose or individual frame in- dex. They can animate the recovered human in novel poses with small variations to the training set. However, as the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16943 blending weights are pose-dependent, these methods usu- ally over-fit to the seen poses of training data, therefore, lack of generalizability [49]. Notably, one category of ap- proaches [5, 21] solves this problem by learning a pose- independent forward blend weight in canonical space and using the root-finding algorithm to search the backward cor- respondence. However, the root-finding algorithm is time- consuming. [4] proposes to add a forward mapping network to help the learning of backward mapping, though with con- sistent constraint, their backward warping weights still are frame dependent. Some other works [16,32,42] leverage the blend weight from the template model like SMPL [25]. The accuracy of their deformation heavily hinges on the tem- plate model and always fails in the cloth parts as SMPL does not model it. How to learn an accurate and generaliz- able deformation field is still an open problem. To reconstruct and animate a photo-realistic avatar that can generalize to unseen poses from the monocular video, we conclude three key observations to achieve generaliz- ability from recent studies: 1)The deformation weight field should be defined in canonical space and as pose- independent as possible [5,48,49]; 2)An ideal deformation field should unify forward and backward deformation to al- leviate ambiguous correspondence on novel poses; 3)Direct appearance reference from input observation helps improve the fidelity of rendering. We propose MonoHuman , a novel framework that en- joys the above strengths to reconstruct an animatable digi- tal avatar from only monocular video. Concretely, we show how to learn more correct and general deformation from such limited video data and how to render realistic results from deformed points in canonical space. We first intro- duce a novel Shared Bidirectional Deformation module to graft into the neural radiance fields, which disentangles the backward and forward deformation into one shared skeletal motion and two separate residual non-rigid motions. The shared motion basis encourages the network to learn more general rigid transformation weights. The separate residual non-rigid motions guarantee the expressiveness of the accu- rate recovery of pose-dependent deformation. With the ac- curate learned deformation, we further build an observation bank consisting of information from sparse keyframes, and present a Forward Correspondence Search module to search observation correspondence from the bank which helps cre- ate high fidelity and natural human appearance, especially the invisible part in the current frame. With the above de- signs, our framework can synthesize a human at any view- point and any pose with natural shape and appearance. To summarize, our main contributions are three-fold: • We present a new approach MonoHuman that can syn- thesize the free viewpoint and novel pose sequences of a performer with explicit pose control, only requiring a monocular video as supervision.• We propose the Shared Bidirectional Deformation module to achieve generalizable consistent forward and backward deformation, and the Forward Search Correspondence module to query correspondence ap- pearance features to guide the rendering step. • Extensive experiments demonstrate that our frame- work MonoHuman renders high-fidelity results and outperforms state-of-the-art methods.

Yang_Neural_Vector_Fields_Implicit_Representation_by_Explicit_Learning_CVPR_2023

Abstract Deep neural networks (DNNs) are widely applied for nowadays 3D surface reconstruction tasks and such meth- ods can be further divided into two categories, which re- spectively warp templates explicitly by moving vertices or represent 3D surfaces implicitly as signed or unsigned dis- tance functions. Taking advantage of both advanced ex- plicit learning process and powerful representation abil- ity of implicit functions, we propose a novel 3D repre- sentation method, Neural Vector Fields (NVF). It not only adopts the explicit learning process to manipulate meshes directly, but also leverages the implicit representation of unsigned distance functions (UDFs) to break the barri- ers in resolution and topology. Specifically, our method first predicts the displacements from queries towards the surface and models the shapes as Vector Fields. Rather than relying on network differentiation to obtain direction fields as most existing UDF-based methods, the produced vector fields encode the distance and direction fields both and mitigate the ambiguity at “ridge” points, such that the calculation of direction fields is straightforward and differentiation-free. The differentiation-free characteristic enables us to further learn a shape codebook via Vector Quantization, which encodes the cross-object priors, accel- erates the training procedure, and boosts model generaliza- tion on cross-category reconstruction. The extensive exper- iments on surface reconstruction benchmarks indicate that our method outperforms those state-of-the-art methods in different evaluation scenarios including watertight vs non- watertight shapes, category-specific vs category-agnostic reconstruction, category-unseen reconstruction, and cross- domain reconstruction. Our code is released at https: //github.com/Wi-sc/NVF .

1. Introduction Reconstructing continuous surfaces from unstructured, discrete and sparse point clouds is an emergent but non- trivial task in nowadays robotics, vision and graphics appli- cations, since the point clouds are hard to be deployed into (b) Mesh (c) Voxel/Occupancy (e) Unsigned Distance (f) Vector Fields (a) Point Cloud (d) Signed Distance (g) Mesh Deformation via Vector FieldsFigure 1. Common 3D representations. Explicit representations: (a) point clouds, (b) meshes, (c) voxels. Implicit representa- tions: (c) occupancy, (d) reconstruction from the signed distance functions, and (e) reconstruction from unsigned distance func- tions. Our method represents continuous surfaces through (f) vec- tor fields. (g) Vector fields can deform meshes (red) as explicit representation methods. the downstream applications without recovering to high- resolution surfaces [5, 7, 38, 42]. With the tremendous success of Deep Neural Networks (DNNs), a few DNN-based surface reconstruction meth- ods have already achieved promising reconstruction perfor- mance. These methods can be roughly divided into two cat- egories according to whether their output representations are explicit or implicit. As shown in Fig. 1, explicit rep- resentation methods including mesh andvoxel based ones denote the exact location of a surface, which learn to warp templates [3, 4, 19, 26, 29, 68] or predict voxel grids [10, 30, 59]. Explicit representations are friendly to downstream ap- plications, but they are usually limited by resolution and topology. On the other hand, implicit representations such This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16727 asOccupancy andSigned Distance Functions (SDFs) repre- sent the surface as an isocontour of a scalar function, which receives increasing attention due to their capacity to repre- sent surfaces with more complicated topology and at arbi- trary resolution [12, 14, 22, 35, 44, 48]. However, most im- plicit representation methods usually require specific pre- processing to close non-watertight meshes and remove in- ner structures. To free from the above pre-processing re- quirements for implicit representation, Chibane et al. [15] introduced Neural Unsigned Distance Fields (NDF) , which employs the Unsigned Distance Functions (UDFs) for neu- ral implicit functions (NIFs) and models continuous sur- faces by predicting positive scalar between query locations and the target surface. Despite certain advantages, UDFs require a more complicated surface extraction process than other implicit representation methods ( e.g., SDFs). Such a process using Ball-Pivoting Algorithm [5] or gradient-based Marching Cube [28,83] relies on model differentiation dur- ing inference ( i.e., differentiation-dependent). Moreover, UDFs leave gradient ambiguities at “ridge” points, where the gradients1used for surface extraction cannot accurately point at target points as illustrated by Fig. 2a. In this work, we propose a novel 3D representation method, Neural Vector Fields (NVF) , which leverages the explicit learning process of direct manipulation on meshes and the implicit representation of UDFs to enjoy the ad- vantages of both approaches. That is, NVF can directly manipulate meshes as those explicit representation meth- ods as Fig. 1g, while representing the shapes with arbi- trary resolution and topology as those implicit represen- tation methods. Specifically, NVF models the 3D shapes as vector fields and computes the displacement between a pointq∈R3and its nearest-neighbor point on the surface ˆ q∈R3by using a learned function f(q) =∆q=ˆ q−q: R3⇒R3. Therefore, NVF could serve both as an implicit function and an explicit deformation function, since the dis- placement output of the function could be directly used to deform source meshes ( i.e., Fig. 1g). In general, it encodes both distance and direction fields within vector fields, which can be straightforwardly obtained from the vector fields. Different from existing UDF-based methods, our NVF representation avoids the comprehensive inference process by skipping the gradient calculation during the surface ex- traction procedure1, and mitigates ambiguities by directly learning displacements as illustrated by Fig. 2b. Such one-pass forward-propagation nature frees NVF from dif- ferentiation dependency, significantly reduces the infer- ence time and memory cost, and allows our model to learn a shape codebook consisting of un-differentiable dis- crete shape codes in the embedded feature space. The 1Learning-based methods calculate the gradients of distance fields via model differentiation. The opposite direction of gradients should point to the nearest-neighbor point on the target surface. 𝒅 𝒅 Surface Gradient Query(a) NDF [15]. Surface Vector Query (b) NVF. Figure 2. Gradient ambiguities. (a) NDF [15] cannot guarantee to pull points onto surfaces ( i.e., ambiguity of gradient), while (b) our NVF address this issue by direct displacement learning. learned shape codebook further provides cross-object priors to consequently improve the model generalization on cross- category reconstruction, and accelerates the training proce- dure as a regularization term during training. We use VQ as an example to demonstrate that the differentiation- free property of NVF provides more flexibility in model design in this paper. We conduct extensive experiments on two surface reconstruction benchmark datasets: a synthetic dataset ShapeNet [8] and a real scanned dataset MGN [6]. Besides category-specific reconstruction [15,76] as demonstrated in most reconstruction methods, we also evaluate our frame- work by category-agnostic reconstruction, category-unseen reconstruction, and cross-domain reconstruction tasks to exploit the model generalization. Our experimental results indicate that our NVF can significantly reduce the inference time compared with other UDF-based methods as we avoid the comprehensive surface extraction step and circumvent the requirement of gradient calculation at query locations. Also, using the shape codebook, we observe a significant performance improvement and a better model generaliza- tion across categories. Our contributions are summarized as follows. • We propose a 3D representation NVF for better 3D field representation, which bridges the explicit learn- ing and implicit representations, and benefits from both of their advantages. Our method can obtain the displacement of a query location in a differentiation- free way, and thus it significantly reduces the infer- ence complexity and provides more flexibility in de- signing network structures which may include non- differentiable components. • Thanks to our differentiation-free design, we further propose a learned shape codebook in the feature space, which uses VQ strategy to provide cross-object priors. In this way, each query location is encoded as a com- position of discrete codes in feature space and further used to learn the NVF. • We conduct the extensive experiments to evaluate the effectiveness of our proposed method. It consistently shows promising performance on two benchmarks 16728 across different evaluation scenarios: water-tight vs non-water-tight shapes, category-specific vs category- agnostic reconstruction, category-unseen reconstruc- tion, and cross-domain reconstruction.

Zhang_DA-DETR_Domain_Adaptive_Detection_Transformer_With_Information_Fusion_CVPR_2023

Abstract The recent detection transformer (DETR) simplifies the object detection pipeline by removing hand-crafted designs and hyperparameters as employed in conventional two- stage object detectors. However, how to leverage the sim- ple yet effective DETR architecture in domain adaptive ob- ject detection is largely neglected. Inspired by the unique DETR attention mechanisms, we design DA-DETR, a do- main adaptive object detection transformer that introduces information fusion for effective transfer from a labeled source domain to an unlabeled target domain. DA-DETR introduces a novel CNN-Transformer Blender (CTBlender) that fuses the CNN features and Transformer features inge- niously for effective feature alignment and knowledge trans- fer across domains. Specifically, CTBlender employs the Transformer features to modulate the CNN features across multiple scales where the high-level semantic information and the low-level spatial information are fused for accu- rate object identification and localization. Extensive experi- ments show that DA-DETR achieves superior detection per- formance consistently across multiple widely adopted do- main adaptation benchmarks.

1. Introduction Object detection aims to predict a bounding box and a class label for interested objects in images and it has been a longstanding challenge in the computer vision re- search. Most existing work adopts a two-stage detection pipeline that involves heuristic anchor designs, complicated post-processing such as non-maximum suppression (NMS), etc. The recent detection transformer (DETR) [5] has at- tracted increasing attention which greatly simplifies the two-stage detection pipeline by removing hand-crafted an- chors [21, 22, 49] and NMS [21, 22, 49]. Despite its great detection performance under a fully supervised setup, how to leverage the simple yet effective DETR architecture in domain adaptive object detection is largely neglected. *Equal contribution, {jingyi.zhang, jiaxing.huang }@ntu.edu.sg. †Corresponding author, [email protected]. Figure 1. The vanilla Deformable-DETR [81] trained with la- beled source data cannot handle target data well due to cross- domain shift. The introduction of adversarial feature alignment inDeformable-DETR + Direct-align [19] improves the detection clearly. The proposed DA-DETR fuses CNN features and trans- former features ingeniously which achieves superior unsupervised domain adaptation consistently across four widely adopted bench- marks including Cityscapes →Foggy cityscapes in (a), SIM 10k →Cityscapes in (b), KITTI →Cityscapes in (c) and PASCAL VOC→Clipart1k in (d). Different from the conventional CNN-based detection architectures such as Faster RCNN [49], DETR has a CNN backbone followed by a transformer head consisting of an encoder-decoder structure. The CNN backbone and the transformer head learn different types of features [17,48,69] - the former largely captures low-level localization features (e.g., edges and lines around object boundaries) while the latter largely captures global inter-pixel relationship and high-level semantic features. At the other end, many prior studies show that fusing different types of features often is often helpful in various visual recognition tasks [9, 11]. Hence, it is very meaningful to investigate how to fuse the two types of DETR features to address the domain adaptive object detection challenge effectively. We design DA-DETR, a simple yet effective Domain Adaptive DETR that introduces information fusion into the DETR architecture for effective domain adaptive ob- ject detection. The core design is a CNN-Transformer Blender (CTBlender) that employs the high-level semantic This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23787 features in the Transformer head to conditionally modulate the low-level localization features in the CNN backbone. CTBlender consists of two sequential fusion components, including split-merge fusion (SMF) that fuses CNN and Transformer features within an image and scale aggregation fusion (SAF) that fuses the SMF features across multiple feature scales. Different from the existing weight-and-sum fusion [9, 11], SMF first splits CNN features into multiple groups with different semantic information as captured by the Transformer head and then merges them with channel shuffling for effective information communication among different groups. The SMF features of each scale are then aggregated by SAF for fusing both semantic and localiza- tion information across multiple feature scales. Hence, CT- Blender captures both semantic and localization features in- geniously which enables comprehensive and effective inter- domain feature alignment with a single discriminator. The main contributions of this work can be summarized in three aspects. First , we propose DA-DETR, a simple yet effective domain adaptive detection transformer that intro- duces information fusion for effective domain adaptive ob- ject detection. To the best of our knowledge, this is the first work that explores information fusion for domain adaptive object detection. Second , we design a CNN-Transformer Blender that fuses the CNN features and Transformer fea- tures ingeniously for effective feature alignment and knowl- edge transfer across domains. Third , extensive experi- ments show that DA-DETR achieves superior object detec- tion over multiple widely studied domain adaptation bench- marks as compared with the state-of-the-art as shown in Fig. 1.

Yin_Hi4D_4D_Instance_Segmentation_of_Close_Human_Interaction_CVPR_2023

Abstract We propose Hi4D, a method and dataset for the auto- matic analysis of physically close human-human interaction under prolonged contact. Robustly disentangling several in-contact subjects is a challenging task due to occlusions and complex shapes. Hence, existing multi-view systems typically fuse 3D surfaces of close subjects into a single, connected mesh. To address this issue we leverage i) in- dividually fitted neural implicit avatars; ii) an alternating optimization scheme that refines pose and surface through periods of close proximity; and iii) thus segment the fused raw scans into individual instances. From these instances we compile Hi4D dataset of 4D textured scans of 20 sub- ject pairs, 100 sequences, and a total of more than 11K frames. Hi4D contains rich interaction-centric annotations in 2D and 3D alongside accurately registered parametric body models. We define varied human pose and shape esti- mation tasks on this dataset and provide results from state- of-the-art methods on these benchmarks. Hi4D dataset can be found at https://ait.ethz.ch/Hi4D .

1. Introduction While computer vision systems have made rapid progress in estimating the 3D body pose and shape of individuals and well-spaced groups, currently there are no methods that can robustly disentangle and reconstruct closely interacting people. This is in part due to the lack of suitable datasets. While some 3D datasets exist that contain human-human interactions, like ExPI [72] and CHI3D [22], they typically lack high-fidelity dynamic textured geometry, do not always provide registered parametric body models and do not always provide rich contact information and are therefore not well suited to study closely interacting people. Taking a first step towards future AI systems that are able to interpret the interactions of multiple humans in close physical interaction and under strong occlusion, we pro- pose a method and dataset that enables the study of this new setting. Specifically, we propose Hi4D, a compre- hensive dataset that contains segmented, yet complete 4D textured geometry of closely interacting humans, along- side corresponding registered parametric human models, in- stance segmentation masks in 2D and 3D, and vertex-level contact annotations (see Fig. 1). To enable research to- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17016 Dataset Multi-view Images TemporalReference Data Modalities 3D Pose Format Textured Scans Contact Annotations Instance Masks ShakeFive2 [23] ✓ Joint Positions MuPoTS-3D [53] ✓ ✓ Joint Positions ExPI [72] ✓ ✓ Joint Positions ✓ MultiHuman [77] Parametric Body Model†✓ CHI3D [22] ✓ ✓ Parametric Body Model region-level (631 events) Hi4D (Ours) ✓ ✓ Parametric Body Model ✓ vertex-level ( >6K events) ✓ Table 1. Comparison of datasets containing close human interaction.†In [77] registrations are not considered as ground-truth. wards automated analysis of close human interactions, we contribute experimental protocols for computer vision tasks that are enabled by Hi4D. Capturing such a dataset and the corresponding anno- tations is a very challenging endeavor in itself. While multi-view, volumetric capture setups can reconstruct high- quality 4D textured geometry of individual subjects, even modern multi-view systems typically fuse 3D surfaces of spatially proximal subjects into a single, connected mesh (see Fig. 1, A). Thus deriving and maintaining complete, per subject 4D surface geometry, parametric body registra- tion, and contact information from such reconstructions is non-trivial. In contrast to the case of rigid objects, simple tracking schemes fail due to very complex articulations and thus strong changes in terms of geometry. Moreover, con- tact itself will further deform the shape. To address these problems, we propose a novel method to track and segment the 4D surface of multiple closely in- teracting people through extended periods of dynamic phys- ical contact. Our key idea is to make use of emerging neu- ral implicit surface representations for articulated shapes, specifically SNARF [13], and create personalized human avatars of each individual (see Fig. 2, A). These avatars then serve as strong personalized priors to track and thus segment the fused geometry of multiple interacting people (see Fig. 2, B). To this end, we alternate between pose opti- mization and shape refinement (see Fig. 3). The optimized pose and refined surfaces yield precise segmentations of the merged input geometry. The tracked 3D instances (Fig. 1, B) then provide 2D and 3D instance masks (Fig. 1, C), vertex-level contact annotations (Fig. 1, B), and can be used to register parametric human models (Fig. 1, D). Equipped with this method, we capture Hi4D, which stands for H umans i nteracting in 4D , a dataset of humans in close physical interaction alongside high-quality 4D an- notations. The dataset contains 20 pairs of subjects (24 male, 16 female), and 100 sequences with more than 11K frames. To our best knowledge, ours is the first dataset con- taining rich interaction-centric annotations and high-quality 4D textured geometry of closely interacting humans. To provide baselines for future work, we evaluate several state-of-the-art methods for multi-person pose and shape modeling from images on Hi4D in different settings suchas monocular and multi-view human pose estimation and detailed geometry reconstruction. Our baseline experi- ments show that our dataset provides diverse and challeng- ing benchmarks, opening up new directions for research. In summary, we contribute: • A novel method based on implicit avatars to track and segment 4D scans of closely interacting humans. • Hi4D, a dataset of 4D textured scans with correspond- ing multi-view RGB images, parametric body models, instance segmentation masks and vertex-level contact. • Several experimental protocols for computer vision tasks in the close human interaction setting.

Xu_Uncovering_the_Missing_Pattern_Unified_Framework_Towards_Trajectory_Imputation_and_CVPR_2023

Abstract Trajectory prediction is a crucial undertaking in under- standing entity movement or human behavior from observed sequences. However, current methods often assume that the observed sequences are complete while ignoring the potential for missing values caused by object occlusion, scope limitation, sensor failure, etc. This limitation in- evitably hinders the accuracy of trajectory prediction. To address this issue, our paper presents a unified framework, the Graph-based Conditional Variational Recurrent Neural Network (GC-VRNN), which can perform trajectory impu- tation and prediction simultaneously. Specifically, we in- troduce a novel Multi-Space Graph Neural Network (MS- GNN) that can extract spatial features from incomplete ob- servations and leverage missing patterns. Additionally, we employ a Conditional VRNN with a specifically designed Temporal Decay (TD) module to capture temporal depen- dencies and temporal missing patterns in incomplete trajec- tories. The inclusion of the TD module allows for valuable information to be conveyed through the temporal flow. We also curate and benchmark three practical datasets for the joint problem of trajectory imputation and prediction. Ex- tensive experiments verify the exceptional performance of our proposed method. As far as we know, this is the first work to address the lack of benchmarks and techniques for trajectory imputation and prediction in a unified manner.

1. Introduction Modeling and predicting future trajectories play an in- dispensable role in various applications, i.e., autonomous driving [23, 25, 65], motion capture [57, 59], behavior un- derstanding [20, 38], etc. However, accurately predicting movement patterns is challenging due to their complex and *Work done during Yi’s internship at Honda Research Institute, under Chiho Choi’s supervision. 𝑻𝑻=𝒕𝒕𝟎𝟎 𝑻𝑻=𝒕𝒕𝟐𝟐 𝑻𝑻 𝒕𝒕𝟏𝟏𝒕𝒕𝟐𝟐 𝒕𝒕𝟎𝟎 𝒕𝒕𝒑𝒑 𝑻𝑻=𝒕𝒕𝟏𝟏Figure 1. A typical example of incomplete observed trajectory. The black car (ego-vehicle) is waiting at the intersection at time stept0, and the yellow car is moving. At time step t1, the red car is occluded by the yellow car, and the blue car appears. The bottom right figure indicates the “visibility” of four cars, where dark means visible and light color means invisible. subtle nature. Despite significant attention and numerous proposed solutions [5, 34, 62, 67–69] to the trajectory pre- diction problem, existing methods often assume that agent observations are entirely complete, which is too strong an assumption to satisfy in practice. For example, in sports games such as football or soccer, not all players are always visible in the live view due to the limitation of the camera view. In addition, the live camera always tracks and moves along the ball, resulting in some players appearing and disappearing throughout the view, depending on their relative locations to the ball. Fig. 2 illus- trates this common concept. Similar situations also arise in autonomous driving where occlusion or sensor failure can cause missing data. As illustrated in Fig. 1, at time step This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9632 𝑻𝑻=𝟏𝟏 𝑻𝑻=𝟐𝟐 𝑻𝑻=𝟑𝟑Figure 2. Three continuous frames of a “throw-and-catch” sequence in a live football match, where the ball is being circled. During this attacking play, the camera follows and zooms in on the ball. Only a subset of players is visible within the camera’s field of view. t0, there is no observation of the blue car. At time step t1, the red car is occluded by the yellow car from the black car perspective, and the blue car appears at the intersection to turn right. Predicting future trajectories of entities under these circumstances will no doubt hinder the performance and negatively influence behavior understanding of moving agents or vehicle safety operations. Although various recent works [15,17,37,53,73,74] have investigated the time series imputation problem, most are autoregressive models that impute current missing values from previous time steps, making them highly susceptible to compounding errors for long-range temporal modeling. Additionally, commonly used benchmarks [29,50,54,82] do not contain interacting entities. Some recent works [36, 78] have studied the imputation problem in a multi-agent sce- nario. Although these methods have achieved promising imputation performance, they fail to explore the relevance between the trajectory imputation and the prediction task. In fact, complete trajectory observation is essential for pre- diction, and accurate trajectory prediction can offer valu- able information about the temporal correlation between past and future, ultimately aiding in the imputation process. In this paper, we present a unified framework, Graph- based Conditional Variational Recurrent Neural Network (GC-VRNN), that simultaneously handles the trajectory im- putation and prediction. Specifically, we introduce a novel Multi-Space Graph Neural Network (MS-GNN) to extract compact spatial features of incomplete observations. Mean- while, we adopt a Conditional VRNN (C-VRNN) to model the temporal dependencies, where a Temporal Decay (TD) module is designed to learn the missing patterns of incom- plete observations. The critical idea behind our method is to acquire knowledge of the spatio-temporal features of miss- ing patterns, and then unite these two objectives through shared parameters. Sharing valuable information allows these two tasks to support and promote one another for bet- ter performance mutually. In addition, to support the joint evaluation of multi-agent trajectory imputation and pre- diction, we curate and benchmark three practical datasets from different domains, Basketball-TIP ,Football-TIP , and Vehicle-TIP , where the incomplete trajectories are gener- ated via reasonable and practical strategies. The main con-tributions of our work can be summarized as follows: • We investigate the multi-agent trajectory imputation and prediction problem and develop a unified frame- work, GC-VRNN, for imputing missing observations and predicting future trajectories simultaneously. • We propose a novel MS-GNN that can extract com- prehensive spatial features of incomplete observations and adopt a C-VRNN with a specifically designed TD module for better learning temporal missing patterns, and valuable information is shared via temporal flow. • We curate and benchmark three datasets for the multi- agent trajectory imputation and prediction problem. Strong baselines are set up for this joint problem. • Thorough experiments verify the consistent and excep- tional performance of our proposed method.

Yuan_Devil_Is_in_the_Queries_Advancing_Mask_Transformers_for_Real-World_CVPR_2023

Abstract Real-world medical image segmentation has tremen- dous long-tailed complexity of objects, among which tail conditions correlate with relatively rare diseases and are clinically significant. A trustworthy medical AI algo- rithm should demonstrate its effectiveness on tail condi- tions to avoid clinically dangerous damage in these out-of- distribution (OOD) cases. In this paper, we adopt the con- cept of object queries in Mask Transformers to formulate semantic segmentation as a soft cluster assignment. The queries fit the feature-level cluster centers of inliers dur- ing training. Therefore, when performing inference on a medical image in real-world scenarios, the similarity be- tween pixels and the queries detects and localizes OOD re- gions. We term this OOD localization as MaxQuery. Fur- thermore, the foregrounds of real-world medical images, whether OOD objects or inliers, are lesions. The differ- ence between them is less than that between the foreground and background, possibly misleading the object queries to focus redundantly on the background. Thus, we propose a query-distribution (QD) loss to enforce clear boundaries between segmentation targets and other regions at the query level, improving the inlier segmentation and OOD indi- cation. Our proposed framework is tested on two real- world segmentation tasks, i.e., segmentation of pancreatic and liver tumors, outperforming previous state-of-the-art algorithms by an average of 7.39% on AUROC, 14.69% on AUPR, and 13.79% on FPR95 for OOD localization. On the other hand, our framework improves the performance of inlier segmentation by an average of 5.27% DSC when compared with the leading baseline nnUNet.

1. Introduction Image segmentation is a fundamental task in medical im- age analysis. With the recent advancements in computer * Corresponding author. ([email protected]) †Work was done during an internship at Alibaba DAMO Academy IPMN PDAC SCN CP MCN SPT PNETCommon Pancreatic DiseasesIn-distribution AC DC Peri-pancreatic Diseases …Rare DiseasesUnseen Diseases(Near) Out-of-distribution Real -world Trustworthy Medical AIData Distribution Inlier Segmentation OOD Localization Figure 1. Real-world medical image segmentation. Real-world medical outliers (unseen, usually rare, tumors) are “near” to the inliers (labeled lesions), forming a typical near-OOD problem. A real-world medical OOD detection/localization model should fo- cus more on subtle differences between outliers and inliers than the significant difference between foreground and background.. vision and deep learning, automated medical image seg- mentation has reached expert-level performance in various applications [3, 28, 54]. Most medical image segmenta- tion methods are based on supervised machine learning that heavily relies on collecting and annotating training data. However, real-world medical images are long-tailed dis- tributed. The tail conditions are outliers and inadequate (or even unable) to train a reliable model [35, 61, 63]. Yet, the model trained with inliers is risky for triggering failures or errors in real-world clinical deployment [43]. For exam- ple, in pancreatic tumor image analysis, a miss-detection of metastatic cancer will directly threaten life; an erroneous recognition of a benign cyst as malignant will lead to unnec- essary follow-up tests and patient anxiety. Medical image segmentation models should thus demonstrate the ability to detect and localize out-of-distribution (OOD) conditions, especially in some safety-critical clinical applications. Previous studies have made valuable attempts on med- ical OOD localization [48, 65], including finding lesions This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23879 apart from normal cases or simulating OOD conditions for model validation. However, the real-world clinical scenario, such as tumor segmentation, is more complex, where either in-distribution or OOD cases have multiple types of tumors. Establishing a direct relationship between image pixels and excessive semantics (types of tumors) is difficult for real- world medical image segmentation. Using this relationship to distinguish inliers and outliers is even more challenging. Fortunately, several works about Mask Transformers [5, 9] have inspired us to split segmentation as a two-stage pro- cess of per-pixel cluster assignment and cluster classifica- tion [57, 58]. A well-defined set of inlier clusters may greatly benefit in identifying the OOD conditions from the medical images. Therefore, we propose MaxQuery, a med- ical image semantic segmentation framework that advances Mask Transformers to localize OOD targets. The frame- work adopts learnable object queries to iteratively fit inlier cluster centers. Since the affinity between OODs and an inlier cluster center should be less than that within the clus- ter (between inliers and cluster centers), MaxQuery uses the negative of such affinity as an indicator to detect OODs. Several recent works further define real-world medical image OOD localization as a near-OOD problem [43, 51], where the distribution gaps between inlier and OOD tumors are overly subtle, as shown in Fig. 1. Thus, the near-OOD problems are more difficult. Our pilot experiments show that the cluster centers redundantly represent the large re- gions of background and organ rather than tumors, compro- mising the necessary variability of the cluster assignments for OOD localization. To solve this issue, we propose the query-distribution (QD) loss to regularize specific quantities of object queries on background, organ, and tumors. This enforces the diversity of the cluster assignments, benefiting the segmentation and recognition of OOD tumors. We curate two real-world medical image datasets of (pancreatic and liver) tumor images from 1,088 patients for image segmentation and OOD localization. Specifically, we collect consecutive patients’ contrast-enhanced 3D CT imaging with a full spectrum of tumor types confirmed by pathology. In these scenarios, the OOD targets are rare tu- mors and diseases. Our method shows robust performance across two datasets, significantly outperforming the previ- ous leading OOD localization methods by an average of 7.39% in AUROC, 14.69% in AUPR, 13.79% in FPR95 for localization, and 3.42% for case-level detection. Mean- while, our framework also improves the performance of in- lier segmentation by an average of 5.27% compared with the strong baseline nnUNet [24]. We summarize our main contributions as follows: • To the best of our knowledge, we are the first to explore the near-OOD detection and localization problem in medical image segmentation. The proposed method has a strong potential for utility in clinical practice.• We propose a novel approach, MaxQuery, using the maximum score of query response as a major indicator for OOD localization. • A query-distribution (QD) loss is proposed to con- centrate the queries on important foreground regions, demonstrating superior effectiveness for near-OOD problems. • We curate two medical image datasets for tumor se- mantic segmentation/detection of real-world OODs. Our proposed framework substantially outperforms previous leading OOD localization methods and im- proves upon the inlier segmentation performance.

Ye_PVO_Panoptic_Visual_Odometry_CVPR_2023

Abstract We present PVO, a novel panoptic visual odometry frame- work to achieve more comprehensive modeling of the scene motion, geometry, and panoptic segmentation information. Our PVO models visual odometry (VO) and video panop- tic segmentation (VPS) in a unified view, which makes the two tasks mutually beneficial. Specifically, we introduce a panoptic update module into the VO Module with the guidance of image panoptic segmentation. This Panoptic- Enhanced VO Module can alleviate the impact of dynamic objects in the camera pose estimation with a panoptic-aware dynamic mask. On the other hand, the VO-Enhanced VPS Module also improves the segmentation accuracy by fusing the panoptic segmentation result of the current frame on the fly to the adjacent frames, using geometric information such as camera pose, depth, and optical flow obtained from the VO Module. These two modules contribute to each other through recurrent iterative optimization. Extensive exper- iments demonstrate that PVO outperforms state-of-the-art methods in both visual odometry and video panoptic segmen- tation tasks. ∗indicates equal contribution.†indicates the corresponding author.

1. Introduction Understanding the motion, geometry, and panoptic seg- mentation of the scene plays a crucial role in computer vision and robotics, with applications ranging from autonomous driving to augmented reality. In this work, we take a step to- ward solving this problem to achieve a more comprehensive modeling of the scene with monocular videos. Two tasks have been proposed to address this problem, namely visual odometry (VO) and video panoptic segmen- tation (VPS). In particular, VO [9, 11, 36] takes monocular videos as input and estimates the camera poses under the static scene assumption. To handle dynamic objects in the scene, some dynamic SLAM systems [2, 43] use instance segmentation network [14] for segmentation and explicitly filter out certain classes of objects, which are potentially dynamic, such as pedestrians or vehicles. However, such ap- proaches ignore the fact that potentially dynamic objects can actually be stationary in the scene, such as a parked vehicle. In contrast, VPS [17, 42, 49] focuses on tracking individual instances in the scene across video frames given some ini- tial panoptic segmentation results. Current VPS methods do not explicitly distinguish whether the object instance is moving or not. Although existing approaches broadly solve these two tasks independently, it is worth noticing that dy- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9579 namic objects in the scene can make both tasks challenging. Recognizing this relevance between the two tasks, some methods [5, 7, 19, 21] try to tackle both tasks simultaneously and train motion-semantics networks in a multi-task man- ner, shown in Fig. 2. However, the loss functions used in these approaches may contradict each other, thus leading to performance drops. In this work, we propose a novel panoptic visual odom- etry (PVO) framework that tightly couples these two tasks using a unified view to model the scene comprehensively. Our insight is that VPS can adjust the weight of VO with panoptic segmentation information (the weights of the pixels of each instance should be correlated) and VO can convert the tracking and fusion of video panoptic segmentation from 2D to 3D. Inspired by the seminal Expectation-Maximization algorithm [26], recurrent iterative optimization strategy can make these two tasks mutually beneficial. Our PVO consists of three modules, an image panoptic segmentation module, a Panoptic-Enhanced VO Module, and a VO-Enhanced VPS Module. Specifically, the panoptic segmentation module (see Sec. 3.1) takes in single images and outputs the image panoptic segmentation results, which are then fed into the Panoptic-Enhanced VO Module as ini- tialization. Note that although we choose PanopticFPN [20], any segmentation model can be used in the panoptic segmen- tation module. In the Panoptic-Enhanced VO Module (see Sec. 3.2), we propose a panoptic update module to filter out the interference of dynamic objects and hence improve the accuracy of pose estimation in the dynamic scene. In the VO-Enhanced VPS Module (see Sec. 3.3), we introduce an online fusion mechanism to align the multi-resolution fea- tures of the current frame to the adjacent frames based on the estimated pose, depth, and optical flow. This online fusion mechanism can effectively solve the problem of multiple object occlusion. Experiments show that the recurrent itera- tive optimization strategy improves the performance of both VO and VPS. Overall, our contributions are summarized as four-fold. •We present a novel Panoptic Visual Odometry (PVO) framework, which can unify VO and VPS tasks to model the scene comprehensively. •A panoptic update module is introduced and incorpo- rated into the Panoptic-Enhanced VO Module to im- prove pose estimation. •An online fusion mechanism is proposed in the VO- Enhanced VPS Module, which helps to improve video panoptic segmentation. •Extensive experiments demonstrate that the proposed PVO with recurrent iterative optimization is superior to state-of-the-art methods in both visual odometry and video panoptic segmentation tasks. Figure 2. Illustration. Our PVO unifies visual odometry and video panoptic segmentation so that the two tasks can be mutually reinforced by iterative optimization. In contrast, methods such as SimVODIS [19] optimize motion and semantic information in a multi-task manner.

Yang_Paint_by_Example_Exemplar-Based_Image_Editing_With_Diffusion_Models_CVPR_2023

Abstract Language-guided image editing has achieved great suc- cess recently. In this paper , we investigate exemplar-guidedimage editing for more precise control. We achieve thisgoal by leveraging self-supervised training to disentangleand re-organize the source image and the exemplar . How-ever , the naive approach will cause obvious fusing artifacts.We carefully analyze it and propose a content bottleneckand strong augmentations to avoid the trivial solution of di-rectly copying and pasting the exemplar image. Meanwhile,to ensure the controllability of the editing process, we de- sign an arbitrary shape mask for the exemplar image and leverage the classifier-free guidance to increase the similar-ity to the exemplar image. The whole framework involvesa single forward of the diffusion model without any itera-tive optimization. We demonstrate that our method achieves an impressive performance and enables controllable edit- ing on in-the-wild images with high fidelity. The code and *Work is done during the internship at Microsoft Research Asia. †Corresponding author.pretrained models are available at https://github. com/Fantasy-Studio/Paint-by-Example .

1. Introduction Creative editing for photos has become a ubiquitous need due to the advances in a plethora of social media plat-forms. AI-based techniques [ 37] significantly lower the barrier of fancy image editing that traditionally requiresspecialized software and labor-intensive manual operations.Deep neural networks can now produce compelling re-sults for various low-level image editing tasks, such as im-age inpainting [ 61,68], composition [ 41,67,72], coloriza- tion [ 53,71,74] and aesthetic enhancement [ 7,11], by learn- ing from richly available paired data. A more challeng-ing scenario, on the other hand, is semantic image editing,which intends to manipulate the high-level semantics of im-age content while preserving image realism. Tremendousefforts [ 2,5,36,49,55,57] have been made along this way, mostly relying on the semantic latent space of generativemodels, e.g., GANs [ 16,28,70], yet the majority of existing works are limited to specific image genres. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18381 Recent large-scale language-image (LLI) models, based on either auto-regressive models [ 12,69] or diffusion mod- els [ 19,45,50,54], have shown unprecedented generative power in modeling complex images. These models en-able various image manipulation tasks [ 3,22,30,52] previ- ously unassailable, allowing image editing for general im-ages with the guidance of text prompt. However, even thedetailed textual description inevitably introduces ambiguity and may not accurately reflect the user-desired effects; in-deed, many fine-grained object appearances can hardly bespecified by the plain language. Hence, it is crucial to de- velop a more intuitive approach to ease fine-grained image editing for novices and non-native speakers. In this work, we propose an exemplar-based image edit- ingapproach that allows accurate semantic manipulation on the image content according to an exemplar image provided by users or retrieved from the database. As the saying goes,“a picture is worth a thousand words”. We believe imagesbetter convey the user’s desired image customization in amore granular manner than words. This task is completelydifferent from image harmonization [ 20,64] that mainly fo- cuses on color and lighting correction when compositing the foreground objects, whereas we aim for a much more com-plex job: semantically transforming the exemplar, e.g., pro- ducing a varied pose, deformation or viewpoint, such thatthe edited content can be seamlessly implanted accordingto the image context. In fact, ours automates the traditional image editing workflow where artists perform tedious trans-formations upon image assets for coherent image blending. To achieve our goal, we train a diffusion model [ 24,59] conditioned on the exemplar image. Different from text- guided models, the core challenge is that it is infeasibleto collect enough triplet training pairs comprising sourceimage, exemplar and corresponding editing ground truth. One workaround is to randomly crop the objects from theinput image, which serves as the reference when training the inpainting model. The model trained from such a self- reference setting, however, cannot generalize to real exem- plars, since the model simply learns to copy and paste thereference object into the final output. We identify severalkey factors that circumvent this issue. The first is to utilizea generative prior. Specifically, a pretrained text-to-image model has the ability to generate high-quality desired re- sults, we leverage it as initialization to avoid falling intothe copy-and-paste trivial solution. However, a long timeof finetuning may still cause the model to deviate from theprior knowledge and ultimately degenerate again. Hence,we introduce the content bottleneck for self-reference con-ditioning in which we drop the spatial tokens and only re-gard the global image embedding as the condition. In thisway, we enforce the network to understand the high-levelsemantics of the exemplar image and the context from thesource image, thus preventing trivial results during the self-supervised training. Moreover, we apply aggressive aug- mentation on the self-reference image which can effectivelyreduce the training-test gap. We further improve the editability of our approach in two aspects. One is that our training uses irregular randommasks so as to mimic the casual user brush used in practi-cal editing. We also prove that classifier-free guidance [ 25] is beneficial to boost both the image quality and the style resemblance to the reference. The proposed method, Paint by Example , well solves semantic image composition problem where the reference is semantically transformed and harmonized before blend- ing into another image, as shown in Figure 1. Our method shows a significant quality advantage over prior works in asimilar setting. Notably, our editing just involves a singleforward of the diffusion model without any image-specificoptimization, which is a necessity for many real applica-tions. To summarize, our contributions are as follows: • We propose a new image editing approach, Paint by Ex- ample , which semantically alters the image content based on an exemplar image. This approach offers fine-grainedcontrol while being convenient to use. • We solve the problem with an image-conditioned diffu- sion model trained in a self-supervised manner. We pro-pose a group of techniques to tackle the degenerate chal-lenge. • Our approach performs favorably over prior arts for in- the-wild image editing, as measured by both quantitativemetrics and subjective evaluation.

Zhang_Lookahead_Diffusion_Probabilistic_Models_for_Refining_Mean_Estimation_CVPR_2023

Abstract We propose lookahead diffusion probabilistic models (LA-DPMs) to exploit the correlation in the outputs of the deep neural networks (DNNs) over subsequent timesteps in diffusion probabilistic models (DPMs) to refine the mean estimation of the conditional Gaussian distributions in the backward process. A typical DPM first obtains an estimate of the original data sample xby feeding the most recent state ziand index iinto the DNN model and then computes the mean vector of the conditional Gaussian distribution for zi1. We propose to calculate a more accurate estimate for xby performing extrapolation on the two estimates of x that are obtained by feeding (zi+1,i+1) and(zi,i)into the DNN model. The extrapolation can be easily integrated into the backward process of existing DPMs by introducing an additional connection over two consecutive timesteps, and fine-tuning is not required. Extensive experiments showed that plugging in the additional connection into DDPM, DDIM, DEIS, S-PNDM, and high-order DPM-Solvers leads to a significant performance gain in terms of Fr ´echet in- ception distance (FID) score. Our implementation is avail- able at https://github.com/guoqiang-zhang- x/LA-DPM .

1. Introduction As one type of generative model, diffusion probabilis- tic models (DPMs) have made significant progress in recent years. The pioneering work [ 17] applied non-equilibrium statistical physics to estimating probabilistic data distribu- tions. In doing so, a Markov forward diffusion process is constructed by systematically inserting additive noise in the data until essentially only noise remains. The data distribu- tion is then gradually restored by a reverse diffusion pro- cess starting from a simple parametric distribution. The main advantage of DPMs over classic tractable models (e.g., HMMs, GMMs, see [ 5]) is that they can accurately modelboth the high and low likelihood regions of the data distribu- tion via the progressive estimation of noise-perturbed data distributions. In comparison to generative adversarial net- works (GANs) [ 1,8,9], DPMs exhibit more stable training dynamics by avoiding adversarial learning. The work [ 10] focuses on a particular type of DPM, namely a denoising diffusion probabilistic model (DDPM), and shows that after a sufficient number of timesteps (or equivalently iterations) in the backward process, DDPM can achieve state-of-the-art performance in image genera- tion tasks by the proper design of a weighted variational bound (VB). In addition, by inspection of the weighted VB, it is found that the method score matching with Langevin dy- namics (SMLD) [ 19,20] can also be viewed as a DPM. The recent work [ 21] interprets DDPM and SMLD as search of approximate solutions to stochastic differential equations. See also [ 15] and [ 7] for improved DPMs that lead to better log-likelihood scores and sampling qualities, respectively. One inconvenience of a standard DPM is that the asso- ciated deep neural network (DNN) needs to run for a suf- ficient number of timesteps to achieve high sampling qual- ity while the generative model of a GAN only needs to run once. This has led to an increasing research focus on re- ducing the number of reverse timesteps in DPMs while re- taining a satisfactory sampling quality (see [ 22] for a de- tailed overview). Song et al. proposed the so-called de- noising diffusion implicit model (DDIM) [ 18] as an exten- sion of DDPM from a non-Markov forward process point of view. The work [ 11] proposed to learn a denoising schedule in the reverse process by explicitly modeling the signal-to- noise ratio in the image generation task. [ 6] and [ 12] consid- ered effective audio generation by proposing different noise scheduling schemes in DPMs. Differently from the above methods, the recent works [ 4] and [ 3] proposed to estimate the optimal variance of the backward conditional Gaussian distribution to improve sampling qualities for both small and large numbers of timesteps. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1421 Another approach for improving the sampling quality of DPMs with a limited computational budget is to exploit high-order methods for solving the backward ordinary dif- ferential equations (ODEs) (see [ 21]). The authors of [ 13] proposed pseudo numerical methods for diffusion models (PNDM), of which high-order polynomials of the estimated Gaussian noises {ˆ✏✓(zi+j,i+j)|rj0}are intro- duced to better estimate the latent variable zi1at iteration i, where ˆ✏✓represents a pre-trained neural network model for predicting the Gaussian noises. The work [ 23] further extends [ 13] by refining the coefficients of the high-order polynomials of the estimated Gaussian noises, and proposes the diffusion exponential integrator sampler (DEIS). Re- cently, the authors of [ 14] considered solving the ODEs of a diffusion model differently from [ 23]. In particular, a high- order Taylor expansion of the estimated Gaussian noises was employed to better approximate the continuous solu- tions of the ODEs, where the developed sampling methods are referred to as DPM-Solvers. We note that the computation of zi1at timestep iin the backward process of existing DPMs (including the high- order ODE solvers) can always be reformulated in terms of an estimate ˆxfor the original data sample xin combination with other terms. In principle, as the timestep idecreases, the estimate ˆxwould become increasingly accurate. In this paper, we aim to improve the estimation accuracy of xat each timestep iin computation of the mean vector for the la- tent variable zi1. To do so, we propose to make an extrap- olation from the two most recent estimates of xobtained at timestep iandi+1. The extrapolation allows the backward process to look ahead towards a noisy direction targeting x, thus improving the estimation accuracy. The extrapola- tion can be realized by simply introducing additional con- nections between two consecutive timesteps, which can be easily plugged into existing DPMs with negligible computa- tional overhead. We refer to the improved diffusion models as Lookahead-DPMs (LA-DPMs). We conducted an extensive evaluation by plugging in the additional connection into the backward process of DDPM, DDIM, DEIS, S-PNDM, and DPM-Solver. Interestingly, it is found that the performance gain of LA-DPMs is more significant for a small number of timesteps. This makes it attractive for practical applications as it is computationally preferable to run the backward process in a limited number of timesteps.

Zhang_MD-VQA_Multi-Dimensional_Quality_Assessment_for_UGC_Live_Videos_CVPR_2023

Abstract User-generated content (UGC) live videos are often bothered by various distortions during capture procedures and thus exhibit diverse visual qualities. Such source videos are further compressed and transcoded by media server providers before being distributed to end-users. Because of the flourishing of UGC live videos, effective video qual- ity assessment (VQA) tools are needed to monitor and per- ceptually optimize live streaming videos in the distributing process. In this paper, we address UGC Live VQA prob- lems by constructing a first-of-a-kind subjective UGC Live VQA database and developing an effective evaluation tool. Concretely, 418 source UGC videos are collected in real live streaming scenarios and 3,762 compressed ones at dif- ferent bit rates are generated for the subsequent subjective VQA experiments. Based on the built database, we de- velop a M ulti-D imensional VQA (MD-VQA) evaluator to measure the visual quality of UGC live videos from seman- tic, distortion, and motion aspects respectively. Extensive experimental results show that MD-VQA achieves state-of- the-art performance on both our UGC Live VQA database and existing compressed UGC VQA databases.

1. Introduction With the rapid development of social media applications and the advancement of video shooting and processing tech- nologies, more and more ordinary people are willing to tell their stories, share their experiences, and have their voice heard on social media or streaming media platforms such as Twitch, Tiktok, Taobao, etc. However, due to the lack of photography skills and professional equipment, the vi- *These authors contributed equally to this work. The database is avail- able at https://tianchi.aliyun.com/dataset/148818?t=1679581936815. †Corresponding author. Users UGC Distortion UploadDistribute Live Platforms Viewers Transcoding Quality Monitoring OptimizationFigure 1. The distributing process of UGC live videos, where the users upload the videos degraded by the UGC distortions to the live platforms and the distorted videos are further compressed be- fore being distributed to the viewers. The VQA models can mon- itor the quality changes of the compressed UGC live videos and adaptively optimize the transcoding setting. sual quality of user-generated content (UGC) videos may be degraded by in-the-wild distortions [51]. What’s more, in common live platforms, live videos are encoded and dis- tributed to end-users with very low delay, where compres- sion algorithms have a significant influence on the visual quality of live videos because they can greatly reduce trans- mission bandwidth. As illustrated in Fig. 1, video quality assessment (VQA) tools play an important role in monitor- ing, optimizing, and further improving the Quality of Expe- rience (QoE) of end-users in UGC live streaming systems. In the literature, many UGC VQA databases have been carried out [14, 21,36,45,51] to address the impact of gen- eral in-the-wild distortions on video quality, while some compression VQA databases [22, 34,40] are proposed to study the influence of compression artifacts. Then some compressed UGC VQA databases [1, 21,46] are further constructed to solve the problem of assessing the quality of UGC videos with compression distortions. However, they are either small in scale or employ high-quality UGC videos as the sources, and all of the mentioned databases lack videos in live streaming scenes. Therefore, there is a lack of a proper UGC Live VQA database to develop and validate the video quality measurement tools for live streaming systems. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1746 Table 1. Review of common VQA databases, where ’UGC+Compression’ refers to manually encoding the UGC videos with different compression settings. Database Year Duration/s Ref. Num. Scale Scope Subjective Evaluating Format CVD2014 [31] 2014 10-25 - 234 In-capture In-lab LIVE-Qualcomm [12] 2016 15 - 208 In-capture In-lab KoNViD-1k [14] 2017 8 - 1,200 In-the-wild Crowdsourced LIVE-VQC [36] 2018 10 - 585 In-the-wild Crowdsourced YouTube-UGC [45] 2019 20 - 1,500 In-the-wild Crowdsourced LSVQ [51] 2021 5-12 - 39,075 In-the-wild Crowdsourced UGC-VIDEO [21] 2019 >10 50 550 UGC + Compression In-lab LIVE-WC [53] 2020 10 55 275 UGC + Compression In-lab YT-UGC+(Subset) [46] 2021 20 189 567 UGC + Compression In-lab ICME2021 [1] 2021 - 1,000 8,000 UGC + Compression In-lab TaoLive(proposed) 2022 8 418 3,762 UGC + Compression In-lab To address UGC Live VQA problems, we first con- struct a large-scale database named TaoLive , consisting 418 source UGC videos from the TaoBao [2] live streaming platform and the corresponding 3,762 compressed videos at various bit rates. Then we perform a subjective experiment in a well-controlled environment. Afterward, we propose a no-reference (NR) M ulti-D imensional VQA (MD-VQA ) model to measure the visual quality of UGC live videos in terms of semantic, distortion, and motion aspects. The semantic features are extracted by pretrained convolutional neural network (CNN) model; the distortion features are ex- tracted by specific handcrafted image distortion descriptors (i.e. blur, noise, block effect, exposure, and colorfulness); and the motion features are extracted from video clips by pretrained 3D-CNN models. Compared with existing UGC VQA algorithms, MD-VQA measures visual quality from multiple dimensions, and these dimensions correspond to key factors affecting live video quality, which thereby has better interpretability and performance. The contributions of this paper are summarized as below: •We build a large-scale UGC Live VQA database targeted at the compression artifacts on the UGC live videos. We collect 418 raw UGC live videos that are diverse in content, distortion, and quality. Then 8 encoding settings are used, which provides 3,762 com- pressed UGC live videos in total. •We carry out a well-controlled in-lab subjective ex- periment. 44 participants are invited to participate in the subjective experiment and a total of 165,528 sub- jective annotations are gathered. •A multi-dimensional NR-VQA model is proposed, using pretrained 2D-CNN, handcrafted distortion de- scriptors, and pretrained 3D-CNN for the semantic, distortion, and motion features extraction respectively. The extracted features are then spatio-temporally fused to obtain the video-level quality score. The extensive experimental results validate the effectiveness of the proposed method. $YJ  6WG$YJ 6WG(a) /uni00000015/uni00000011/uni00000018 /uni00000016/uni00000011/uni00000013 /uni00000016/uni00000011/uni00000018 /uni00000017/uni00000011/uni00000013 /uni00000017/uni00000011/uni00000018 /uni00000018/uni00000011/uni00000013 MOS/uni00000013/uni00000011/uni00000013/uni00000013/uni00000013/uni00000011/uni00000013/uni00000018/uni00000013/uni00000011/uni00000014/uni00000013/uni00000013/uni00000011/uni00000014/uni00000018/uni00000013/uni00000011/uni00000015/uni00000013/uni00000013/uni00000011/uni00000015/uni00000018/uni00000013/uni00000011/uni00000016/uni00000013Probability/uni00000037/uni00000044/uni00000052/uni0000002f/uni0000004c/uni00000059/uni00000048/uni00000042/uni00000035/uni00000048/uni00000049 /uni0000002c/uni00000026/uni00000030/uni00000028/uni00000015/uni00000013/uni00000015/uni00000014/uni00000042/uni00000035/uni00000048/uni00000049 (b) Figure 2. Comparison of the quality distribution of reference videos between the TaoLive and ICME2021 [1] databases. The source UGC videos in the ICME2021 database are centered on high-quality levels while the quality levels of the source UGC videos in the TaoLive database are more diverse.

Zhang_Weakly_Supervised_Segmentation_With_Point_Annotations_for_Histopathology_Images_via_CVPR_2023

Abstract Image segmentation is a fundamental task in the field of imaging and vision. Supervised deep learning for segmen- tation has achieved unparalleled success when sufficient training data with annotated labels are available. However, annotation is known to be expensive to obtain, especially for histopathology images where the target regions are usu- ally with high morphology variations and irregular shapes. Thus, weakly supervised learning with sparse annotations of points is promising to reduce the annotation workload. In this work, we propose a contrast-based variational model to generate segmentation results, which serve as reliable complementary supervision to train a deep segmentation model for histopathology images. The proposed method considers the common characteristics of target regions in histopathology images and can be trained in an end-to-end manner. It can generate more regionally consistent and smoother boundary segmentation, and is more robust to un- labeled ‘novel’ regions. Experiments on two different his- tology datasets demonstrate its effectiveness and efficiency in comparison to previous models. Code is available at: https://github.com/hrzhang1123/CVM_WS_ Segmentation .

1. Introduction Histopathology images are of great importance for clini- cal diagnosis and prognosis of diseases. With the thriving of artificial intelligence (AI) techniques over the past decade, especially deep learning, automatic analysis of histopathol- ogy images in some tasks has achieved comparable or even †: Equal contribution; ∗: Corresponding author Figure 1. Two examples of histopathology images with target re- gions (e.g. tumor (top row) and stroma (bottom row)) annotated by contours (b) or in-target and out-of-target points (c). surpassing performance in comparison with human pathol- ogists’ reviewing [2, 13, 28, 50]. However, most competent methods are based on supervised learning, and their perfor- mances critically rely on a large number of training samples with detailed annotations. Yet such annotations usually re- quire experienced pathologists and are expensive (in terms of cost and time consumption) to obtain, and also subject to human errors. The annotation problem for histopathol- ogy images is particularly demanding, not only due to the large size of such an image but also resulting from irregular shapes of target tissues to be annotated (See Figure.1). Weakly supervised learning is a promising solution to alleviate the issues of obtaining annotations. The anno- tations for the “weakly supervised” can specifically re- fer to image-level labelling (multiple instance learning) [23, 25, 27, 41, 48, 51], partial annotations within an image (point or scribble annotation) [26], or full annotation in par- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15630 tial images (semi-supervised learning) [31]. Amongst these three categories, learning by partial annotations has excel- lent target localization capability, yet requires comparably less cost to annotate. Interactive segmentation with scribbles or points has been widely studied for a few decades [4]. Conventional methods relied on user interactive input for object segmen- tation, such as grab-cut [39], Graph-cut [4], active contour- based selective segmentation [37], random walker [17]. In recent years it has been a hot topic to develop segmentation models that can be trained by utilizing only the scribble or point annotations, formulated as partially-supervised learn- ing problem [8, 12, 22, 24, 26, 34, 36, 46, 49, 52]. Yet, exist- ing partially-supervised methods were designed mainly for natural images or medical images with relatively regular- shaped objects and consistent textures, and very few are di- rectly applicable to histopathology images given the above challenges. In this work, we focus on partially-supervised learning for histopathology image segmentation based on in-target and out-of-target point annotations, where in-target points refer to those labelled inside the target regions and out- of-target points refer to the outside ones. Histopathology images are significantly distinguished from other types of images. In many histopathology images, the target objects present distinct regional characteristics. Specifically, as the example shown in Figure.1, the tumors usually cluster in- side large regions, in which morphological features or tex- tures are similar and visually different to the outside re- gions, and there exist comparably clear boundaries. Ex- isting works on partially-supervised learning do not well utilize this characteristic. Moreover, a histopathology im- age scanned from a tissue section often contains some non- target regions that are visually or morphologically unique to those in other images. If such regions are not labelled, they will be ‘novel’ to a trained machine learning model, and the predicted categories of them will depend on their similarity to the neutral tissue and to the target tissue. If such novel re- gions are more similar to target tissues, they will be wrongly predicted as the target category, leading to false detections. Existing methods are limited in tackling this situation, es- pecially for those methods based on consistency training on data augmentation [15, 30], as consistency supervision may amplify such errors. Based on the above observations and insights, we pro- pose to adopt a variational segmentation model to provide complementary supervision information to assist the train- ing of a deep segmentation model. This variational seg- mentation model itself will be guided by annotated in-target and out-of-target points for the segmentation of target re- gions in the images. Variational methods are powerful tools for segmenting structures in an image. Often posed as an optimisation problem, energy functionals can be carefullyconstructed to satisfy certain desired properties, for exam- ple, maintaining consistency inside the evolving contour, or constraining the length of the boundary to ensure a smooth boundary [11, 40]. The uniqueness of variational methods highly fits the characteristic of target regions in histopathology images, as mentioned above. However, existing variational methods cannot be directly applied to high-dimensional deep fea- tures, which contain higher-level semantic information. To tackle this problem, we introduce the concept of contrast maps derived from deep feature correlation maps and for- mulate a variational model applied to the obtained contrast maps. Specifically, a set of correlation maps are generated based on the annotated points on an image. The correspond- ing contrast maps can then be obtained by the pairs of corre- lation maps of in-target and out-of-target points through the subtraction operation. A variational formulation is used to aggregate the obtained contrast maps for the final segmen- tation result. Finally, the variational segmentation provides complementary supervision to the deep segmentation model through the uncertainty-aware Kullback–Leibler (KL) di- vergence. Besides, the proposed model can alleviate the aforementioned issue of unlabeled novel tissue regions, re- sulting from the subtraction operation in obtaining the con- trast maps. In summary, the main contribution of this work is the for- mulated contrast-based variational model, used as reliable complementary supervision for training a deep segmenta- tion model from weak point annotations for histopathology images. The variational model is based on the proposed new contrast maps, which incorporate the correlations between each location in an image and the annotated in-target and out-of-target points. The proposed model is well suited for the segmentation of histopathology images and is robust to unlabeled novel regions. The effectiveness and efficiency of the proposed method are empirically proven by the experi- mental results on two different histology datasets.

Yang_DeCo_Decomposition_and_Reconstruction_for_Compositional_Temporal_Grounding_via_Coarse-To-Fine_CVPR_2023

Abstract Understanding dense action in videos is a fundamen- tal challenge towards the generalization of vision models. Several works show that compositionality is key to achiev- ing generalization by combining known primitive elements, especially for handling novel composited structures. Com- positional temporal grounding is the task of localizing dense action by using known words combined in novel ways in the form of novel query sentences for the actual grounding. In recent works, composition is assumed to be learned from pairs of whole videos and language embeddings through large scale self-supervised pre-training. Alternatively, one can process the video and language into word-level prim- itive elements, and then only learn fine-grained semantic correspondences. Both approaches do not consider the gran- ularity of the compositions, where different query granularity corresponds to different video segments. Therefore, a good compositional representation should be sensitive to different video and query granularity. We propose a method to learn a coarse-to-fine compositional representation by decomposing the original query sentence into different granular levels, and then learning the correct correspondences between the video and recombined queries through a contrastive rank- ing constraint. Additionally, we run temporal boundary prediction in a coarse-to-fine manner for precise ground- ing boundary detection. Experiments are performed on two datasets, Charades-CG and ActivityNet-CG, showing the superior compositional generalizability of our approach.

1. Introduction Over the last years, robust results have been shown for the detection of predefined, simpler action classes in video [11, 38, 41, 48]. On the other hand, the detection of dense ac- tion, e.g. action contents described by a rich description, still poses a significant challenge due to the large diversity of pos- sible language descriptions and semantic associations [16]. * Work done while Lijin Yang was an Intern at Woven by Toyota.Recent work in this area treats complex actions as monolithic events via end-to-end predictions [27, 44]. They produce a single label to densely describe a long video sequence, and they are difficult to scale up to more varied patterns. Fundamentally, most complex actions consist of a series of simpler events, and thus a dense action can be treated as a composition of known event primitives [13, 25, 26]. Such a compositional representation can allow for a higher degree of model generalization. By learning from a finite number of action composites, and recombining their constituent event primitives in novel ways for unseen action descriptions, the representation can expand to large numbers of novel sce- narios that have not been observed in the original action space. In this case, the action description is not treated as a single label but as a language modality that allows to learn finer-grained video and language correspondence. The task of temporal video grounding concentrates on the described setting. Given a video and an action-related query sentence, the model has to output the start and end times of the specific moment that semantically corresponds to the given query sentence. On top of temporal grounding, Compositional Temporal Grounding is a new task for testing whether the model can generalize to sentences that contain novel compositions of seen words. Several recent methods [12, 47, 50] encode both sentence and video segments into unstructured global repre- sentations and then devise specific cross-modal interaction modules to fuse them for final prediction. Unfortunately, such global representations fail to explicitly model video structure and language compositions, and fail for cases where higher granularity is required. Similar to the above approach of correspondence learning via global information, compos- ite representations learn from paired monolithic video and language respectively, through large-scale self-supervised pre-training [32]. In either case, the main challenge of utiliz- ing datasets of limited action size to achieve generalization towards novel actions remains challenging due to its combi- natorial complexity. Differently to the global representation approaches, VISA [17] introduces a compositional method toward the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23130 “News reporter in blue are giving news about people cutting dogs’ hair and trimming dogs’ nails.”48.8s10.7sTemporal boundary “people cuts dogs’ hair”“people trimsdogs’ hair”???“reporter in blue”𝑊!𝑊"𝑊#[Verb][Noun]Learnable promptQuery A = selected global sentenceQuery B = different global sentence from A Sub-sentences (primitive events)Query AQuery BQuery A“News reporter in blue are giving news about people cutting dogs’ hair and trimming dogs’ nails.”Masking Positive anchor ANegative AB[Verb] from A, [Noun] from BNegative B[Verb] from B, [Noun] from B[Verb] from A, [Noun] from AReconstructionlogitsQueryA >positive A>negative AB>negative BFigure 1. Our proposed method of decomposition and reconstruction of masked words with a coarse-to-fine scheme. temporal grounding task. They parse video and language into several primitive elements, and then learn fine-grained semantic correspondences. Specifically, the composite rep- resentation is extracted from a graph that is structured with primitive elements, and these elements serve as a common representation for vision-language correspondence learning. However, both the global and compositional approaches dis- regard the granularity of the action composition during the learning phase, and can have issues when generalizing to more varied action spaces. Based on these considerations, we argue that a good com- posite representation should be sensitive to different levels of granularity of both the action and the query language. We provide an overview of our idea in Fig. 1. Concretely, we propose to learn a composite representation in a coarse- to-fine manner, where we first decompose the whole query sentence into several simple phrases as subsentences, and then learn correspondences not only globally across query and video, but also between the subsentences and their re- lated primitive sequences from the video. Since there is no ground truth to relate subsentences and their corresponding sequences, we propose to learn correspondence by decom- posing and reconstructing them under weak supervision. For each subsentence as an anchor sample, we generate a tem- poral proposal and learn the positive representation through the mining of negative samples. To structure the overall weak supervision, we mask the words in a given anchor query, and use the embedding of the positive and negative query with the related pseudo temporal segments from the video respectively to do the reconstruc- tion for the masked words. The negative sample is another subsentence but includes the words from the other query action description, to allow the model sensitive to the word level variation for composition. According to the fact that the negative subsentence contains the novel word compared to the positive anchor sub-sentence, we could rank the recon-struction quality according to the given query sub-sentence as a natural prior constraint during the training. Additionally, the ground truth that links the temporal boundary to the global sentence provides supervision for the coarse compositional representation. The generated subsen- tences from the same global query are recombined into a new sentence that should be as informative as the original global query, and then this recombined sentence is used with the original query sentence to estimate the temporal bound- ary via supervised learning. The whole process is trained end-to-end. Our contributions are summarized below: (i). We argue that a good compositional representation should be sensitive to action granularity of video and query language, and propose a coarse-to-fine decomposition ap- proach to this end. (ii). To learn a compositional representation without composite-level labels, we decompose events from global queries and learn primitive-level correspondences between video and language with a weakly-supervised contrastive ranking in form of word-masked reconstruction. (iii). The decomposed events are recombined to a novel query along with the original query for temporal boundary estimation supervised, jointly learned end-to-end way with a coarse-to-fine structure. (iv). Experiments on Charades-CG and ActivityNet- CG [17] demonstrate our method significantly outperforms existing methods about compositional temporal grounding tasks.

Yu_PanelNet_Understanding_360_Indoor_Environment_via_Panel_Representation_CVPR_2023

Abstract Indoor 360 panoramas have two essential properties. (1) The panoramas are continuous and seamless in the hori- zontal direction. (2) Gravity plays an important role in in- door environment design. By leveraging these properties, we present PanelNet, a framework that understands indoor envi- ronments using a novel panel representation of 360 images. We represent an equirectangular projection (ERP) as consec- utive vertical panels with corresponding 3D panel geometry. To reduce the negative impact of panoramic distortion, we in- corporate a panel geometry embedding network that encodes both the local and global geometric features of a panel. To capture the geometric context in room design, we introduce Local2Global Transformer, which aggregates local informa- tion within a panel and panel-wise global context. It greatly improves the model performance with low training overhead. Our method outperforms existing methods on indoor 360 depth estimation and shows competitive results against state- of-the-art approaches on the task of indoor layout estimation and semantic segmentation.

1. Introduction Understanding indoor environments is an important topic in computer vision as it is crucial for multiple practical appli- cations such as room reconstruction, robot navigation, and virtual reality applications. Early methods focus on mod- eling indoor scenes using perspective images [9, 10, 19]. With the development of CNNs and omnidirectional pho- tography, many works turn to understand indoor scenes us- ing panorama images. Compared to the perspective images, panorama images have a larger field-of-view (FoV) [43] and provide the geometric context of the indoor environment in a continuous way [24]. There are several 360 input formats used in indoor scene understanding. One of the most commonly-used formats is the equirectangular projection (ERP). Modeling the holistic scene from an ERP is challenging. The ERP distortion in- creases when pixels are close to the zenith or nadir of the image, which may decrease the power of the convolutional PanelNet Indoor Scene UnderstandingPanel RepresentationSemanticDepth LayoutFigure 1. An overview of the proposed system. We present Panel- Net, a network that learns the indoor environment using a novel panel representation of ERP. We formulate the panel representation as consecutive ERP panels with corresponding global and local geometry. By slightly modifying the network structure, PanelNet is capable of tackling major 360 indoor understanding tasks such as depth estimation, semantic segmentation and layout prediction. structures designed for distortion-free perspective images. To eliminate the negative effects of ERP distortion, recent works [8, 21, 28] focus on decomposing the whole panorama into perspective patches, i.e., tangent images. However, par- titioning a panorama into discontinuous patches breaks the local continuity of gravity-aligned scenes and objects which limits the performance of these works. To reduce the im- pact of distortion while preserving the local continuity, we present PanelNet, a novel network to understand the indoor scene from equirectangular projection. We design our PanelNet based on two essential properties of equirectangular projection. (1) The ERP is continuous and seamless in the horizontal direction. (2) Gravity plays an important role in indoor environment design, which makes the gravity-aligned features crucial for indoor 360 under- standing [24, 31]. Following these two properties, we tackle the challenges above through a novel panel representation of ERP. We represent an ERP as consecutive panels with cor- responding global and local 3D geometry, which preserves the gravity-aligned features within a panel and maintains the global continuity of the indoor structure across panels. Inspired by Omnifusion [21], we design a geometry embed- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 878 ding network for panel representations that encodes both local and global features of panels to reduce the negative effects of ERP distortion without adding further explicit dis- tortion fixing modules. We further introduce Local2Global Transformer as a feature processor. Considering the nature of panel representation, we design this Transformer with Window Blocks for local information aggregation and Panel Blocks for panel-wise context capturing. The main contribu- tions of our work are: •We represent the ERP as consecutive vertical panels with corresponding 3D geometry. We introduce Panel- Net, a novel indoor panorama understanding pipeline using the panel representation. Following the essential geometric properties of the indoor equirectangular pro- jection, our framework outperforms existing methods on the task of indoor 360 depth estimation and shows competitive results on other indoor scene understand- ing tasks such as semantic segmentation and layout prediction. •We propose a panel geometry embedding network that encodes both local and global geometric features of pan- els and reduces the negative impact of ERP distortion implicitly while preserving the geometric continuity. •We design Local2Global Transformer as a feature pro- cessor, which greatly enhances the continuity of geo- metric features and improves the model performance by successfully aggregating the local information within a panel and capturing panel-wise context accurately.

Yu_Fusing_Pre-Trained_Language_Models_With_Multimodal_Prompts_Through_Reinforcement_Learning_CVPR_2023

Abstract Language models are capable of commonsense reason- ing: while domain-specific models can learn from ex- plicit knowledge ( e.g. commonsense graphs [6], ethical norms [25]), and larger models like GPT-3 [7] mani- fest broad commonsense reasoning capacity. Can their knowledge be extended to multimodal inputs such as im- ages and audio without paired domain data? In this work, we propose ESPER (Extending Sensory PErception with Reinforcement learning) which enables text-only pre- trained models to address multimodal tasks such as visual commonsense reasoning. Our key novelty is to use rein- forcement learning to align multimodal inputs to language model generations without direct supervision: for example, our reward optimization relies only on cosine similarity de- rived from CLIP [52] and requires no additional paired (image, text) data. Experiments demonstrate that ESPER outperforms baselines and prior work on a variety of mul- timodal text generation tasks ranging from captioning to commonsense reasoning; these include a new benchmark we collect and release, the ESP dataset, which tasks mod- els with generating the text of several different domains for each image. Our code and data are publicly released at https://github.com/JiwanChung/esper .

1. Introduction Collecting multimodal training data for new domains can be a Herculean task. Not only is it costly to assemble mul- timodal data, but also curated datasets cannot completely cover a broad range of skills, knowledge, and form ( e.g. free text, triplets, graphs, etc.). Ideally, we want to endow *denotes equal contribution ESPERGenerateNews :Dialog :Story : Prompt PLM CLIPA: Is the cat sleeping? B: YesDogs with vegan diets are healthier, study suggestsI was walking to the veterinarian. My dog was having a bad day.0.770.410.69CLIPRewardFigure 1. The intuition of ESPER ,Extending Sensory PErception with Reinforcement learning. To better align knowledge in CLIP and pretrained language models (PLM), we use CLIP as a reward for the pairs of images and self-generated text. multimodal models with diverse reasoning capacity ( e.g. ethics [25], commonsense [57], etc.) without undertaking a separate multimodal annotation effort each time. In this work, we propose Extending Sensory PErception with Reinforcement learning( ESPER ), a new framework that enables a pre-trained language model to accept multimodal inputs like images and audio. ESPER extends diverse skills embodied by the pre-trained language model to similarly diverse multimodal capabilities, all without requiring ad- ditional visually paired data. In a zero-shot fashion, our model generates text conditioned on an image: using this interface, we show that ESPER is capable of a diverse range of skills, including visual commonsense [50], news [39], di- alogues [60], blog-style posts [28], and stories [22]. ESPER combines insights from two previously disjoint lines of work: multimodal prompt tuning and reinforce- ment learning . Like prior multimodal prompt tuning work, ESPER starts from a base language-only model (e.g., GPT- 2 [53], COMET [6]), keeps most of its parameters frozen, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10845 and trains a small number of encoder parameters to map visual features into the embedding space of the language model [40, 45, 68]. Unlike prior works, however, ESPER does not train these parameters using maximum likelihood estimation over a dataset of aligned (image, caption) pairs. Instead, it uses a reinforcement learning objective. During training, the model is first queried for completions condi- tioned on visual features. Then, the lightweight vision-to- text encoder is updated using proximal policy optimization (PPO) [59] to maximize a similarity score computed with a secondary pre-trained image-caption model, CLIP [52]. As a result, the frozen language model can interpret the mul- timodal inputs in the same context as the text embedding space without additional human-annotated paired data. Reinforcement learning has two advantages over maxi- mum likelihood objectives.

Yu_Rotation-Invariant_Transformer_for_Point_Cloud_Matching_CVPR_2023

Abstract The intrinsic rotation invariance lies at the core of matching point clouds with handcrafted descriptors. How-ever , it is widely despised by recent deep matchers thatobtain the rotation invariance extrinsically via data aug-mentation. As the finite number of augmented rotationscan never span the continuous SO(3) space, these meth- ods usually show instability when facing rotations that are rarely seen. To this end, we introduce RoITr, a Rotation- Invariant Transformer to cope with the pose variations in the point cloud matching task. We contribute both onthe local and global levels. Starting from the local level, we introduce an attention mechanism embedded with Point Pair Feature (PPF)-based coordinates to describe the pose-invariant geometry, upon which a novel attention-basedencoder-decoder architecture is constructed. We furtherpropose a global transformer with rotation-invariant cross-frame spatial awareness learned by the self-attention mech-anism, which significantly improves the feature distinctive-ness and makes the model robust with respect to the lowoverlap. Experiments are conducted on both the rigid andnon-rigid public benchmarks, where RoITr outperforms allthe state-of-the-art models by a considerable margin in thelow-overlapping scenarios. Especially when the rotations are enlarged on the challenging 3DLoMatch benchmark,RoITr surpasses the existing methods by at least 13 and 5percentage points in terms of Inlier Ratio and RegistrationRecall, respectively. Code is publicly available 1.

1. Introduction The correspondence estimation between a pair of partially-overlapping point clouds is a long-standing taskthat lies at the core of many computer vision applica-tions, such as tracking [ 17,18], reconstruction [ 22,23,39], pose estimation [ 19,27,50] and 3D representation learn- ing [ 15,16,46], etc. In a typical solution, geometry is first encoded into descriptors, and correspondences are then es-tablished between two frames by matching the most similar descriptors. As the two frames are observed from different views, depicting the same geometry under different trans- 1https://github.com/haoyu94/RoITrFigure 1. Feature Matching Recall (FMR) on 3DLoMatch [ 19] and Rotated 3DLoMatch. Distance to the diagonal represents the robustness against rotations. Among all the state-of-the-art ap- proaches, RoITr not only ranks first on both benchmarks but alsoshows the best robustness against the enlarged rotations. formations identically, i.e., the pose-invariance, becomes the key to success in the point cloud matching task. Since the side effects caused by a global translation can always be easily eliminated, e.g., by aligning the barycen- ter with the origin, the attention naturally shifts to coping with the rotations. In the past, handcrafted local descrip- tors [ 11,30,31,41] were designed to be rotation-invariant so that the same geometry observed from different viewscan be correctly matched. With the emergence of deep neu-ral models for 3D point analysis, e.g., multilayer percep- trons (MLPs)-based like PointNet [ 25,26], convolutions- based like KPConv [ 6,40], and the attention-based like PointTransformer [ 33,53], recent approaches [ 1,7,9,10,13, 19,20,27,32,48–51] propose to learn descriptors from raw points as an alternative to handcrafted features that are lessrobust to occlusion and noise. The majority of deep pointmatchers [ 7,10,19,20,27,34,48,50–52] is sensitive to ro- tations. Consequently, their invariance to rotations must beobtained extrinsically via augmented training to ensure thatthe same geometry under different poses can be depictedsimilarly. However, as the training cases can never span thecontinuous SO(3) space, they always suffer from instabil- ity when facing rotations that are rarely seen during train-ing. This can be observed by a significant performance dropunder enlarged rotations at inference time. (See Fig. 1.) There are other works [ 1,9,13,32,44] that only lever- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5384 age deep neural networks to encode the pure geometry with the intrinsically-designed rotation invariance. However, theintrinsic rotation invariance comes at the cost of losingglobal context. For example, a human’s left and right halvesare almost identically described, which naturally degradesthe distinctiveness of features. Most recently, RIGA [ 49] is proposed to enhance the distinctiveness of the rotation-invariant descriptors by incorporating a global context, e.g., the left and right halves of a human become distinguishableby knowing there is a chair on the left while a table on theright. However, it lacks a highly-representative geometryencoder since it relies on PointNet [ 25], which accounts for an ineffective local geometry description. Moreover, as de-picting the cross-frame spatial relationships is non-trivial,previous works [ 19,27,34,50] merely leverage the contex- tual features in the cross-frame context aggregation, whichneglects the positional information. Although RIGA pro- poses to learn a rotation-invariant position representation by leveraging an additional PointNet, this simple design is hardto model the complex cross-frame positional relationshipsand leads to less distinctive descriptors. In this paper, we present Rotation- Invariant Transformer (RoITr) to tackle the problem of point cloud matching un-der arbitrary pose variations. By using Point Pair Fea-tures (PPFs) as the local coordinates, we propose an at-tention mechanism to learn the pure geometry regardless of the varying poses. Upon it, attention-based layers are further proposed to compose the encoder-decoder architec-ture for highly-discriminative and rotation-invariant geom-etry encoding. We demonstrate its superiority over Point-Transformer [ 53], a state-of-the-art attention-based back- bone network, in terms of both efficiency and efficacy inFig. 8and Tab. 4(a), respectively. On the global level, the cross-frame position awareness is introduced in a rotation-invariant fashion to facilitate feature distinctiveness. We il-lustrate its significance over the state-of-the-art design [ 27] in Tab. 4(d). Our main contributions are summarized as: • An attention mechanism designed to disentangle the geometry and poses, which enables the pose-agnosticgeometry description. • An attention-based encoder-decoder architecture that learns highly-representative local geometry in arotation-invariant fashion. • A global transformer with rotation-invariant cross- frame position awareness that significantly enhancesthe feature distinctiveness.

Zhang_LVQAC_Lattice_Vector_Quantization_Coupled_With_Spatially_Adaptive_Companding_for_CVPR_2023

Abstract Recently, numerous end-to-end optimized image com- pression neural networks have been developed and proved themselves as leaders in rate-distortion performance. The main strength of these learnt compression methods is in powerful nonlinear analysis and synthesis transforms that can be facilitated by deep neural networks. However, out of operational expediency, most of these end-to-end methods adopt uniform scalar quantizers rather than vector quantiz- ers, which are information-theoretically optimal. In this pa- per, we present a novel Lattice Vector Quantization scheme coupled with a spatially Adaptive Companding (LVQAC) mapping. LVQ can better exploit the inter-feature depen- dencies than scalar uniform quantization while being com- putationally almost as simple as the latter. Moreover, to improve the adaptability of LVQ to source statistics, we couple a spatially adaptive companding (AC) mapping with LVQ. The resulting LVQAC design can be easily embedded into any end-to-end optimized image compression system. Extensive experiments demonstrate that for any end-to-end CNN image compression models, replacing uniform quan- tizer by LVQAC achieves better rate-distortion performance without significantly increasing the model complexity.

1. Introduction In the past five years, the research on end-to-end CNN image compression has made steady progress and led to the birth of a new class of image compression methods [1, 3, 4, 8, 21, 25, 26, 34–36, 40–44]. The CNN compres- sion can now match and even exceed the rate-distortion performance of the previous best image compression meth- ods [6,33,37,45], which operate in the traditional paradigm of linear transform, quantization and entropy coding. The advantages of the CNN approach of data compres- sion come from the nonlinearity of its analysis and syn- thesis transforms of the autoencoder architecture, the end- to-end joint optimization of the nonlinear transforms, uni-form quantization of the latent space and conditional en- tropy coding (context-based arithmetic coding) of quantized features. Apparently, using uniform scalar quantizer in the above CNN image compression framework is motivated by oper- ational expediency more than other considerations. Only at very high bit rate uniform quantization can approach the rate-distortion optimality [15]. It is very difficult to di- rectly adopt and optimize a vector quantizer (VQ) in the end-to-end CNN architecture design for data compression, because VQ is a discrete decision process and it is not com- patible with variational backpropagation that is necessary to the end-to-end CNN training. In [1], Agustsson et al. tried to circumvent the difficulty by a so-called soft-to-hard vec- tor quantization scheme. Their technique is a soft (continu- ous) relaxation of discrete computations of VQ and entropy so that their effects can be approximated in the end-to-end training. However, in [1] the quantization centers are opti- mized along with the other modules, which make the whole system quite cumbersome and more difficult to train. In this paper, we p ropose a novel Lattice Vector Quantiza- tion scheme coupled with a spatially Adaptive Companding (LVQAC) mapping. LVQ can better exploit the inter-feature dependencies than scalar uniform quantization while be- ing computationally almost as simple as the latter. Even if the features to be compressed are statistically independent, LVQ is still a more efficient coding strategy than scalar uni- form quantization. This is because the former offers a more efficient covering of high-dimensional space than the lat- ter, as proven by the theory of sphere packings, lattices and groups [11]. Moreover, to improve the adaptability of LVQ to source statistics, we couple a spatially adaptive compand- ing mapping with LVQ. The resulting LVQAC design is computationally as simple and as amenable to the end-to- end training of the CNN compression model as in the orig- inally proposed framework of [3]. Consequently, for any end-to-end CNN image com- pression models, replacing uniform quantizer by LVQAC achieves better rate-distortion performance without sig- nificantly increasing the model complexity; the simpler This CVPR paper is the Open Access version, provided

Zhang_MetaPortrait_Identity-Preserving_Talking_Head_Generation_With_Fast_Personalized_Adaptation_CVPR_2023

Abstract In this work, we propose an ID-preserving talking head generation framework, which advances previous methods intwo aspects. First, as opposed to interpolating from sparse flow, we claim that dense landmarks are crucial to achiev- ing accurate geometry-aware flow fields. Second, inspiredby face-swapping methods, we adaptively fuse the source identity during synthesis, so that the network better pre-serves the key characteristics of the image portrait. Al- though the proposed model surpasses prior generation fi-delity on established benchmarks, personalized fine-tuning is still needed to further make the talking head generation qualified for real usage. However , this process is rather computationally demanding that is unaffordable to stan-dard users. To alleviate this, we propose a fast adaptationmodel using a meta-learning approach. The learned modelcan be adapted to a high-quality personalized model as fast as 30 seconds. Last but not least, a spatial-temporal en- hancement module is proposed to improve the fine details while ensuring temporal coherency. Extensive experiments *Equal contribution, interns at Microsoft Research. †Joint corresponding authors.prove the significant superiority of our approach over the state of the arts in both one-shot and personalized settings.

1. Introduction Talking head generation [ 1,6,7,24,27,31,33,39,41,47] has found extensive applications in face-to-face live chat,virtual reality and virtual avatars in games and videos. Inthis paper, we aim to synthesize a realistic talking head with a single source image (one-shot) that provides the appear-ance of a given person while being animatable according to the motion of the driving person. Recently, considerable progress has been made with neural rendering techniques, bypassing the sophisticated 3D human modeling processand expensive driving sensors. While these works attain increasing fidelity and higher rendering resolution, identity preserving remains a challenging issue since the human vi-sion system is particularly sensitive to any nuanced devia-tion from the person’s facial geometry. Prior arts mainly focus on learning a geometry-aware warping field, either by interpolating from sparse 2D/3Dlandmarks or leveraging 3D face prior, e.g., 3D morphable This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22096 face model (3DMM) [ 2,3]. However, fine-grained facial geometry may not be well described by a set of sparse land-marks or inaccurate face reconstruction. Indeed, the warp-ing field, trained in a self-supervised manner rather than us-ing accurate flow ground truth, can only model coarse ge- ometry deformation, lacking the expressivity that captures the subtle semantic characteristics of the portrait. In this paper, we propose to better preserve the portrait identity in two ways. First, we claim that dense facial land- marks are sufficient for an accurate warping field predictionwithout the need for local affine transformation. Specifi-cally, we adopt a landmark prediction model [ 43] trained on synthetic data [ 42], yielding 669 head landmarks that of- fer significantly richer information on facial geometry. Inaddition, we build upon the face-swapping approach [ 23] and propose to enhance the perceptual identity by atten-tionally fusing the identity feature of the source portraitwhile retaining the pose and expression of the intermedi- ate warping. Equipped with these two improvements, our one-shot model demonstrates a significant advantage over prior works in terms of both image quality and perceptualidentity preservation when animating in-the-wild portraits. While our one-shot talking head model has achieved state-of-the-art quality, it is still infeasible to guarantee sat-isfactory synthesis results because such a one-shot setting isinherently ill-posed—one may never hallucinate the person- specific facial shape and occluded content from a singlephoto. Hence, ultimately we encounter the uncanny val- ley [32] that a user becomes uncomfortable as the synthe- sis results approach to realism. To circumvent this, one workaround is to finetune the model using several minutesof a personal video. Such personalized training has been widely adopted in industry to ensure product-level quality,yet this process is computationally expensive, which greatly limits its use scenarios. Thus, speeding up this personal- ized training , a task previously under-explored, is of great significance to the application of talking head synthesis. We propose to achieve fast personalization with meta- learning. The key idea is to find an initialization modelthat can be easily adapted to a given identity with limited training iterations. To this end, we resort to a meta-learning approach [ 9,26] that finds success in quickly learning dis- criminative tasks, yet is rarely explored in generative tasks. Specifically, we optimize the model for specific personal data with a few iterations. In this way, we get a slightlyfine-tuned personal model towards which we move the ini-tialization model weight a little bit. Such meta-learned ini- tialization allows us to train a personal model within 30 sec-onds, which is 3 times faster than a vanilla pretrained modelwhile requiring less amount of personal data. Moreover, we propose a novel temporal super-resolution network to enhance the resolution of the generated talkinghead video. To do this, we leverage the generative priorto boost the high-frequency details for portraits and mean- while take into account adjacent frames that are helpful toreduce temporal flickering. Finally, we reach temporallycoherent video results of 512×512 resolution with com- pelling facial details. In summary, this work innovates inthe following aspects: • We propose a carefully designed framework to signif- icantly improve the identity-preserving capability whenanimating a one-shot in-the-wild portrait. • To the best of our knowledge, we are the first to explore meta-learning to accelerate personalized training, thusobtaining ultra-high-quality results at affordable cost. • Our novel video super-resolution model effectively en- hances details without introducing temporal flickering.

Yeh_Meta-Personalizing_Vision-Language_Models_To_Find_Named_Instances_in_Video_CVPR_2023

Abstract Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they cur- rently struggle with personalized searches for moments in a video where a specific object instance such as “My dog Biscuit” appears. We present the following three contribu- tions to address this problem. First, we describe a method tometa-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM’s token vocabulary by learn- ing novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM’s embedding space. Finally, we intro- duce This-Is-My, a personal video instance retrieval bench- mark. We evaluate our approach on This-Is-My and Deep- Fashion2 and show that we obtain a 15% relative improve- ment over the state of the art on the latter dataset.

1. Introduction The recent introduction of large-scale pre-trained vision- language models (VLMs) has enabled many new vision tasks, including zero-shot classification and retrieval [14, 18, 22], image/video generation [12, 24, 25, 27, 28, 32], or language-guided question answering [1, 19, 42]. It is now possible to search not only for specific object categories (e.g., dogs) but also for more specific descriptions of both †Equal advising. ∗Work done during CHY’s summer internship at Adobe Research. 2Czech Institute of Informatics, Robotics and Cybernetics at the Czech Technical University in Prague. Figure 1. Meta-Personalized Vision-Language Model (VLM) to Retrieve Named Instances in Video. Given a video where a user-specific instance, e.g., “My dog Biscuit” is mentioned, our method automatically learns a representation for the user-specific instance in the VLM’s text input space. The personalized VLM can then be used to retrieve the learned instance in other contexts through natural language queries, e.g.,<my dog Biscuit > grabbing a pink frisbee . This result is enabled by meta- personalizing the VLM on a large-scale dataset of narrated videos by pre-learning shared global category tokens (in this example for the category of ’dogs’), which are then easily personalized to user- specific instances from only a few user-given training examples. the object and scene attributes ( e.g., “A small white dog playing at the dog park”). However, we often do not want to search for just any example of a generic category but in- stead to find a specific instance. For example, a user might want to search their personal video library for all the scenes that show their dog “Biscuit grabbing a pink frisbee”, as illustrated in Figure 1. Since VLMs do not have a represen- tation of “Biscuit,” such queries are beyond the capabilities This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19123 of off-the-shelf VLMs. Recent work [5] proposed a method to extend the lan- guage encoder’s vocabulary with a newly learned token that represents a specific personal instance to address this is- sue. While this approach enables language-guided search for personal instances by placing the learned tokens in the query prompt, their solution assumes a collection of man- ually annotated images showing the individual instance in various contexts for successful token learning. For this ap- proach to work in practice, a user must manually annotate all their important personal instances in various contexts, such that the instance representation does not capture nui- sance features, e.g., the background. We thus identify two key challenges: 1) collecting personal instance examples without explicit human labeling and 2) learning a gener- alizable object-centric representation of personal instances from very few examples. The contributions of this work are three-fold. As our first contribution, we propose a method to automatically identify important personal instances in videos for personalizing a vision-language model without explicit human annotations. Indeed, people often record and refer to personal items or relationships in videos found online. Our approach is thus to identify mentions of personal instances in a video au- tomatically and leverage these moments to build a set of personal instances for training. To this end, we extract the transcripts of videos using speech-to-text models and find candidate moments by looking for occurrences of “this is my *” or similar possessive adjective patterns. The symbol * in this example could represent a single word or sequence of words describing the instance ( e.g., *= “dog Biscuit”). We then use vision-language similarity to filter non-visual examples and to find additional occurrences in the video for training For example, we found more than six thousand named instances in 50K videos randomly sampled from the Merlot Reserve dataset [44]. We call the resulting collection of named instances in videos the This-Is-My dataset. As our second contribution, we propose a novel model and training procedure to learn text tokens representing the named instances in video from possibly very few and noisy training examples. Our method represents each in- stance with learned tokens and models each token as a linear combination of a set of pre-learned category-specific fea- tures shared across different instances. This set of shared category-specific features (similar to object attributes) im- proves the generalization of our method by preventing the instance representations from capturing nuisance features (e.g., the scene background). Furthermore, we show how to pre-train and adapt the shared category features using a large set of automatically collected This-Is-My examples, further improving our model’s few-shot personalization per- formance at test-time. We call this pre-training of shared category features meta-personalization . In contrast to priorwork [5], our method does not require training additional neural network models and requires only the optimization of a contrastive learning objective. As our final contribution, we demonstrate and evalu- ate our model on an existing fashion item retrieval bench- mark, DeepFashion2, and our new challenging This-Is-My video instance retrieval dataset4depicting specific object instances across different videos and contexts. Our ex- periments demonstrate that our method outperforms sev- eral baselines and prior approaches on these challenging language-guided instance retrieval tasks.

Ye_Affordance_Diffusion_Synthesizing_Hand-Object_Interactions_CVPR_2023

Abstract Recent successes in image synthesis are powered by large-scale diffusion models. However, most methods are currently limited to either text- or image-conditioned gen- eration for synthesizing an entire image, texture transfer or inserting objects into a user-specified region. In contrast, in this work we focus on synthesizing complex interactions (i.e., an articulated hand) with a given object. Given an RGB image of an object, we aim to hallucinate plausible images of a human hand interacting with it. We propose a two-step generative approach: a LayoutNet that samples an articulation-agnostic hand-object-interaction layout, and a ContentNet that synthesizes images of a hand grasping the object given the predicted layout. Both are built on top of a large-scale pretrained diffusion model to make use of its latent representation. Compared to baselines, the proposed method is shown to generalize better to novel objects and perform surprisingly well on out-of-distribution in-the-wild scenes of portable-sized objects. The resulting system al- lows us to predict descriptive affordance information, such as hand articulation and approaching orientation. *Yufei was an intern at NVIDIA during the project.

1. Introduction Consider the bottles, bowls and cups shown in the left column of Figure 1. How might a human hand interact with such objects? Not only is it easy to imagine, from a single image, the types of interactions that might occur (e.g., ‘grab/hold’), and the interaction locations that might happen ( e.g. ‘handle/body’), but it is also quite natural to hallucinate—in vivid detail— several ways in which a hand might contact and use the objects. This ability to predict and hallucinate hand-object-interactions (HOI) is critical to functional understanding of a scene, as well as to visual im- itation and manipulation. Can current computer vision algorithms do the same? On the one hand, there has been a lot of progress in image gen- eration, such as synthesizing realistic high-resolution im- ages spanning a wide range of object categories [43, 73] from human faces to ImageNet classes. Newer diffusion models such as Dall-E 2 [65] and Stable Diffusion [66] can generate remarkably novel images in diverse styles. In fact, highly-realistic HOI images can be synthesized from simple text inputs such as “a hand holding a cup” [65, 66]. On the other hand, however, such models fail when con- ditioned on an image of a particular object instance. Given This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22479 an image of an object, it remains an extremely challeng- ing problem to generate realistic human object interaction. Solving this problem requires (at least implicitly) an under- standing of physical constraints such as collision and force stability, as well as modeling the semantics and functional- ity of objects — the underlying affordances [19]. For ex- ample, the hand should prefer to grab the kettle handle but avoid grabbing the knife blade. Furthermore, in order to produce visually plausible results, it also requires modeling occlusions between hands and objects, their scale, lighting, texture, etc. In this work, we propose a method for interaction syn- thesis that addresses these issues using diffusion models. In contrast to a generic image-conditioned diffusion model, we build upon the classic idea of disentangling where to inter- act (layout ) from how to interact ( content ) [25,30]. Our key insight is that diverse interactions largely arise from hand- object layout, whereas hand articulations are driven by lo- cal object geometry. For example, a mug can be grasped by either its handle or body, but once the grasping location is determined, the placement of the fingers depends on the object’s local surface and the articulation will exhibit only subtle differences. We operationalize this idea by proposing a two-step stochastic procedure: 1) a LayoutNet that gener- ates 2D spatial arrangements of hands and objects, and 2) aContentNet that is conditioned on the query object im- age and the sampled HOI layout to synthesize the images of hand-object interactions. These two modules are both implemented as image-conditioned diffusion models. We evaluate our method on HOI4D and EPIC- KITCHEN [11, 48]. Our method outperforms generic im- age generation baselines, and the extracted hand poses from our HOI synthesis are favored in user studies against base- lines that are trained to directly predict hand poses. We also demonstrate surprisingly robust generalization ability across datasets, and we show that our model can quickly adapt to new hand-object-interactions with only a few ex- amples. Lastly, we show that our proposed method enables editing and guided generation from partially specified lay- out parameters. This allows us to reuse heatmap prediction from prior work [13, 56] and to generate consistent hand sizes for different objects in one scene. Our main contributions are summarized below: 1) we propose a two-step method to synthesize hand-object in- teractions from an object image, which allows affordance information extracted from it; 2) we use inpainting tech- inuqes to supervise the model with paired real-world HOI and object-only images and propose a novel data augmenta- tion method to alleviate overfit to artifacts; and 3) we show that our approach generates realistic HOI images along with plausible 3D poses and generalizes surprisingly well on out- of-distribution scenes. 4) We also highlight several applica- tions that would benefit from such a method.

Yin_1_VS_100_Parameter-Efficient_Low_Rank_Adapter_for_Dense_Predictions_CVPR_2023

Abstract Fine-tuning large-scale pre-trained vision models to downstream tasks is a standard technique for achieving state-of-the-art performance on computer vision bench- marks. However, fine-tuning the whole model with millions of parameters is inefficient as it requires storing a same- sized new model copy for each task. In this work, we pro- pose LoRand, a method for fine-tuning large-scale vision models with a better trade-off between task performance and the number of trainable parameters. LoRand gener- ates tiny adapter structures with low-rank synthesis while keeping the original backbone parameters fixed, resulting in high parameter sharing. To demonstrate LoRand’s ef- fectiveness, we implement extensive experiments on object detection, semantic segmentation, and instance segmenta- tion tasks. By only training a small percentage (1% to 3%) of the pre-trained backbone parameters, LoRand achieves comparable performance to standard fine-tuning on COCO and ADE20K and outperforms fine-tuning in low-resource PASCAL VOC dataset.

1. Introduction With the rapid development of computer vision, pa- rameters in deep models are surging. Giant models need to be trained with massive resources to achieve superior performance [3, 17, 47, 58], which is often unavailable to many academics and institutions. “Pretrain & Finetun- ing” paradigm is widely used to alleviate this dilemma. Teams with sufficient computation resources utilise enor- mous datasets [2, 9, 40, 50] to train superior backbones [4, 32, 40, 48] and optimise the models with ideal perfor- mances. Models pretrained in this way usually have a su- Corresponding author. yEqual contribution. 46474849505152 0 20 40 60 80mAP Trainable Backbone Params (M)LoRand(Swin -B) 2.5M LoRand(Swin -S) 1.8M LoRand(Swin -T) 0.9MSwin -B 88MSwin -S 50M Swin -T 29M DeiT -S 22M ResNet -50 20MResNeXt -101-64 80MResNeXt -101-32 40M ResNet -101 44MFigure 1. Comparisons of trainable backbone parameters between our methods (red) and fine-tuning (black). In COCO, we achieve advanced performances and outperform most existing backbones with only 0.92.5M new backbone parameters (Cascade-RCNN is employed as the detector). The fine-tuning paradigm produces massive redundant backbone parameters, whereas our approach saves over 97% of hardware resources with competitive perfor- mances. The sizes of the circles intuitively compare the number of trainable parameters. perior understanding of homogeneous data. After that, re- searchers with limited computational resources can trans- fer the understanding capabilities of the pre-trained models to downstream tasks with promising performances by fine- tuning [1, 26, 46, 53]. However, the fine-tuned model will produce a new set of parameters as large as the pre-trained model. New pa- rameters are independent of the pre-trained models and un- shareable, which are very hardware intensive for cloud ser- vice providers [23, 49]. Figure 1 compares the parameter quantities of some remarkable backbones and their perfor- mances on the COCO [28] dataset. Recent advances in natural language processing (NLP) [30, 38] show that large pre-trained models trained with rich data have strong gener- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20116 Layer Norm W/SW - MSA LoRand Module Layer Norm 2x Feed - Forward Layer LoRand Module ` Multi - branch low rank project Nonlinearity Fusion Multi - branch low rank project Fusion ` Swin - Transformer Block LoRand LayerFigure 2. Architecture of the adapter module and its integration with the Transformer. Left: We add two LoRand structures to each SwinBlock located behind the W/SW-MSA and MLP struc- tures respectively. Right: LoRand contains two Multi-branch low- rank projections and nonlinearity. We include skip-connection to LoRand to enhance its robustness. alisability, which means most parameters in the pre-trained models can be shared with the new tasks [22,36,37,44,59]. Moreover, recent literature demonstrates that the feature un- derstanding of pre-trained models could be reduced when they are fine-tuned in low-resource situations [12, 36]. To tackle these issues, NLP researchers propose two new train- ing paradigms based on pre-trained models: Adapter Tun- ing [22] and Prompt Tuning [30], both of which tune the new models by fixing the pre-trained parameters and adding a few trainable structures (less than 10% of the backbone). These paradigms create a new buzz in NLP and achieve im- pressive performances which can be competitive with fine- tuning [12, 22, 30, 36–38, 44, 59]. Advances in NLP also shed new light on computer vision. Jia et al. [24] propose Visual Prompt Tuning (VPT) and demonstrate that VPT can outperform fine-tuning on image classification tasks by training a small number of trainable parameters. Never- theless, VPT shows weakness on more challenging dense predictions like semantic segmentation compared with fine- tuning [24]. To find a parameter-efficient paradigm with promising performance in computer vision, we explore the poten- tial of Adapter Tuning for visual dense predictions. We employ the advanced Swin Transformer [32] trained with ImageNet-22K [9] as the pre-trained model. After that, we add bottleneck adapter structures [22] behind each Swin- Block and freeze the original backbone parameters when training, but this approach cannot achieve comparable per- formance to fine-tuning as mentioned in [24]. In the exper-iments, we find that the models perform better with sparser adapter structures. To improve the performance of Adapter Tuning, we propose Low-Rank Adapter (LoRand) to re- duce the adapter parameters, as shown in Figure 2. LoRand sparsely parameterizes the matrices in adapters by low-rank synthesis. Specifically, the projection matrix of the fully- connected layer (FC) in LoRand is a product of multiple low-rank matrices, which reduces FC parameters by more than 80%. We implement extensive experiments on ob- ject detection (PASCAL VOC [14]), semantic segmentation (ADE20K [62]), and instance segmentation (MS COCO [28]) to verify the capability of LoRand. Experimental re- sults show that LoRand-Tuning is comparable to fine-tuning on multiple tasks with only 1.8% to 2.8% new backbone parameters, which suggests that the pre-trained backbone parameters can be fully shared. More interestingly, our method completely outperforms fine-tuning on the PAS- CAL VOC dataset, illustrating that LoRand-Tuning can re- duce the impairment of fine-tuning on pre-trained models in low-resource configurations. Our method demonstrates that the LoRand-Tuning paradigm can substantially save storage resources and achieve competitive performances on most dense prediction tasks. In summary, our contributions are three-fold: We demonstrate that visual pre-trained models are highly generalisable and shareable. With our training methods, new tasks require only a few trainable pa- rameters to achieve performances comparable to fine- tuning, which can save massive hardware resources. We propose the LoRand structure for sparser adapters based on low-rank synthesis. We demonstrate that the backbone parameters in fine-tuning are highly redun- dant, which can be replaced by 1.8% to 2.8% addi- tional parameters in LoRand. Extensive experiments on object detection, semantic segmentation, and instance segmentation show that LoRand-Tuning can achieve remarkable performances and reduce massive new parameters in challenging dense prediction tasks.

Yu_ACR_Attention_Collaboration-Based_Regressor_for_Arbitrary_Two-Hand_Reconstruction_CVPR_2023

Abstract Reconstructing two hands from monocular RGB images is challenging due to frequent occlusion and mutual con- fusion. Existing methods mainly learn an entangled rep- resentation to encode two interacting hands, which are in- credibly fragile to impaired interaction, such as truncated hands, separate hands, or external occlusion. This pa- per presents ACR (Attention Collaboration-based Regres- sor), which makes the first attempt to reconstruct hands in arbitrary scenarios. To achieve this, ACR explicitly miti- gates interdependencies between hands and between parts by leveraging center and part-based attention for feature extraction. However, reducing interdependence helps re- lease the input constraint while weakening the mutual rea- soning about reconstructing the interacting hands. Thus, based on center attention, ACR also learns cross-hand prior that handle the interacting hands better. We eval- uate our method on various types of hand reconstruction datasets. Our method significantly outperforms the best interacting-hand approaches on the InterHand2.6M dataset while yielding comparable performance with the state-of- the-art single-hand methods on the FreiHand dataset. More qualitative results on in-the-wild and hand-object interac- tion datasets and web images/videos further demonstrate the effectiveness of our approach for arbitrary hand recon- struction. Our code is available at this link1.

1. Introduction 3D hand pose and shape reconstruction based on a single RGB camera plays an essential role in various emerging ap- plications, such as augmented and virtual reality (AR/VR), human-computer interaction, 3D character animation for movies and games, etc. However, this task is highly chal- lenging due to limited labeled data, occlusion, depth am- biguity, etc. Earlier attempts [ 1,2,36,39] level down the problem difficulty and focus on single-hand reconstruction. These methods started from exploring weakly-supervised *Corresponding author. 1https://github.com/ZhengdiYu/Arbitrary-Hands-3D-Reconstruction Figure 1. Given a monocular RGB image, our method makes the first attempt to reconstruct hands under arbitrary scenarios by representation disentanglement and interaction mutual reasoning while the previous state-of-the-art method IntagHand [ 15] failed. learning paradigms [ 2] to designing more advanced network models [ 31]. Although single-hand approaches can be ex- tended to reconstruct two hands, they generally ignore the inter-occlusion and confusion issues, thus failing to handle two interacting hands. To this end, recent research has shifted toward recon- structing two interacting hands. Wang et al. [ 33] extract multi-source complementary information to reconstruct two interacting hands simultaneously. Rong et al. [ 27] and Zhang et al. [ 35] first obtain initial prediction and stack in- termediate results together to refine two-hand reconstruc- tion. The latest work [ 15] gathers pyramid features and two-hand features as input for a GCN-based network that regresses two interacting hands unitedly. These methods share the same principle: treating two hands as an integral This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12955 and learning a unified feature to ultimately refine or regress the interacting-hand model. The strategy delivers the ad- vantage of explicitly capturing the hands’ correlation but in- evitably introduces the input constraint of two hands. This limitation also makes the methods particularly vulnerable and easily fail to handle inputs containing imperfect hand interactions, including truncation or external occlusions. This paper takes the first step toward reconstructing two hands in arbitrary scenarios. Our first key insight is lever- aging center and part attention to mitigate interdependen- cies between hands and between parts to release the input constraint and eliminate the prediction sensitivity to a small occluded or truncated part. To this end, we propose At- tention Collaboration-based Regressor (ACR). Specifically, it comprises two essential ingredients: Attention Encoder (AE) and Attention Collaboration-based Feature Aggrega- tor (ACFA). The former learns hand-center and per-part at- tention maps with a cross-hand prior map, allowing the net- work to know the visibility of both hands and each part be- fore the hand regression. The latter exploits the hand-center and per-part attention to extract global and local features as a collaborative representation for regressing each hand inde- pendently and subsequently enhance the interaction model- ing by cross-hand prior reasoning with an interaction field. In contrast to the existing method, our method provides more advantages, such as hand detector free. Furthermore, experiments show that ACR achieves lower error on the In- terHand2.6M dataset than the state-of-the-art interacting- hand methods, demonstrating its effectiveness in handling interaction challenges. Finally, results on in-the-wild im- ages or video demos indicate that our approach is promis- ing for real-world application with the powerful aggregated representation for arbitrary hands reconstruction. Our key contributions are summarized as: (1)we take the first step toward reconstructing two hands at arbi- trary scenarios .(2)We propose to leverage both cen- ter and part based representation to mitigate interde- pendencies between hands and between parts and re- lease the input constraint. (3)In terms of modeling for in- teracting hands, we propose a cross-hand prior reason- ing module with an interaction field to adjust the de- pendency strength. (4)Our method outperforms exist- ing state-of-the-art approaches significantly on the In- terHand2.6M benchmark . Furthermore, ACR is the most practical method for various in-the-wild application scenes among all the prior arts of hand reconstruction.

Zhang_Document_Image_Shadow_Removal_Guided_by_Color-Aware_Background_CVPR_2023

Abstract Existing works on document image shadow removal mostly depend on learning and leveraging a constant back- ground (the color of the paper) from the image. However, the constant background is less representative and frequent- ly ignores other background colors, such as the printed col- ors, resulting in distorted results. In this paper, we present a color-aware background extraction network (CBENet) for extracting a spatially varying background image that ac- curately depicts the background colors of the documen- t. Furthermore, we propose a background-guided docu- ment images shadow removal network (BGShadowNet) us- ing the predicted spatially varying background as auxil- iary information, which consists of two stages. At Stage I, a background-constrained decoder is designed to pro- mote a coarse result. Then, the coarse result is refined with a background-based attention module (BAModule) to maintain a consistent appearance and a detail improvement module (DEModule) to enhance the texture details at Stage II. Experiments on two benchmark datasets qualitatively and quantitatively validate the superiority of the proposed approach over state-of-the-arts.

1. Introduction Documents, such as textbooks, newspapers, leaflets, and receipts, are available daily, often saved as electronic doc- uments for digital document archives or online message transfer. Since the wide use and convenience of mobile phones, people currently prefer to use mobile phones for digital document copying. However, the captured docu- ment images become highly susceptible to shadows when the light sources are blocked. The low brightness in shadow regions reduces the quality and readability of the documen- Corresponding author: Chunxia Xiao ([email protected]). (a) Shadow image (b) Our background (c) Our result (d) Result of [2] (e) Background of [24] (f) Result of [24] Figure 1. Document image shadow removal. With the assumption of constant background, results of [2] and [24] may cause color distortion or shadow remnant. By using a spatially varying back- ground, our method can produce more desirable result. t image, resulting in illegible content and unpleasant user experience [2, 10, 12, 19, 28, 29]. Thus, shadow removal for document images is a required image processing task in var- ious vision applications [3, 4, 33, 37, 46, 48]. Although natural image shadow removal has made sub- stantial progress [16, 20, 22, 40, 44, 45], these approach- es generally perform poorly on document pictures due to their drastically different features from natural images. Natural image, for example, emphasizes background con- tent (shadow-free image) without a shadow layer [6, 7, 13, 27, 41], whereas document image emphasizes text content [2, 19, 30]. Without considering the particular properties of the document image, traditional approaches to natural im- age generally yield incorrect result when they are applied to document image, as well as learning-based methods (see Figure 9(c-f)) due to the less attention to the content struc- tures. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1818 Several approaches on document image shadow removal are currently available, which dig into the specific charac- teristics of the document image [3, 21, 30, 49]. However, these approaches may cause color distortion or shadow rem- nant for image with complex backgrounds, as illustrated in Figure 1(d). Recently, Lin et al. [24] estimate a constant background for the image and propose BEDSR-Net for doc- ument image shadow removal. The constant background is the color of the paper (see Figure 1(e)), which ignores some other background colors, such as the printed colors. The constant background may provide inaccurate information for the shadow removal task, resulting in unsatisfactory re- sults (see Figure 1(f)). To address this problem, we propose a color-aware background extraction network (CBENet) to extract a spatially varying background, which can preserve various background colors of the original image (see Figure 1(b)). The spatially varying background can provide more useful color information, which contributes to image shad- ow removal, as shown in Figure 1(c). With the background image, we propose a background- guided shadow removal network (BGShadowNet) for doc- ument image that exploits the background image as aux- iliary information. Figure 2 presents the framework of the proposed BGShadowNet, which removes shadows in a two-stage process. First, we introduce a background- constrained decoder to combine background features with image features, which can help to promote the realistic coarse shadow-removal result. Then, we refine the coarse result with a background-based attention module (BAMod- ule) and a detail enhancement module (DEModule). In par- ticular, BAModule is designed to eliminate the illumination and color inconsistency in the image by using the attention mechanism. Inspired by image histogram equalization, DE- Module aims to enhance the detail features at low-level s- cales. Due to the lack of large-scale real document image datasets, we construct a new dataset comprised of real doc- ument images to facilitate the performance of document image shadow removal. Experiments on extensive docu- ment images and evaluations on two datasets verify that our BGShadowNet outperforms existing approaches. In summary, our main contributions are three-fold: We present a color-aware background extraction net- work (CBENet) for estimating a spatially varying background image that guides the shadow removal of document image. We propose a framework named BGShadowNet for document image shadow removal, which takes full ad- vantage of the background image and is able to robust- ly produce high-quality shadow removal results. We design a background-based attention module (BAModule) to maintain a consistent appearance and a detail enhancement module (DEModule) to enhance texture details.

Yoon_Cross-Guided_Optimization_of_Radiance_Fields_With_Multi-View_Image_Super-Resolution_for_CVPR_2023

Abstract Novel View Synthesis (NVS) aims at synthesizing an im- age from an arbitrary viewpoint using multi-view images and camera poses. Among the methods for NVS, Neural Radiance Fields (NeRF) is capable of NVS for an arbitrary resolution as it learns a continuous volumetric representa- tion. However, radiance fields rely heavily on the spectral characteristics of coordinate-based networks. Thus, there is a limit to improving the performance of high-resolution novel view synthesis (HRNVS). To solve this problem, we propose a novel framework using cross-guided optimiza- tion of the single-image super-resolution (SISR) and radi- ance fields. We perform multi-view image super-resolution (MVSR) on train-view images during the radiance fields op- timization process. It derives the updated SR result by fus- ing the feature map obtained from SISR and voxel-based un- certainty fields generated by integrated errors of train-view images. By repeating the updates during radiance fields op- timization, train-view images for radiance fields optimiza- tion have multi-view consistency and high-frequency de- tails simultaneously, ultimately improving the performance of HRNVS. Experiments of HRNVS and MVSR on various benchmark datasets show that the proposed method signifi- cantly surpasses existing methods.

1. Introduction Novel View Synthesis (NVS) is an approach to synthe- sizing an image from an arbitrary viewpoint using multi- view images and camera poses. This is an essential task in computer vision and graphics, and it can be actively used in street-view navigation, AR/VR, and robotics. Recently, Neural Radiance Fields [28] (NeRF) significantly improved the performance of NVS by learning multi-layer perceptron (MLP) from 5d coordinate input. Since then, many studies have been conducted to shorten the long learning time of NeRF [4, 10, 29, 36, 42, 43, 48], increase the performance of NVS using depth priors [5,8,32,41], and enable NVS from few-shot views [13, 16, 31, 49]. Figure 1. Cross-guided optimization between single image super- resolution and radiance fields. They complement weaknesses of one another with their respective strengths by using the SR update module, rendered train-view RGBs, and uncertainty maps. Continuous scene representations such as NeRF [28] can be rendered at arbitrary resolution. Thus, there are many studies to improve the performance of multi-scale scene representation. Mip-NeRF [2] proposes scale-dependent positional encoding, which makes a network be trained on multiple scales. In addition, BACON [22] proposes a net- work capable of band-limited multi-scale decomposition by giving a constraint to the bandwidth of network outputs. Both papers showed significant down-scaling performance on volume rendering. On the other hand, NeRF-SR [39] improves the performance of high-resolution novel view synthesis (HRNVS) by learning in an unsupervised manner through super-sampling in the radiance fields optimization process. Radiance fields have the ability to find scene geometry and optimize 5D functions simultaneously. Still, radiance fields have a low ability to perform super-resolution, and even if they synthesize high-resolution (HR) images, they This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12428 only depend on the characteristic of continuous scene repre- sentation. On the other hand, single-image super-resolution (SISR) generally specializes in learning the inverse func- tion of image degradation. Therefore, SISR could be benefi- cial to HRNVS by super-resolving train-view images; how- ever, SISR is an ill-posed problem for which multiple so- lutions exist, and multi-view consistency cannot be main- tained when multi-view images are processed separately. To solve this problem, we propose a novel framework us- ing cross-guided optimization between radiance fields and SISR. As shown in Fig. 1, our framework aims to ensure that radiance fields are guided by superior high-frequency details from SISR, and conversely, SISR is guided by multi- view consistency from radiance fields. We perform train- view synthesis during the radiance fields optimization pro- cess. Then, we generate voxel-based uncertainty fields to obtain uncertainty maps to find reliable regions in rendered train-view RGB images. The rendered train-view outputs and feature maps from the SISR network make it possible to do multi-view image super-resolution (MVSR) through the SR update module (SUM). Then, we continue optimizing the radiance fields using the updated SR outputs. Repeating the update process makes train-view images for radiance fields optimization have multi-view consistency and high- frequency details simultaneously, ultimately improving the performance of HRNVS. Our method shows that the performance of HRNVS and MVSR on various benchmark datasets significantly sur- passes existing methods. It also shows consistent perfor- mance improvements for various SISR models and radiance fields models in our method. In summary, our contributions are as follows: • We propose a novel framework for performing cross- guided radiance fields optimization using the SISR model for HRNVS. • We propose voxel-based uncertainty fields to find reliable regions of synthesized images. • We propose an SR update module (SUM) using voxel- based uncertainty fields and train-view synthesis outputs for MVSR. • Experiments on various benchmark datasets show that the proposed method significantly surpasses existing meth- ods in terms of performance for HRNVS and MVSR.

Yang_Efficient_On-Device_Training_via_Gradient_Filtering_CVPR_2023

Abstract Despite its importance for federated learning, continu- ous learning and many other applications, on-device train- ing remains an open problem for EdgeAI. The problem stems from the large number of operations (e.g., floating point multiplications and additions) and memory consump- tion required during training by the back-propagation al- gorithm. Consequently, in this paper, we propose a new gradient filtering approach which enables on-device CNN model training. More precisely, our approach creates a special structure with fewer unique elements in the gradi- ent map, thus significantly reducing the computational com- plexity and memory consumption of back propagation dur- ing training. Extensive experiments on image classification and semantic segmentation with multiple CNN models (e.g., MobileNet, DeepLabV3, UPerNet) and devices (e.g., Rasp- berry Pi and Jetson Nano) demonstrate the effectiveness and wide applicability of our approach. For example, com- pared to SOTA, we achieve up to 19 ×speedup and 77.1% memory savings on ImageNet classification with only 0.1% accuracy loss. Finally, our method is easy to implement and deploy; over 20 ×speedup and 90% energy savings have been observed compared to highly optimized baselines in MKLDNN and CUDNN on NVIDIA Jetson Nano. Conse- quently, our approach opens up a new direction of research with a huge potential for on-device training.1

1. Introduction Existing approaches for on-device training are neither efficient nor practical enough to satisfy the resource con- straints of edge devices (Figure 1). This is because these methods do not properly address a fundamental problem in on-device training, namely the computational and memory complexity of the back-propagation (BP) algorithm. More precisely, although the architecture modification [6] and layer freezing [18, 20] can help skipping the BP for some layers, for other layers, the complexity remains high. Gra- 1Code: https://github.com/SLDGroup/GradientFilter-CVPR23dient quantization [4, 7] can reduce the cost of arithmetic operations but cannot reduce the number of operations ( e.g., multiplications); thus, the speedup in training remains lim- ited. Moreover, gradient quantization is not supported by existing deep-learning frameworks (e.g., CUDNN [9], MKLDNN [1], PyTorch [25] and Tensorflow [2]). To en- able on-device training, there are two important questions must be addressed: •How can we reduce the computational complexity of back propagation through the convolution layers? •How can we reduce the data required by the gradient computation during back propagation? In this paper, we propose gradient filtering , a new research direction, to address both questions. By addressing the first question, we reduce the computational complexity of train- ing; by addressing the second question, we reduce the mem- ory consumption. In general, the gradient propagation through a convolu- tion layer involves multiplying the gradient of the output variable with a Jacobian matrix constructed with data from either the input feature map or the convolution kernel. We aim at simplifying this process with the new gradient filter- ing approach proposed in Section 3. Intuitively, if the gradi- ent map w.r.t. the output has the same value for all entries, then the computation-intensive matrix multiplication can be greatly simplified, and the data required to construct the Ja- cobian matrix can be significantly reduced. Thus, our gra- dient filtering can approximate the gradient w.r.t. the output by creating a new gradient map with a special ( i.e., spatial) structure and fewer unique elements. By doing so, the gra- dient propagation through the convolution layers reduces to cheaper operations, while the data required (hence memory) for the forward propagation also lessens. Through this fil- tering process, we trade off the gradient precision against the computation complexity during BP. We note that gradi- ent filtering does not necessarily lead to a worse precision, i.e., models sometimes perform better with filtered gradi- ents when compared against models trained with vanilla BP. In summary, our contributions are as follows: This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3811 • We propose gradient filtering , which reduces the com- putation and memory required for BP by more than two orders of magnitude compared to the exact gradi- ent calculation. • We provide a rigorous error analysis which shows that the errors introduced by the gradient filtering have only a limited influence on model accuracy. • Our experiments with multiple CNN models and com- puter vision tasks show that we can train a neural net- work with significantly less computation and memory costs, with only a marginal accuracy loss compared to baseline methods. Side-by-side comparisons against other training acceleration techniques also suggest the effectiveness of our method. • Our method is easy to deploy with highly optimized deep learning frameworks ( e.g., MKLDNN [1] and CUDNN [9]). Evaluations on resource-constrained edge (Raspberry Pi and Jetson Nano) and high- performance devices (CPU/GPU) show that our method is highly suitable for real life deployment. The paper is organized as follows. Section 2 reviews rel- evant work. Section 3 presents our method in detail. Sec- tion 4 discusses error analysis, computation and memory consumption. Experimental results are presented in Section 5. Finally, Section 6 summarizes our main contributions.

You_UTM_A_Unified_Multiple_Object_Tracking_Model_With_Identity-Aware_Feature_CVPR_2023

Abstract Recently, Multiple Object Tracking has achieved great success, which consists of object detection, feature embed- ding, and identity association. Existing methods apply the three-step or two-step paradigm to generate robust trajec- tories, where identity association is independent of other components. However, the independent identity associa- tion results in the identity-aware knowledge contained in the tracklet not be used to boost the detection and embed- ding modules. To overcome the limitations of existing meth- ods, we introduce a novel Unified Tracking Model (UTM) to bridge those three components for generating a positive feedback loop with mutual benefits. The key insight of UTM is the Identity-Aware Feature Enhancement (IAFE), which is applied to bridge and benefit these three components by utilizing the identity-aware knowledge to boost detec- tion and embedding. Formally, IAFE contains the Identity- Aware Boosting Attention (IABA) and the Identity-Aware Erasing Attention (IAEA), where IABA enhances the consis- tent regions between the current frame feature and identity- aware knowledge, and IAEA suppresses the distracted re- gions in the current frame feature. With better detections and embeddings, higher-quality tracklets can also be gener- ated. Extensive experiments of public and private detections on three benchmarks demonstrate the robustness of UTM.

1. Introduction Multiple Object Tracking (MOT) aims at locating and identifying all of the moving objects in the video, which has broad application prospects in visual surveillance, human- computer interaction, virtual reality, and unmanned vehi- cles. With the rapid development of object detection [12, 34, 35, 66], Tracking-By-Detection has became a favorite *indicates corresponding author: Changsheng Xu. (a) Separate Detection and Embedding (SDE) Detector DetectionEmbedding model Embedding TrackletIdentity Association (b) Joint Detection and Embedding (JDE) BackboneTracklet Identity Association Detection Embedding (c) Unified Tracking Model (UTM) BackboneIAFE Detection Embedding TrackletFigure 1. Comparison of different MOT frameworks in existing methods. (a) Separate Detection and Embedding (SDE) [2]. (b) Joint Detection and Embedding (JDE) [50]. (c) The proposed Uni- fied Tracking Model (UTM) that constructed with Identity-Aware Feature Enhancement (IAFE) module. paradigm in MOT. Recently, a number of Tracking-By- Detection approaches have been proposed, which can be divided into two paradigms: Separate Detection and Em- bedding (SDE) [1, 2, 4, 5, 8, 10, 16, 25, 38, 43, 52, 54], and Joint Detection and Embedding (JDE) [15, 29, 45, 50, 64]. As illustrated in Figure 1(a), SDE can be divided into three independent components: object detection, feature embedding, and identity association1. Candidate bound- ing boxes (bboxes) are obtained by standard detectors in each frame first, then identity embedding of each bbox is extracted by the re-identification algorithms, finally linked across frames through identity association to generate tra- 1In this paper, identity association includes affinity computation and data association. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21876 jectories. Therefore, SDE mainly exploits refining detec- tion [2, 16], enhancing identity embedding [1, 25, 38, 52], or designing robust association algorithms [5, 8, 54] to im- prove tracking performance. With the maturity of multi- task learning, some approaches [45, 50, 64] propose JDE framework to reduce the computation cost. Different from SDE, JDE applies the one-shot tracker to generate object detections and corresponding visual embeddings simultane- ously, thus treating MOT as two independent components, shown in Figure 1(b). Since the identity association in the above two paradigms is independent of object detection and feature embedding, the significant clues contained in the tracklet cannot be applied to enhance the detection and em- bedding modules. To address the above problem, a feasible idea is in- troducing an auxiliary module to associate and propagate identity-aware knowledge between the identity association and the other two components. Therefore, we design a novel Identity-Aware Feature Enhancement (IAFE) module to achieve information interaction between different com- ponents, shown in Figure 1(c). In detail, IAFE propagates the identity-aware knowledge generated by identity asso- ciation module to enhance the backbone feature of object detection. Meanwhile, the enhanced backbone feature can be utilized by the feature embedding module to generate discriminative embeddings. With more accurate detections and robust embeddings, the identity association module can produce higher-quality tracklets. Therefore, object detec- tion, feature embedding, and identity association are in- volved with each other, thus forming a positive feedback loop with mutual benefits. As shown in Figure 2, the proposed Unified Tracking Model (UTM) is composed of IAFE, detection branch, embedding branch, identity association branch, and mem- ory aggregation module. The detection and embedding branches are applied to locate and identify each object of the current frame, and the identity association branch ap- plies graph matching to associate the candidate bboxes with history tracklets. To achieve information interaction, IAFE is proposed to bridge and benefit these three branches, which utilizes the identity-aware knowledge to boost the detection and embedding. Specifically, IAFE consists of two modules: Identity-Aware Boosting Attention (IABA) and Identity-Aware Erasing Attention (IAEA), where IABA boosts the consistent regions between the current frame fea- ture and identity-aware knowledge, and IAEA erases the distracted regions in the current frame feature. With the designed IAFE, UTM constructs a positive feedback loop among all the three branches to improve the performance of MOT. Furthermore, we introduce a memory aggregation module to capture identity-aware knowledge through adap- tively selecting features of history frames, which can allevi- ate the effect of identity switches.The main contributions of the proposed method can be summarized as follows: • We design an Identity-Aware Feature Enhancement (IAFE) module to bridge and benefit different com- ponents in MOT. Specifically, it utilizes the identity- aware knowledge to boost backbone feature, which is further used to enhance the detection and embedding. • With the proposed IAFE, we construct a Unified Tracking Model (UTM) to form a positive feedback loop with mutual benefits. • The evaluation of public and private detections on three benchmarks verifies the effectiveness and gen- eralization ability of the proposed UTM.

Zhang_LOGO_A_Long-Form_Video_Dataset_for_Group_Action_Quality_Assessment_CVPR_2023

Abstract Action quality assessment (AQA) has become an emerg- ing topic since it can be extensively applied in numerous scenarios. However, most existing methods and datasets fo- cus on single-person short-sequence scenes, hindering the application of AQA in more complex situations. To address this issue, we construct a new multi-person long-form video dataset for action quality assessment named LOGO. Dis- tinguished in scenario complexity, our dataset contains 200 videos from 26 artistic swimming events with 8 athletes in each sample along with an average duration of 204.2 sec- onds. As for richness in annotations, LOGO includes for- mation labels to depict group information of multiple ath- letes and detailed annotations on action procedures. Fur- thermore, we propose a simple yet effective method to model relations among athletes and reason about the potential *indicates the corresponding author.temporal logic in long-form videos. Specifically, we de- sign a group-aware attention module, which can be eas- ily plugged into existing AQA methods, to enrich the clip- wise representations based on contextual group informa- tion. To benchmark LOGO, we systematically conduct in- vestigations on the performance of several popular meth- ods in AQA and action segmentation. The results reveal the challenges our dataset brings. Extensive experiments also show that our approach achieves state-of-the-art on the LOGO dataset. The dataset and code will be released at https://github.com/shiyi-zh0408/LOGO .

1. Introduction Action quality assessment (AQA) is applicable to many real-world contexts where people evaluate how well a spe- cific action is performed such as sports events [13, 18, 29– This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2405 Table 1. Comparisons of LOGO and existing datasets of action quality assessment (upper part of the table) and group activity recognition (lower part of the table). Score indicates the score annotations; Action denotes action types and temporal boundaries; Act.Label indicates the action types of both individuals and groups. Bbox indicates bounding boxes for actors. Formation represents formation annotations. Temp. indicates temporal boundary, Spat. indicates spatial localization. Dataset Duration Avg.Dur. Anno.Type Samples Events Form.Anno. Year MTL Dive [33] 15m,54s 6.0s Score 159 1 ✗ 2014 UNLV Dive [31] 23m,26s 3.8s Score 370 1 ✗ 2017 AQA-7-Dive [29] 37m,31s 4.1s Score 549 6 ✗ 2019 MTL-AQA [30] 96m,29s 4.1s Action,Score 1412 16 ✗ 2019 Rhythmic Gymnastics [53] 26h,23m,20s 1m,35s Score 1000 4 ✗ 2020 FSD-10 [26] - 3-30s Action,Score 1484 - ✗ 2020 FineDiving [50] 3h,30m,0s 4.2s Action,Score 3000 30 ✗ 2022 Collective Activity [7] - - Act.Label 44 - ✗ 2009 NCAA [35] 16h,9m,52s 4s Bbox,Act.Label 11436 - ✗ 2016 V olleyball [16] 2h,12m,1s 1.64s Bbox,Act.Label 4830 - ✗ 2016 FineGym [38] 161h,1m,45s 45.7s Act.Label,Temp. 12685 10 ✗ 2020 Multisports [25] 18h,34m,40s 20.9s Act.Label,Temp.,Spat. 3200 247 ✗ 2021 LOGO (Ours) 11h,20m,41s 3m,24s Action,Formation,Score 200 26 ✓(15764) 33, 46], healthcare [28, 39, 55–58], art performances, mil- itary parades, and others. Due to the extensive applica- tion of AQA, many efforts have been made over the past few years. Although some existing works have achieved promising performances in several simple scenarios, the ap- plication of AQA in many situations is still difficult to im- plement. In the data-driven era, the richness of the dataset largely determines whether the model can be applied to a wide range of scenarios or not. Inspired by this, we re- viewed the existing datasets in AQA and concluded that they are not rich enough for the following two reasons: Simplicity of Scenarios. We argue that the complex- ity of the AQA application scenes is reflected in two as- pects, the number of people and the duration of the videos. In snowboarding, for example, there is only one performer, while there are multiple actors in a military parade. In div- ing, the duration is only about 5 seconds, while in artis- tic swimming and dance performances, the duration of the action reaches several minutes. However, most existing datasets contain a single performer in each sample and many of them collect videos of 3-8s [29–31, 33, 50], which makes it difficult for existing methods to model complex scenes with more actors and longer duration. In such cases, just focusing on how well actions are performed by each individual may be insufficient. Relations among actors should be built, and the potential temporal logic in long- term videos should be modeled. Coarse-grained Annotations. Though there have been some long-form video datasets in AQA [49, 53] which pro- vide more complex scenes, they typically contain the score as the only annotation. Such coarse-grained annotationmakes it difficult for models to learn deeper information, es- pecially in more complex situations. Simply judging action quality via regressing a score for a long-term video could be confusing since we cannot figure out how the model de- termines whether an action is well-performed or not. To address these issues, we propose a multi-person long- form video dataset, LOGO (short for LOng-form GrOup). With 8 actors in each sample, LOGO contains videos with an average duration of 204.2s, much longer than most ex- isting datasets in AQA, making the scenes more complex. Besides, as shown in Figure 1, LOGO contains fine-grained annotations including frame-wise action type labels and temporal boundaries of actions for each video. We also devise formation labels to depict relations among actors. Specifically, we use a convex polygon to represent the for- mation actors perform, which reflects their position infor- mation and group information. In general, LOGO is distin- guished by its scenario complexity, while it also provides richer annotations compared to most existing datasets in AQA [29–31, 33, 49, 50, 53]. Furthermore, we build a plug-and-play module, GOAT (short for GrOup-aware ATtention), to bridge the gap be- tween single-person short-sequence scenarios and multi- person long-sequence scenarios. Specifically, in the spa- tial domain, by building a graph for actors, we model the relations among them. The nodes of this graph represent actors’ features extracted from a CNN, and the edges repre- sent the relations among actors. Then we use a graph con- volution network (GCN) [19] to model the group features from the graph. The optimized features of the graph then serve as “ queries ” and “ keys” for GOAT. In the temporal do- 2406 main, after feature extraction by the video feature backbone, the clip-wise features serve as “ values ” for GOAT. Instead of fusing the features simply using the average pooling as most previous works [44,50,52], GOAT learns the relations among clips and models the temporal features of the long- term videos based on the spatial information in every clip. The contributions of this paper can be summarized as: (1) We construct the first multi-person long-form video dataset, LOGO, for action quality assessment. To the best of our knowledge, LOGO stands out for its longer aver- age duration, the larger number of people, and richer an- notations when compared to most existing datasets. Ex- perimental results also reveal the challenges our proposed dataset brings. (2) We propose a plug-and-play group-aware module, GOAT, which models the group information and the temporal contextual relations for input videos, bridg- ing the gap between single-person short-sequence scenarios and multi-person long-sequence scenarios. (3) Experimen- tal results demonstrate that our group-aware approach ob- tains substantial improvements compared to existing meth- ods in AQA and achieves the state-of-the-art.

Yang_Geometry_and_Uncertainty-Aware_3D_Point_Cloud_Class-Incremental_Semantic_Segmentation_CVPR_2023

Abstract Despite the significant recent progress made on 3D point cloud semantic segmentation, the current methods re-quire training data for all classes at once, and are not suit- able for real-life scenarios where new categories are be- ing continuously discovered. Substantial memory storageand expensive re-training is required to update the modelto sequentially arriving data for new concepts. In thispaper , to continually learn new categories using previous knowledge, we introduce class-incremental semantic seg-mentation of 3D point cloud. Unlike 2D images, 3D pointclouds are disordered and unstructured, making it difficult to store and transfer knowledge especially when the pre-vious data is not available. We further face the challengeof semantic shift, where previous/future classes are indis- criminately collapsed and treated as the background in thecurrent step, causing a dramatic performance drop on pastclasses. We exploit the structure of point cloud and pro- pose two strategies to address these challenges. First, wedesign a geometry-aware distillation module that transferspoint-wise feature associations in terms of their geomet- ric characteristics. To counter forgetting caused by the semantic shift, we further develop an uncertainty-awarepseudo-labelling scheme that eliminates noise in uncertainpseudo-labels by label propagation within a local neighbor-hood. Our extensive experiments on S3DIS and ScanNet ina class-incremental setting show impressive results compa- rable to the joint training strategy (upper bound). Code isavailable at: https://github.com/leolyj/3DPC-CISS

1. Introduction The semantic segmentation of 3D point cloud plays a crucial role in applications such as virtual reality, roboticsand autonomous vehicles. In recent years, a number of pointcloud segmentation methods [ 16,28,29,38,39,45]h a v e Corresponding Author: Yinjie Lei ([email protected])Input Base Point Cloud Base ModelTraining on Base Classes { ceiling , floor , wall, table } Training on Novel Classes { chair , clutter } Input Novel Point CloudLabels Geometry-aware Feature- relation Transfer ( GFT ) Sec. 3.2 Mixed Labels ˇ Novel LabelsPesudo Labels=base$ novel$ Novel ModelUncertainty-aware Pseudo- label Generation ( UPG ) Sec. 3.3 Novel model for Inference on both Base and Novel classesPesudo Labels Prediction Branch Figure 1. To continually learn new categories without forgetting the previous ones, we leverage Geometry-aware Feature-relationTransfer (Sec. 3.2) to distill point-wise relationships from the base model and further employ Uncertainty-aware Pseudo-label Gener-ation (Sec. 3.3) to synthesize pseudo labels of old classes with low uncertainty as guidance for novel model training. achieved remarkable performance in the traditional setting where all classes are learned at once. Nevertheless, newcategories are gradually discovered in real-life scenarios, and updating the model to cater for these new categories re- quires large memory storage and expensive re-training. Insuch situations, as illustrated in Fig. 1, class-incremental learning provides a promising paradigm, since it enablesprogressively learning novel knowledge in an efficient man- ner while preserving the previous capabilities. The existing research on class-incremental learning is mostly on 2D image classification [ 17,19,21,32] with some efforts extended to RGB semantic segmentation [ 3,4,10, 41]. These methods employ a strategy based upon regu- larization [11,19,44],rehearsal/replay [2,17,24,32,35] orknowledge distillation [8,20,21] to preserve previous knowledge. At present, only a few works have investigated This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21759 3D point clouds based incremental learning for classifica- tion [ 5,6,9,22,43]. They focus on the classification of an individual object and extend 2D methods to 3D. Unlikeclassification which only considers a single object (not thescene with multiple objects), continually learning to seg- ment 3D point cloud in complex scenes introduces multiplenew challenges and has not been previously studied. 3D point clouds are disordered and unstructured, which makes it difficult to preserve previous knowledge and re-sults in catastrophic forgetting [13,33,36]. This specifically becomes pronounced when old data is not available. Wefurther observe that the 3D class-incremental segmentationfaces the phenomenon of semantic shift, where the pointsbelonging to old classes are indiscriminately collapsed intobackground during the current learning step. The semanticshift further suppresses the capability of the model to rec-ognize old categories, thus exacerbating forgetting. In this paper, we are the first to propose a class- incremental learning approach for 3D point cloud seman-tic segmentation. To prevent forgetting caused by unstruc- tured point clouds, we design a Geometry-aware Feature- relation Transfer (GFT) strategy to transfer the structural re-lationships among point features. Moreover, to address thesemantic shift issue, we assign uncertainty-aware pseudo-labels to the background points. Different from the con-ventional approaches, where pseudo-labels are directly ob-tained from the old model, we estimate uncertainties ac-cording to the distribution characteristics of points, andleverage the neighborhood information to propagate labelsfrom low to high uncertainties. Our Uncertainty-awarePseudo-label Generation (UPG), therefore, assists in elimi-nating the influence of noisy labels and helps tackle the se- mantic shift issue. Note that our approach does not involve any rehearsal or memory replay buffer to store old data orits annotations during the incremental process. We showpromises of our approach through comprehensive evalua-tions on benchmarks defined on public datasets i.e.,S3DIS [1] and ScanNet [ 7]. Our key contributions are: • A class-incremental learning framework for 3D point cloud semantic segmentation, to sequentially adapt tonew classes from previous acquired knowledge. • To transfer previous knowledge and prevent forgetting caused by unstructured nature of the point clouds, wepropose a Geometry-aware Feature-relation Transfer (GFT) module that captures the point-wise feature re-lations based on the geometric information. • To tackle the semantic shift issue where old classes are indiscriminately collapsed into the background, wedesign an Uncertainty-aware Pseudo-label Generation(UPG) strategy to enhance pseudo-labelling quality and thus provide effective guidance for old classes. • Compared with several baselines on multiple bench-marks, our approach achieves promising results for 3D class-incremental semantic segmentation, closer to thejoint training (upper bound) using all data at once.

Yin_Multi-Space_Neural_Radiance_Fields_CVPR_2023

Abstract Existing Neural Radiance Fields (NeRF) methods suf- fer from the existence of reflective objects, often result- ing in blurry or distorted rendering. Instead of calculat- ing a single radiance field, we propose a multi-space neu- ral radiance field (MS-NeRF) that represents the scene us- ing a group of feature fields in parallel sub-spaces, which leads to a better understanding of the neural network to- ward the existence of reflective and refractive objects. Our multi-space scheme works as an enhancement to existing NeRF methods, with only small computational overheads needed for training and inferring the extra-space outputs. We demonstrate the superiority and compatibility of our ap- proach using three representative NeRF-based models, i.e., NeRF , Mip-NeRF , and Mip-NeRF 360. Comparisons are performed on a novelly constructed dataset consisting of 25 synthetic scenes and 7 real captured scenes with complex reflection and refraction, all having 360-degree viewpoints. Extensive experiments show that our approach significantly outperforms the existing single-space NeRF methods for rendering high-quality scenes concerned with complex light paths through mirror-like objects. Our code and dataset will be publicly available at https://zx-yin.github.io/msnerf.

1. Introduction Neural Radiance Fields (NeRF) [25] and its variants are refreshing the community of neural rendering, and the po- tential for more promising applications is still under ex- ploration. NeRF represents scenes as continuous radiance fields stored by simple Multi-layer Perceptrons (MLPs) and renders novel views by integrating the densities and radi- ance, which are queried from the MLPs by points sampled along the ray from the camera to the image plane. Since its first presentation [25], many efforts have been investigated to enhance the method, such as extending to unbounded scenes [2, 50], handling moving objects [29, 30, 37], or re- constructing from pictures in the wild [6, 21, 35, 49]. *Bo Ren is the corresponding author. (a) Mip-NeRF 360 [2], SSIM=0.825 (b) Our Model, SSIM=0.881 Figure 1. (a) Though Mip-NeRF 360 can handle unbounded scenes, it still suffers from reflective surfaces, as the virtual images violate the multi-view consistency, which is of vital importance to NeRF-based methods. (b) Our method can help conventional NeRF-like methods learn the virtual images with little extra cost. However, rendering scenes with mirrors is still a chal- lenging task for state-of-the-art NeRF-like methods. One of the principle assumptions for the NeRF method is the multi- view consistency property of the target scenes [16, 20, 36]. When there are mirrors in the space, if one allows the view- points to move 360-degree around the scene, there is no consistency between the front and back views of a mirror, since the mirror surface and its reflected virtual image are only visible from a small range of views. As a result, it is often required to manually label the reflective surfaces in order to avoid falling into sub-optimal convergences [12]. In this paper, we propose a novel multi-space NeRF- based method to allow the automatic handling of mirror-like objects in the 360-degree high-fidelity rendering of scenes without any manual labeling. Instead of regarding the Eu- clidean scene space as one single space, we treat it as com- posed of multiple virtual sub-spaces, whose composition changes according to location and view direction. We show that our approach using such a multi-space decomposition leads to successful handlements of complex reflections and refractions where the multi-view consistency is heavily vi- olated in the Euclidean real space. Furthermore, we show that the above benefits can be achieved by designing a low- cost multi-space module and replacing the original output layer with it. Therefore, our multi-space approach serves as This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12407 a general enhancement to the NeRF-based backbone, equip- ping most NeRF-like methods with the ability to model complex reflection and refraction, as shown in Fig. 1. Existing datasets have not paid enough attention to the 360-degree rendering of scenes containing mirror-like ob- jects, such as RFFR [12] just has forward-facing scenes, and the Shiny dataset in [42] with small viewpoints changes and cannot exhibit view-dependent effects in large angle scale. Therefore we construct a novel dataset dedicated to evaluation for the 360-degree high-fidelity rendering of scenes containing complex reflections and refractions. In this dataset, we collect 25 synthesized scenes and 7 captured real-world scenes. Each synthesized scene consists of 120 images of 360-degree around reflective or refractive objects, with 100 randomly split for training, 10 for validation, and 10 for evaluation. Each real-world scene is captured ran- domly around scenes with reflective and refractive objects, consisting of 62 to 118 images, and organized under the convention of LLFF [24]. We then demonstrate the supe- riority and compatibility of our approach by comparisons, using three representative baseline models, i.e., NeRF [25], Mip-NeRF [1], and Mip-NeRF 360 [2], with and without our multi-space module. Experiments show that our ap- proach improves performance by a large margin both quan- titatively and qualitatively on scenes with reflection and re- fraction. Our main contributions are as follows: • We propose a multi-space NeRF method that auto- matically handles mirror-like objects in 360-degree high-fidelity scene rendering, achieving significant im- provements over existing representative baselines both quantitatively and qualitatively. • We design a lightweight module that can equip most NeRF-like methods with the ability to model reflection and refraction with small computational overheads. • We construct a dataset dedicated to evaluation for the 360-degree high-fidelity rendering of scenes contain- ing complex reflections and refractions, including 25 synthesized scenes and 7real captured scenes.

Yu_Hint-Aug_Drawing_Hints_From_Foundation_Vision_Transformers_Towards_Boosted_Few-Shot_CVPR_2023

Abstract Despite the growing demand for tuning foundation vision transformers (FViTs) on downstream tasks, fully unleash- ing FViTs’ potential under data-limited scenarios (e.g., few- shot tuning) remains a challenge due to FViTs’ data-hungry nature. Common data augmentation techniques fall short in this context due to the limited features contained in the few-shot tuning data. To tackle this challenge, we first iden- tify an opportunity for FViTs in few-shot tuning: pretrained FViTs themselves have already learned highly representa- tive features from large-scale pretraining data, which are fully preserved during widely used parameter-efficient tun- ing. We thus hypothesize that leveraging those learned fea- tures to augment the tuning data can boost the effective- ness of few-shot FViT tuning. To this end, we propose a framework called Hint -based Data Augmentation ( Hint- Aug), which aims to boost FViT in few-shot tuning by aug- menting the over-fitted parts of tuning samples with the learned features of pretrained FViTs. Specifically, Hint-Aug integrates two key enablers: (1) an Attentive Over-fitting Detector ( AOD ) to detect over-confident patches of founda- tion ViTs for potentially alleviating their over-fitting on the few-shot tuning data and (2) a Confusion-based Feature Infusion ( CFI) module to infuse easy-to-confuse features from the pretrained FViTs with the over-confident patches detected by the above AOD in order to enhance the feature diversity during tuning. Extensive experiments and ablation studies on five datasets and three parameter-efficient tuning techniques consistently validate Hint-Aug’s effectiveness: 0.04%∼32.91% higher accuracy over the state-of-the-art (SOTA) data augmentation method under various low-shot settings. For example, on the Pet dataset, Hint-Aug achieves a 2.22% higher accuracy with 50% less training data over SOTA data augmentation methods.

1. Introduction Foundation vision transformers (FViTs) [16, 41, 54, 55, 64] with billions of floating point operations (FLOPs) and parameters have recently demonstrated significant poten-tial in various downstream tasks [40, 41]. The success of FViTs has ushered in a new paradigm in deep learning: pretraining-then-tuning [16, 40, 67], which first pretrains an FViT on a large-scale dataset, then uses recently developed parameter-efficient tuning methods (e.g., visual prompt tun- ing (VPT) [34], visual prompting [2], LoRA [33], and Adapter [72]) to tune pretrained FViTs on downstream tasks with limited tuning data. However, although it is highly de- sirable, effectively tuning pretrained FViTs for real-world applications, especially under few-shot tuning scenarios, re- mains a particularly challenging task. The reason is that al- though parameter-efficient tuning methods are dedicatedly designed for FViTs and can alleviate the overfitting issue by reducing the number of trainable parameters [2, 34, 72], the data-hungry nature of FViTs [16, 54] is not mitigated and thus the achievable accuracy under data-limited scenarios (e.g., few-shot tuning scenarios) are still limited. Therefore, how to effectively tune pretrained FViTs on various down- stream tasks with few-shot tuning is still an open question. To enhance the effectiveness of parameter-efficient FViT tuning under few-shot settings, one promising direction is to leverage data augmentation techniques to increase the data diversity and thus the feature diversity of the models when being tuned on few-shot data, boosting the achievable accu- racy [12, 31, 68, 71]. Nevertheless, it has been shown that existing data augmentation techniques fall short in boost- ing the model accuracy under few-shot tuning scenarios. This is because most of the existing data augmentation tech- niques are random-based (e.g., RandAugment [13], Au- toAugment [12], color jitter, mixup [71], and cutmix [68]), which only randomly permute existing features in the train- ing data and thus cannot generate new and meaningful features [63]. As illustrated in Fig. 1, we observe that neither the widely-used random-based data augmentation techniques (i.e., a dedicated combination of techniques in- cluding RandAugment [13], color jitter, and random eras- ing [74] as in [72]) nor training without data augmentation can consistently achieve a satisfactory accuracy across dif- †Our code is available at https://github.com/GATECH- EIC/Hint-Aug This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11102 4-shot on Aircraft 12-shot on Aircraft 16-shot on Aircraft 8-shot on Pets 12-shot on Pets16-shot on Pets4-shot on Food8-shot on Food12-shot on Food16-shot on Food 8-shot on AircraftNo-Aug Ours NPS 4-shot on PetsFigure 1. The normalized achieved accuracies when few-shot tun- ing the ViT-base model [16] on various datasets and numbers of tuning shots using (1) vanilla training without augmentation (i.e., No-Aug), (2) the SOTA parameter-efficient tuning technique [72] (i.e., NPS), and (3) our proposed Hint-Aug. ferent datasets under few-shot tuning. Specifically, when being applied to fine-grained classification tasks, e.g., the Aircraft dataset [45], these random-based data augmenta- tion techniques actually hurt the achievable accuracy. The reason is that random-based data augmentation techniques can easily create out-of-manifold samples [68, 71], espe- cially on commonly used fine-grained datasets. Such out- of-manifold samples can largely degrade the achievable ac- curacy given the limited number of training samples under few-shot tuning scenarios [27]. Therefore, it is crucial to develop data augmentation techniques that can adaptively augment the given training samples with diverse, but still within-manifold, features to boost the effectiveness of tun- ing FViTs on various downstream tasks. This work sets out to close the increasing gap be- tween the growing demand for effective few-shot FViT tun- ing and the unsatisfactory achievable accuracy by exist- ing techniques. In particular, we identify that in few-shot parameter-efficient tuning, the pretrained FViTs’ weights are fixed during tuning. Meanwhile, existing works have shown that (1) pretrained transformer models have already learned complex but generalizable features [16, 41, 54] and (2) gradient-based methods can extract the learned features from pretrained models and then add them to the input images [44, 46]. Therefore, we hypothesize that FViTs’ few-shot tuning accuracies can be non-trivially improved by leveraging the learned features in the pretrained FViTs. Specifically, we make the following contributions: • We propose a framework called Hint -based Data Augmentation ( Hint-Aug ), which is dedicated to boosting the achievable accuracy of FViTs under few-shot tuning scenarios by leveraging the learned features of pretrained FViTs to guide the data augmentation strategy used for thetraining dataset in an input-adaptive manner. • Our Hint-Aug framework integrates two enablers: (1) an Attentive Over-fitting Detector (AOD) to identify the over- fitting samples and patches in the given training dataset by making use of the attention maps of pretrained FViTs and (2) a Confusion-based Feature Infusion (CFI) module to adaptively infuse pretrained FViTs’ learned features into the training data to better tuning those models on down- stream tasks, alleviating the commonly recognized chal- lenge of having limited features under few-shot tuning. • Extensive experiments and ablation studies on five datasets and three parameter-efficient tuning techniques consis- tently validate the effectiveness of our proposed Hint-Aug framework, which achieves a 0.04%∼32.91% higher accuracy over state-of-the-art (SOTA) data augmentation methods [72] across different datasets and few-shot set- tings. For example, on the Pets dataset, Hint-Aug achieves a 2.22% higher accuracy with 50% less training data com- pared with the SOTA augmentation method.

Zhang_GeoMVSNet_Learning_Multi-View_Stereo_With_Geometry_Perception_CVPR_2023

Abstract Recent cascade Multi-View Stereo (MVS) methods can efficiently estimate high-resolution depth maps through nar- rowing hypothesis ranges. However, previous methods ig- nored the vital geometric information embedded in coarse stages, leading to vulnerable cost matching and sub-optimal reconstruction results. In this paper, we propose a ge- ometry awareness model, termed GeoMVSNet , to explic- itly integrate geometric clues implied in coarse stages for delicate depth estimation. In particular, we design a two- branch geometry fusion network to extract geometric pri- ors from coarse estimations to enhance structural feature extraction at finer stages. Besides, we embed the coarse probability volumes, which encode valuable depth distri- bution attributes, into the lightweight regularization net- work to further strengthen depth-wise geometry intuition. Meanwhile, we apply the frequency domain filtering to mit- igate the negative impact of the high-frequency regions and adopt the curriculum learning strategy to progres- sively boost the geometry integration of the model. To in- tensify the full-scene geometry perception of our model, we present the depth distribution similarity loss based on the Gaussian-Mixture Model assumption. Extensive exper- iments on DTU and Tanks and Temples (T &T) datasets demonstrate that our GeoMVSNet achieves state-of-the-art results and ranks first on the T &T-Advanced set. Code is available at https://github.com/doubleZ0108/GeoMVSNet.

1. Introduction Multi-View Stereo (MVS) reconstructs the dense ge- ometry representation of a scene from multiple overlap- ping photographs, which is an influential branch of three- dimensional (3D) computer vision and has been extensively studied for decades. Learning-based MVS methods aggre- gate cost volume from different viewpoints and use neu- ral networks for cost regularization, which achieve superior performance compared with traditional methods. Recently, cascade-based architectures [7, 14, 54] havebeen widely applied. They compute different resolution depth maps in a coarse-to-fine manner and progressively narrow hypothesis plane guidance to reduce computational complexity. However, these approaches do not take ad- vantage of valuable insight contained in early phases and only consider the pixel-wise depth attribute. Some meth- ods, e.g. deformable kernel-based [47] and transformer- based [4, 8, 22, 27, 46], introduce finely designed external structures for feature extraction but do not fully exploit the geometric clues embedded in the MVS scenarios. Unlike existing works, we propose to explore the ge- ometric structures embedded in coarse stages for delicate estimations in finer stages. In particular, we build a two- branch fusion network to integrate geometric priors con- tained in coarse depth maps with ordinary features extracted by the classic FPN [23], and the fused geometry awareness features can provide solid foundations for robust aggrega- tion. Meanwhile, coarse probability volumes with abundant geometric structures are embedded into the regularization network, and we replace the heavy 3D convolution with en- hanced 2D regularization without degrading the quality of depth-wise correlation, resulting in lightweight but robust cost matching. However, MVS networks tend to produce severe misestimation at high-frequency clutter textures due to confused matching in coarse stages, which inevitably af- fects explicit geometry perception. We are inspired by the human behavior that a nearsighted person can still perceive a scene well without glasses, even if the texture details can- not be seen clearly. Based on the observation, we refer to the idea of curriculum learning [2] to embed coarse geo- metric priors into finer stages from easy to difficult. Specif- ically, we utilize the frequency domain filtering strategy to effectively alleviate redundant high-frequency textures without producing more learning parameters and leverage geometric structures embedded in different hierarchies of frequency for gradually delicate depth estimation. In addition, depth ranges of MVS scenarios are often concentrated in several intervals, for this, we adopt the Gaussian-Mixture Model to simulate full-scene depth dis- tribution and PauTa Criterion [31] allows us to depict loca- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21508 tions that are too close or too far hidden in the long tailing of the depth distribution curve, e.g. sky. The depth distribution loss is proposed finally for full-scene similarity supervision. In summary, the main contributions are as follows. • We propose the geometric prior guided feature fusion and the probability volume geometry embedding ap- proaches for robust cost matching. • We enhance geometry awareness via the frequency do- main filtering strategy and adopt the idea of curriculum learning for progressively introducing geometric clues from easy to difficult. • We model the depth distribution of MVS scenarios us- ing the Gaussian-Mixture Model assumption and build the full-scene geometry perception loss function. • The proposed method is extensively evaluated on the DTU dataset and both intermediate and advanced sets of Tanks and Temples benchmark, all achieving brand- new state-of-the-art performance.

Ye_Self-Supervised_Super-Plane_for_Neural_3D_Reconstruction_CVPR_2023

Abstract Neural implicit surface representation methods show impressive reconstruction results but struggle to handle texture-less planar regions that widely exist in indoor scenes. Existing approaches addressing this leverage image prior that requires assistive networks trained with large- scale annotated datasets. In this work, we introduce a self-supervised super-plane constraint by exploring the free geometry cues from the predicted surface, which can fur- ther regularize the reconstruction of plane regions without any other ground truth annotations. Specifically, we in- troduce an iterative training scheme, where (i) grouping of pixels to formulate a super-plane (analogous to super- pixels), and (ii) optimizing of the scene reconstruction net- work via a super-plane constraint, are progressively con- ducted. We demonstrate that the model trained with super- planes surprisingly outperforms the one using conventional annotated planes, as individual super-plane statistically oc- cupies a larger area and leads to more stable training. Ex- tensive experiments show that our self-supervised super- plane constraint significantly improves 3D reconstruction quality even better than using ground truth plane segmen- tation. Additionally, the plane reconstruction results from our model can be used for auto-labeling for other vision tasks. The code and models are available at https: //github.com/botaoye/S3PRecon .

1. Introduction Reconstructing 3D scenes from multi-view RGB images is an important but challenging task in computer vision, which has numerous applications in autonomous driving, virtual reality, robotics, etc. Existing matching-based meth- ods [30, 31, 50] estimate per-view depth maps, which are then fused to construct 3D representation. However, they do not recover the depth of the scene in texture-less planar ar- eas well (such as walls, floors, and other solid color planes), which are abundant, especially for indoor scenes. Recent data-driven methods [22, 26, 34, 38] alleviate this problem to some extent by automatically learning geometric priors from large-scale training data, but they either require nu- Input View Surface Reconstruction Plane SegmentationFigure 1. Reconstruction and Plane Segmentation Results. Our method can reconstruct smooth and complete planar regions by employing the super-plane constraint and further obtain plane seg- mentation in an unsupervised manner. merous and expensive 3D supervision ( e.g., depth [22, 38], normal [14], etc.) or lack fine-grain details [26, 34]. Recently, neural implicit representations have gained much attention and shown impressive reconstruction results without 3D supervision [39, 46, 47]. However, these meth- ods purely rely on the photo-consistency to construct the scene, which leads to texture-geometry ambiguities since there are different plausible interpretations to satisfy this objective. Several approaches address this problem by in- troducing additional priors obtained from trained models that can be considered as assistant networks . For instance, ManhattanSDF [9] introduces the Manhattan assumption on the floor and wall regions, which are predicted by a seman- tic segmentation model. NeuRIS [37] and MonoSDF [48] adopt additional normal supervision acquired from normal prediction networks trained on large-scale labeled datasets. Although these methods can regularize the reconstruction process on indoor scenes, they all rely on large-scale an- notated 2D or 3D datasets. In addition, these pretrained geometric prediction networks are sensitive to different scenes and not friendly across different domains or datasets. For example, MonoSDF [48] reports that different normal prediction networks significantly affect the reconstruction quality. Thus, a natural question arises: can we improve the RGB-based reconstruction results on texture-less regions without any implicit supervision from assistant networks? In this work, we propose a novel neural 3D reconstruc- tion framework with the Self-Supervised Super-Plane con- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21415 straint, called S3P, which does not require any labeled data or assistant networks. The intuition behind our approach is simple: the constructed results provide a free geometry cue, i.e., surface normal, which can be utilized to guide the reconstruction process of planar regions. Specifically, pix- els belonging to parallel planes tend to have similar normal directions (shown in Fig 3). We group pixels sharing sim- ilar normal values into the same cluster, which we call a super-plane (analogous to a group of pixels is called super- pixel). A super-plane constraint is then applied to force normal directions within the same super-plane to be con- sistent, thus constraining the reconstruction of large super- plane regions. Due to the ambiguity of the prediction, espe- cially at the early training stages, the grouped super-plane can be inaccurate and introduces noisy self-supervision sig- nals. Therefore, an automatic filtering strategy is further designed to compare the normals of each pixel with the es- timated super-plane normal, i.e., the average normal of all pixels belonging to the same super-plane, and outliers with large angular differences will be filtered out. We also detect the discontinuity of the surface according to the geometry and color features and mask out non-plane edge regions in the super-plane segmentation maps. Notably, by using normals for grouping, multiple paral- lel planes will be grouped together, so that our super-planes are typically larger than individual planes. This property is particularly beneficial for volume rendering-based training process [25] since planes with more pixels will also have more stable and accurate averaged normals when limited pixels are sampled in each iteration. We experimentally verify this benefit: our super-plane constraint yields better results than adopting ground truth plane segmentation. As a by-product, self-supervised plane segmentation of the reconstructed scene can be easily obtained from the super-plane masks by extracting connected components separated by the detected non-plane edge regions. Thus, our approach can be extended to reconstruct planes of a scene without ground truth supervision. It can also be applied to label new scenes automatically for applications that require such training data. The main contributions of this work are: • We introduce a super-plane constraint for neural im- plicit scene reconstruction by first generating super- planes in an unsupervised manner and then performing automatic outlier filtering. • Our super-plane segmentation method can be further extended to get unsupervised plane reconstruction re- sults, which can be used as auto-labeling. • Experimental results show that our method signifi- cantly improves the reconstruction quality of texture- less planar regions, and the unsupervised plane recon- struction results are comparable to those from state-of- the-art supervised methods.MethodsExplicit 3D supervisionImplicit 2D/3D supervisionHandle texture-less Patch Match-based MVS × × × Data-driven MVS ✓ × ✓ NeuS [39], V olSDF [46] × × × ManhattanSDF [9] × ✓(2D) ✓ NeuRIS [37], MonoSDF [48] × ✓(3D) ✓ Ours × × ✓ Table 1. Comparison between different reconstruction meth- ods. Our method can handle texture-less planar regions without implicit supervision provided by assistant networks.

Xu_OmniAvatar_Geometry-Guided_Controllable_3D_Head_Synthesis_CVPR_2023

Abstract We present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images that is capable of synthesizing diverse identity- preserved 3D heads with compelling dynamic details un- der full disentangled control over camera poses, facial ex- pressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control, we first ex- plicitly define a novel semantic signed distance function (SDF) around a head geometry (FLAME) conditioned on the control parameters. This semantic SDF allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. We then leverage the 3D- aware GAN framework (EG3D) to synthesize detailed shape and appearance of 3D full heads in the canonical space, fol- lowed by a volume rendering step guided by the volumetric correspondence map to output into the observation space. To ensure the control accuracy on the synthesized headshapes and expressions, we introduce a geometry prior loss to conform to head SDF and a control loss to conform to the expression code. Further, we enhance the temporal realism with dynamic details conditioned upon varying expressions and joint poses. Our model can synthesize more preferable identity-preserved 3D heads with compelling dynamic de- tails compared to the state-of-the-art methods both qualita- tively and quantitatively. We also provide an ablation study to justify many of our system design choices.

1. Introduction Photo-realistic face image synthesis, editing and ani- mation attract significant interests in computer vision and graphics, with a wide range of important downstream appli- cations in visual effects, digital avatars, telepresence and many others. With the advent of Generative Adversar- ial Networks (GANs) [15], remarkable progress has been achieved in face image synthesis by StyleGAN [23–25] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12814 as well as in semantic and style editing for face im- ages [46, 54]. To manipulate and animate the expressions and poses in face images, many methods attempted to lever- age 3D parametric face models, such as 3D Morphable Models (3DMMs) [6, 40], with StyleGAN-based synthesis models [10, 41, 53]. However, all these methods operate on 2D convolutional networks (CNNs) without explicitly en- forcing the underlying 3D face structure. Therefore they cannot strictly maintain the 3D consistency when synthe- sizing faces under different poses and expressions. Recently, a line of work has explored neural 3D rep- resentations by unsupervised training of 3D-aware GANs from in-the-wild unstructured images [7, 8, 11, 17, 37, 44, 45, 48, 61, 62, 67]. Among them, methods with generative Neural Radiance Fields (NeRFs) [33] have demonstrated striking quality and multi-view-consistent image synthesis [7,11,17,37,48]. The progress is largely due to the integra- tion of the power of StyleGAN in photo-realistic image syn- thesis and NeRF representation in 3D scene modeling with view-consistent volumetric rendering. Nevertheless, these methods lack precise 3D control over the generated faces beyond camera pose, as well as the quality and consistency in control over other attributes, such as shape, expression, neck and jaw pose, leave much to be desired. In this work, we present OmniAvatar , a novel geometry- guided 3D head synthesis model trained from in-the-wild unstructured images. Our model can synthesize a wide range of 3D human heads with full control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled con- trol for 3D human head synthesis, we devise our model learning in two stages . We first define a novel semantic signed distance function (SDF) around a head geometry (i.e. FLAME [29]) conditioned on its control parameters. This semantic SDF fully distills rich 3D geometric prior knowl- edge from the statistical FLAME model and allows us to build a differentiable volumetric correspondence map from theobservation space to a disentangled canonical space from all the control parameters. In the second training stage, we then leverage the state-of-the-art 3D GAN framework (EG3D [7]) to synthesize realistic shape and appearance of 3D human heads in the canonical space, including the mod- eling of hair and apparels. Following that, a volume render- ing step is guided by the volumetric correspondence map to output the geometry and image in the observation space. To ensure the consistency of synthesized 3D head shape with controlling head geometry, we introduce a geometry prior loss to minimize the difference between the synthe- sized neural density field and the FLAME head SDF in ob- servation space. Furthermore, to improve the control ac- curacy, we pre-train an image encoder of the control pa- rameters and formulate a control loss to ensure synthesized images matching the input control code upon encoding. An-other key aspect of synthesis realism is dynamic details such as wrinkles and varying shading as subjects change expres- sions and poses. To synthesize dynamic details, we propose to condition EG3D’s triplane feature decoding with noised controlling expression. Compare to state-of-the-art methods, our method achieves superior synthesized image quality in terms of Frechet Inception Distance (FID) and Kernel Inception Dis- tance (KID). Our method can consistently preserve the iden- tity of synthesized subjects with compelling dynamic de- tails while changing expressions and poses, outperforming prior methods both quantitatively and qualitatively. The contributions of our work can be summarized as: • A novel geometry-guided 3D GAN framework for high-quality 3D head synthesis with full control on camera poses, facial expressions, head shapes, artic- ulated neck and jaw poses. • A novel semantic SDF formulation that defines the vol- umetric correspondence map from observation space to canonical space and allows full disentanglement of control parameters in 3D GAN training. • A geometric prior loss and a control loss to ensure the head shape and expression synthesis accuracy. • A robust noised expression conditioning scheme to en- able dynamic detail synthesis.