Datasets:

bskang
/

CVPR2023_title_abstract_intro

Dataset card Viewer Files Files and versions Community

Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Yao_Teacher-Generated_Spatial-Attention_Labels_Boost_Robustness_and_Accuracy_of_Contrastive_Models_CVPR_2023	Abstract Human spatial attention conveys information about the regions of visual scenes that are important for perform- ing visual tasks. Prior work has shown that the informa- tion about human attention can be leveraged to beneﬁt var- ious supervised vision tasks. Might providing this weak form of supervision be useful for self-supervised represen- tation learning? Addressing this question requires collect- ing large datasets with human attention labels. Yet, col- lecting such large scale data is very expensive. To address this challenge, we construct an auxiliary teacher model to predict human attention, trained on a relatively small la- beled dataset. This teacher model allows us to generate im- age (pseudo) attention labels for ImageNet. We then train a model with a primary contrastive objective; to this stan- dard conﬁguration, we add a simple output head trained to predict the attention map for each image, guided by the pseudo labels from teacher model. We measure the qual- ity of learned representations by evaluating classiﬁcation performance from the frozen learned embeddings as well as performance on image retrieval tasks (see supplementary material). We ﬁnd that the spatial-attention maps predicted from the contrastive model trained with teacher guidance aligns better with human attention compared to vanilla con- trastive models. Moreover, we ﬁnd that our approach im- proves classiﬁcation accuracy and robustness of the con- trastive models on ImageNet and ImageNet-C. Further, we ﬁnd that model representations become more useful for im- age retrieval task as measured by precision-recall perfor- mance on ImageNet, ImageNet-C, CIFAR10, and CIFAR10- C datasets. Figure 1. Illustration. A teacher model is trained to predict human spatial-attention from a small dataset. Then the model is used to provide attention labels for larger dataset, which are used as addi- tional targets for contrastive models.	1. Introduction Deep learning models have made signiﬁcant progress and obtained notable success on various vision tasks. De- spite these promising results, humans continue to perform better than deep learning models in many applications. A notable reason is that deep learning models have a tendency to learn “short-cuts”, i.e., giving signiﬁcance to physically meaningless patterns or exploiting features which are pre- dictive in some settings, but not causal [ 20]. Examples include focusing on less signiﬁcant features such as back- ground and textures [ 13]. These models yield representa- tions that are less generalizable and lead to models that are highly sensitive to small pixel modulations [ 42]. Human vision on the other hand is known to be much more robust and generalizable. One major difference be- tween human and machine vision is that humans tend to ⇤Equal technical contribution. †Equal leadership and advising contribution Correspondence to: [email protected] &[email protected] 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. 23282 focus on speciﬁc regions in visual scene [ 45]. These lo- cations often reﬂect regions salient or useful to perform a speciﬁc vision task. Machines, instead, initially place equal signiﬁcance to all regions. A natural question is: will it be beneﬁcial if machine vision models is guided by human spatial attention? Human spatial attention has been shown to beneﬁt com- puter vision models in supervised tasks, such as classiﬁca- tion [ 32]. Yet, it is still a question whether adding a form of weak supervision in the form of human spatial atten- tion could similarly beneﬁt self-supervised models that are trained end-to-end. Self-supervised models typically need a large amount of data to yield good representations. To test if training weakly supervised models with human spa- tial attention cues, we will need to collect a large volume of human spatial attention labels, which is a very expen- sive process that requires either using trackers to record eye movements [ 5,43,52] or asking humans to highlight regions that they attend to [ 25,27]. This process is prohibitively te- dious and costly for datasets with millions of examples. In this work, we test the hypothesis that a weak super- vision in the form of human spatial attention is beneﬁcial for representation learning for models trained with a con- trastive objective. Inspired by knowledge distillation and self-training using teacher models [ 47,49], we address the challenge of obtaining spatial attention labels on large scale image datasets by using machine pseudo-labeling. We train a teacher model on a set of limited ground truth human spa- tial attention labels, and use this teacher model to gener- ate spatial attention pseudo-labels for the large ImageNet benchmark. We are then able to utilize the generated spa- tial attention maps in the contrastive models, and discover that this approach yields representations that are highly pre- dictive of human spatial attention. Further, we ﬁnd that the learned representations are better as measured by higher ac- curacy and robustness on classiﬁcation downstream tasks, and higher precision and recall on image retrieval tasks. In- terestingly, we ﬁnd that the gains from using teacher mod- els to provide pseudo labels are larger than using the lim- ited ground truth human labels directly when training con- trastive models, and the gains are larger for contrastive mod- els than when applying same method to supervised models. In summary, our contributions are as follows: •We create a dataset with spatial attention maps for the ImageNet [ 37] benchmark by ﬁrst training a teacher model to predict human spatial attention labels from Salicon dataset [ 25] and then use the model to label ImageNet examples •We use spatial-attention labels from the teacher model as an additional prediction target to models trained Trained teacher model is available at: https://github.com/google-research/google-research/tree/master/human attention/with contrastive objective. •We ﬁnd that the proposed method can learn bet- ter representation, leading to better accuracy and ro- bustness for downstream classiﬁcation tasks (on Im- ageNet and ImageNet-C), and better performance on retrieval tasks (on ImageNet, ImageNet-C, CIFAR-10, and CIFAR10-C).
Yang_FreeNeRF_Improving_Few-Shot_Neural_Rendering_With_Free_Frequency_Regularization_CVPR_2023	Abstract Novel view synthesis with sparse inputs is a challeng- ing problem for neural radiance ﬁelds (NeRF). Recent ef- forts alleviate this challenge by introducing external super- vision, such as pre-trained models and extra depth signals, or by using non-trivial patch-based rendering. In this pa- per, we present Frequency regularized NeRF (FreeNeRF), a surprisingly simple baseline that outperforms previous methods with minimal modiﬁcations to plain NeRF . We an- alyze the key challenges in few-shot neural rendering and ﬁnd that frequency plays an important role in NeRF’s train- ing. Based on this analysis, we propose two regularization terms: one to regularize the frequency range of NeRF’s inputs, and the other to penalize the near-camera density ﬁelds. Both techniques are “free lunches” that come at noadditional computational cost. We demonstrate that even with just one line of code change, the original NeRF can achieve similar performance to other complicated methods in the few-shot setting. FreeNeRF achieves state-of-the- art performance across diverse datasets, including Blender, DTU, and LLFF . We hope that this simple baseline will mo- tivate a rethinking of the fundamental role of frequency in NeRF’s training, under both the low-data regime and be- yond. This project is released at FreeNeRF .	1. Introduction Neural Radiance Field (NeRF) [ 21] has gained tremen- dous attention in 3D computer vision and computer graph- ics due to its ability to render high-ﬁdelity novel views. 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. 8254 However, NeRF is prone to overﬁtting to training views and struggles with novel view synthesis when only a few inputs are available. We term this view synthesis from sparse in- puts problem as a few-shot neural rendering problem. Existing methods address this challenge using different strategies. Transfer learning methods, e.g., PixelNerf [ 37] and MVSNeRF [ 4], pre-train on large-scale curated multi- view datasets and further incorporate per-scene optimiza- tion at test time. Depth-supervised methods [ 6,29] in- troduce estimated depth as an external supervisory signal, leading to a complex training pipeline. Patch-based reg- ularization methods impose regularization from different sources on rendered patches, e.g., semantic consistency reg- ularization [ 11], geometry regularization [ 8,22], and ap- pearance regularization [ 22], all at the cost of computation overhead since an additional, non-trivial number of patches must be rendered during training [ 8,11,22]. In this work, we ﬁnd that a plain NeRF can work sur- prisingly well with none of the above strategies in the few- shot setting by adding (approximately) as few as oneline of code (see Fig. 1). Concretely, we analyze the common failure modes in training NeRF under a low-data regime. Drawing on this analysis, we propose two regularization terms. One is frequency regularization, which directly reg- ularizes the visible frequency bands of NeRF’s inputs to stabilize the learning process and avoid catastrophic over- ﬁtting at the start of training. The other is occlusion reg- ularization, which penalizes the near-camera density ﬁelds that cause “ﬂoaters,” another failure mode in the few-shot neural rendering problem. Combined, we call our method Frequency regularized NeRF (FreeNeRF), which is “free” in two ways. First, it is dependency-free because it requires neither costly pre-training [ 4,11,22,37] nor extra supervi- sory signals [ 6,29]. Second, it is overhead-free as it requires no additional training-time rendering for patch-based regu- larization [ 8,11,22]. We consider FreeNeRF a simple baseline (with mini- mal modiﬁcations to a plain NeRF) in the few-shot neural rendering problem, although it already outperforms exist- ing state-of-the-art methods on multiple datasets, including Blender, DTU, and LLFF, at almost no additional computa- tion cost. Our contributions can be summarized as follows: •We reveal the link between the failure of few-shot neu- ral rendering and the frequency of positional encoding, which is further veriﬁed by an empirical study and ad- dressed by our proposed method. To our knowledge, our method is the ﬁrst attempt to address few-shot neural ren- dering from a frequency perspective. •We identify another common failure pattern in learning NeRF from sparse inputs and alleviate it with a new oc- clusion regularizer. This regularizer effectively improves performance and generalizes across datasets. •Combined, we introduce a simple baseline, FreeNeRF,that can be implemented with a few lines of code mod- iﬁcation while outperforming previous state-of-the-art methods. Our method is dependency-free and overhead- free, making it a practical and efﬁcient solution to this problem. We hope the observations and discussions in this paper will motivate people to rethink the fundamental role of fre- quency in NeRF’s positional encoding.
Yoshimura_Rawgment_Noise-Accounted_RAW_Augmentation_Enables_Recognition_in_a_Wide_Variety_CVPR_2023	Abstract Image recognition models that work in challenging en- vironments (e.g., extremely dark, blurry, or high dynamic range conditions) must be useful. However, creating train- ing datasets for such environments is expensive and hard due to the difficulties of data collection and annotation. It is desirable if we could get a robust model without the need for hard-to-obtain datasets. One simple approach is to ap- ply data augmentation such as color jitter and blur to stan- dard RGB (sRGB) images in simple scenes. Unfortunately, this approach struggles to yield realistic images in terms of pixel intensity and noise distribution due to not consider- ing the non-linearity of Image Signal Processors (ISPs) and noise characteristics of image sensors. Instead, we propose a noise-accounted RAW image augmentation method. In essence, color jitter and blur augmentation are applied to a RAW image before applying non-linear ISP , resulting in re- alistic intensity. Furthermore, we introduce a noise amount alignment method that calibrates the domain gap in the noise property caused by the augmentation. We show that our proposed noise-accounted RAW augmentation method doubles the image recognition accuracy in challenging en- vironments only with simple training data.	1. Introduction Although image recognition has been actively studied, its performance in challenging environments still needs im- provement [15]. Sensitive applications such as mobility sensing and head-mounted wearables need to be robust to various kinds of difficulties, including low light, high dy- namic range (HDR) illuminance, motion blur, and cam- era shake. One possible solution is to use image enhance- ment and restoration methods. A lot of DNN-based low- light image enhancement [12, 20, 29, 30, 46, 54], denois- ing [32, 43, 53], and deblurring [43, 48, 52] methods are proposed to improve the pre-captured sRGB image qual- ity. While they are useful for improving pre-captured image unrealistic intensity ISP(Contrast) augmentation (Contrast) augmentation ISP unrealistic noise (a) Usual training pipeline RAW image luminance+noise realistic intensity realistic noise (b) Noise-accounted RAW augmentation pipeline RAW image luminance+noise Noise correction Figure 1. The concept of the proposed noise-accounted RAW aug- mentation. Conventional augmentation (a) is applied to the output of an ISP; due to the nonlinear operations in the ISP, it produces images that cannot be captured at any ambient light intensities. Instead, ours (b) applies augmentation before an ISP. It generates realistic pixel intensity distribution that can be captured when the light intensity is different. Moreover, the noise amount is also cor- rected to minimize the domain gap between real and augmented ones. quality, a recent work [15] shows that using them as prepro- cessing for image recognition models has limited accuracy gains since they already lost some information, and restor- ing the lost information is difficult. Another possible solution is to prepare a dataset for dif- ficult environments [3, 33]. However, these datasets only cover one or a few difficulties, and creating datasets in various environments is too expensive. Especially, man- ual annotation of challenging scenes is difficult and time- consuming. For example, we can see almost nothing in usual sRGB images under extremely low-light environ- ments due to heavy noise. In addition, some regions in HDR scenes suffer from halation or blocked-up shadows because the 8-bit range of usual sRGB images cannot fully preserve the real world, which is 0.000001 [cd/m2]under starlight and 1.6 billion [cd/m2]under direct sunlight [37]. Heavy 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. 14007 motion blur and camera shake also make annotation diffi- cult. Some works capture paired short-exposure and long- exposure images, and the clean long-exposure images are used for annotation or ground truth [15–17, 19]. The limi- tation is that the target scene needs to be motionless if the pairs are taken sequentially with one camera [15], and posi- tional calibration is required if the pairs are taken with syn- chronized cameras [16,17]. Some works use a beam splitter to capture challenging images and their references without calibration [19, 45]. However, they are difficult to apply in dark scenes. Moreover, HDR images cannot be taken in the same way because some regions become overexposed or underexposed in both cameras. To this end, we aim to train image recognition mod- els that work in various environments only using a train- ing dataset in simple environments like bright, low dynamic range, and blurless. In this case, image augmentation or do- main adaptation is important to overcome the domain gap between easy training data and difficult test data. However, we believe usual augmentations on sRGB space are inef- fective because it does not take into account the nonlinear mapping of ISPs. In particular, tone mapping significantly changes the RAW image values, which are roughly propor- tional to physical brightness [44]. Contrast, brightness, and hue augmentation on sRGB space result in unrealistic im- ages that cannot be captured under any ambient light inten- sity as shown in Fig. 1(a). In contrast, we propose augmen- tation on RAW images. In other words, augmentation is applied before ISPs to diminish the domain shift as shown in Fig. 1(b). Oher possible sources of the domain gap are differences in noise amount and noise distribution. To tackle these problems, we propose a method to align both light inten- sity and noise domains. Recent works show that adding physics-based realistic noise improves the performance of DNN-based denoisers [2, 44, 47, 50] and dark image recog- nition [4, 15]. Although their proposed sensor noise mod- elings are accurate, they assume that the original bright im- ages are noise free. In contrast, we propose to modify the noise amount after contrast, brightness, and hue conversion considering the noise amount in the original images. It en- ables a more accurate alignment of the noise domain. Even bright images may have dark areas due to shadows or object colors, and their prior noise cannot be ignored. Another merit of our method is that it is possible to take dark im- ages that already contain a lot of noise as input. In addition to noise alignment after color jitter augmentation, we show the importance of noise alignment after blur augmentation, which is proposed for the first time in this paper. Our contributions are as follows: • It is the first work to emphasize the importance of aug- mentation before ISP for image recognition to the best of our knowledge.• Noise amount alignment method is proposed to reduce the noise domain gap after RAW image augmentation. In contrast to previous works, our proposed method takes into account prior noise in the input image. It en- ables more accurate alignment and use of any strength of augmentation and even already noisy input. • We show qualitative analysis for the validity of our sensor noise modeling and corresponding noise- accounted augmentation. We prove that our proposed noise-accounted RAW augmentation has the edge over the previous methods.
You_Castling-ViT_Compressing_Self-Attention_via_Switching_Towards_Linear-Angular_Attention_at_Vision_CVPR_2023	Abstract Vision Transformers (ViTs) have shown impressive per- formance but still require a high computation cost as com- pared to convolutional neural networks (CNNs), one rea- son is that ViTs’ attention measures global similarities and thus has a quadratic complexity with the number of in- put tokens. Existing efficient ViTs adopt local attention or linear attention, which sacrifice ViTs’ capabilities of cap- turing either global or local context. In this work, we ask an important research question: Can ViTs learn both global and local context while being more efficient during inference? To this end, we propose a framework called Castling-ViT , which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attention during inference. Our Castling-ViT leverages angular ker- nels to measure the similarities between queries and keys via spectral angles. And we further simplify it with two tech- niques: (1) a novel linear-angular attention mechanism: we decompose the angular kernels into linear terms and high-order residuals, and only keep the linear terms; and (2) we adopt two parameterized modules to approximate high-order residuals: a depthwise convolution and an aux- iliary masked softmax attention to help learn global and lo- cal information, where the masks for softmax attention are regularized to gradually become zeros and thus incur no overhead during inference. Extensive experiments validate the effectiveness of our Castling-ViT, e.g., achieving up to a 1.8% higher accuracy or 40% MACs reduction on classifi- cation and 1.2higher mAP on detection under comparable FLOPs, as compared to ViTs with vanilla softmax-based at- tentions. Project page is available at here.	1. Introduction Vision Transformers (ViTs) have made significant progress in image classification, object detection, and many *Equal contribution. †Work done while interning at Meta Research. 0.02.55.07.510.012.515.017.5 MACs (G) ×109758085Top-1 Acc. (%) Image Classification on ImageNet LeViT EfficientNet MobileNetV2 DeiT RegNetYSwin CSwin MViTv2 Autoformer Castling-ViT(Ours) 1.02.03.04.05.06.0 MACs (G) ×109253035mAP (%) Object Detection on COCO YOLOv5 YOLOX MobileDet-DSP EfficientDet FBNetv5 Castling-ViT(Ours)Figure 1. Castling-ViT over SOTA baselines on (1) ImageNet [18] image classification and (2) COCO [36] object detection. other applications. It is well recognized that the supe- rior performance achieved by ViTs is largely attributed to their self-attention modules that can better capture global context [20, 57, 64]. Nevertheless, ViTs’ powerful self- attention module comes at the cost of quadratic complex- ity with the number of input tokens, causing a major effi- ciency bottleneck to ViTs’ achievable runtime (i.e., infer- ence latency) [3, 8, 32, 58, 66, 69, 76]. To mitigate this is- sue, linear attention designs have been developed to alle- viate the vanilla ViT attention’s quadratic complexity. In particular, existing efforts can be categorized into two clus- ters: (1) ViTs with local attention by restricting the atten- tion window size [38, 53], sharing the attention queries [2], or representing the attention queries/keys with low rank ma- trices [58]; and (2) ViTs with kernel-based linear attention, which approximate the non-linearity softmax function by decomposing it into separate kernel embeddings. This en- ables a change in the matrix computation order for a re- duced computational complexity [5,6,14,29,37,39,43,66]. Despite their promise in alleviating ViTs’ complexity 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. 14431 and thus inference runtime, both the local and linear at- tention compromise ViTs’ performance due to the lack of capabilities to capture global or local context. To marry the best of both worlds, we advocate training ViTs with both (1) efficient but less powerful linear attention, i.e., without the high-order residuals in angular kernel expansion, and (2) powerful yet costly softmax-based masked attention. The latter helps approximate high-order residuals at the early training stage while being dropped during inference, based on an assumption that the remaining networks can gradu- ally learn the high-order components at the later training stage [67]. This concept resembles the “castling” move in chess when two pieces are moved at once. While it sounds promising, there are still two challenges to achieve this. First , existing linear attention modules still underperform their vanilla softmax-based counterparts. Therefore, a bet- ter linear attention is crucial for the final performance. We find that angular kernels perform equally as softmax-based attentions in terms of similarity measurements. While they still suffer from a quadratic complexity, they can be divided into linear terms and high-order residuals. The challenge is how to construct ViTs with only the linear terms .Second , doing so would require that the trained ViTs merely rely on the linear terms towards the end of training, which would call for an approximation of the above high-order residuals. The challenge is that how we can resort to costly but pow- erful modules to approximate high-order residuals during training but does not incur extra inference cost . In this work, we develop techniques to tackle those chal- lenges, and make the following contributions: • We propose a framework called Castling-ViT , which trains ViTs using both linear-angular attention and masked softmax-based quadratic attention, but then switches to having only linear-angular attentions dur- ing ViT inference to save computational costs. • We develop a new linear-angular attention leveraging angular kernels to close the accuracy gap between lin- ear attention and softmax-based attention. It expands angular kernels where linear terms are kept while com- plex high-order residuals are approximated. • We use two parameterized modules to approximate the high-order residuals above: a depthwise convolu- tion and an auxiliary masked softmax-based attention, where the latter’s attention masks are regularized to gradually become zeros to avoid inference overhead. • We conduct extensive experiments to validate the ef- fectiveness of the proposed Castling-ViT. Results on classification, detection, and segmentation tasks con- sistently demonstrate its superior performance ( ↑1.8% top-1 accuracy or ↑1.2 mAP) or efficiency (40% MACs savings) over state-of-the-art (SOTA) CNNs and ViTs.
Zeng_ConZIC_Controllable_Zero-Shot_Image_Captioning_by_Sampling-Based_Polishing_CVPR_2023	Abstract Zero-shot capability has been considered as a new rev- olution of deep learning, letting machines work on tasks without curated training data. As a good start and the only existing outcome of zero-shot image captioning (IC), ZeroCap abandons supervised training and sequentially searches every word in the caption using the knowledge of large-scale pre-trained models. Though effective, its autoregressive generation and gradient-directed searching mechanism limit the diversity of captions and inference speed, respectively. Moreover, ZeroCap does not consider the controllability issue of zero-shot IC. To move forward, we propose a framework for Controllable Zero-shot IC, named ConZIC . The core of ConZIC is a novel sampling- based non-autoregressive language model named Gibbs- BERT, which can generate and continuously polish every word. Extensive quantitative and qualitative results demon- strate the superior performance of our proposed ConZIC for both zero-shot IC and controllable zero-shot IC. Espe- cially, ConZIC achieves about 5 ×faster generation speed than ZeroCap, and about 1.5 ×higher diversity scores, with accurate generation given different control signals. Our code is available at https://github.com/joeyz0z/ConZIC.	1. Introduction Image captioning (IC) is a visual-language task, which targets at automatically describing an image by generating a coherent sentence. By performing supervised learning on human-annotated datasets, such as MS-COCO [43], many methods [22, 33, 49, 50] have achieved impressive evalu- ation scores on metrics like BLEU [52], METEOR [7], CIDERr [66], and SPICE [3]. However, these methods still lag behind human capability of zero-shot IC. Specifically, those supervised methods extremely rely on well-designed image-captions pairs. However, it is likely *Equal contribution. †Corresponding authors GRIT: A drawing of graffiti on a wall. ViTCap : A picture of a sheep and a cow. CLIPCap : A black and white photo of a group of cows. ZeroCap :Image of a cows drawing. Ours: Two cows face each other on a pasture with various flowers in sequence . GRIT: A woman sitting on a fountain in the sea. ViTCap :A painting of a woman laying on a bed. CLIPCap: A painting of a woman in on a surfboard. ZeroCap :Image of a girl sleeping in the sea . Ours: A painting of the princess submerged in delicate poses with water background . Positive: 1.A very cute cheerful white bird accompanies a happy tiny elephant. 2.A cute little white duck enjoys amazed chatting with an elephant. 3.A white henand an extremely small beautiful elephant play happily on ground. 4.A h ealthy elephant enthusiastically admires anawesome adorable southern bird. 5.A g orgeous elephant walks beside a white goose with miniature smiling . N egative: 1.A b adly elephant stars at a scared small bird. 2. A scared little bird is afraid that the vicious elephant would eat it. 3. Image of a sad libelous chicken moping alongside a small lonely elephant is shown. 4.A stra y worried white duck meets a stealthy hungry elephant. 5.A b rown solitary elephant roams with an lonely white sparrow nearby.+ Positive Negativeor(a) Examples of zero-shot image captioning. GRIT: A drawing of graffiti on a wall. ViTCap : A picture of a sheep and a cow. CLIPCap : A black and white photo of a group of cows. ZeroCap :Image of a cows drawing. Ours: Two cowsface each other on a pasture with various flowers in sequence . GRIT: A woman sitting on a fountain in the sea. ViTCap :A painting of a woman laying on a bed. CLIPCap: A painting of a woman in on a surfboard. ZeroCap :Image of a girl sleeping in the sea. Ours: A painting of the princess submerged in delicate poses with water background . Positive: 1. A very cute cheerful white bird accompanies a happy tiny elephant. 2. A cute little white duck enjoys amazed chatting with an elephant. 3. A white henand an extremely small beautiful elephant play happily on ground. 4. A healthy elephant enthusiastically admires anawesome adorable southern bird. 5. A gorgeous elephant walks beside a white goose with miniature smiling . Negative:1. A badly elephant stars at a scared small bird. 2. A scared little bird is afraid that the vicious elephant would eat it. 3. Image of a sad libelous chicken moping alongside a small lonely elephant is shown. 4. A stray worried white duck meets a stealthy hungry elephant. 5. A brown solitary elephant roams with an lonely white sparrow nearby.+ Positive Negativeor (b) Diversity of ConZIC. Figure 1. The highlights of our proposed method. (a) shows two examples of zero-shot image captioning on several SOTA meth- ods. Specifically, GRIT [50] and ViTCAP [22] are two supervised methods without pre-trained models. ClipCap [49] is a super- vised method using pre-trained CLIP. GRIT, ViTCAP, and CLIP- Cap are firstly trained on MSCOCO and then do testing. ZeroCap [65] is the zero-shot method without any training. (b) shows the diversity of our proposed ConZIC, which manifests two aspects: semantic (diverse words: different colors denoting different parts- of-speech) and syntactic (diverse sentence patterns). impossible to construct a large enough dataset, including paired images and high-quality captions covering various styles/contents. As a result, it is challenging for the machine to caption images that are outliers with respect to the train- ing distribution, which is common in real applications (see 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. 23465 examples in Fig. 1a). On the contrary, humans can perform IC without any specific training, i.e., realizing zero-shot IC. Because humans can integrate what they see, i.e., the image, and what they know, i.e., the knowledge. Recently, large-scale pretraining models have shown a strong capability of learning knowledge from super-large- scale data, showing great potential in various downstream tasks [10,27,54,57,63]. Equipped with the visual-language knowledge learned by CLIP [57] and linguistic knowledge from GPT-2 [58], ZeroCap [65] is the first and the only zero- shot IC method, which proposes a searching-based strategy and is free of training on extra supervised data. Specifically, ZeroCap searches the caption words one by one and from left to right, guided by CLIP-induced score for image-text matching and GPT-2 word distribution for caption fluency. ZeroCap is a good start and inspires us to explore how to search for the optimal caption in a better way. i) More flexible. ZeroCap utilizes GPT-2 to perform left- to-right autoregressive generation. Once a word is fixed, there is no chance to modify it when we move to the next position. In other words, such generation order is not flexi- ble enough to consider the full context information. ii) More efficient. The searching at every position is real- ized by iteratively updating the parameters of GPT-2, which is time-consuming, as shown in Fig. 3c. iii) More diverse. IC is an open problem. Given an image, different persons may have different visual atten- tions [14] and language describing styles [24, 47, 73], thus resulting in diverse descriptions. ZeroCap employs beam search to generate several candidate sentences, which, how- ever, have similar syntactic patterns (see Appendix D). iv) More controllable. To endow captioning models with human-like controllability, e.g., sentiment, personality, a re- cent surge of efforts [12,19,24,47] resort to introducing ex- tra control signals as constraints of the generated captions, called Controllable IC. However, controllable zero-shot IC has not been explored yet. Bearing all these four-aspect concerns in mind, we propose a novel framework for controllable zero-shot IC, named ConZIC, as shown in Fig. 2. Specifically, after ana- lyzing the relationship between Gibbs sampling and masked language models (MLMs, currently we use BERT) [11, 20, 70], we firstly develop a new language model (LM) called Gibbs-BERT to realize the zero-shot IC by sampling-based search. Compared with autoregressive models, Gibbs- BERT has more a flexible generation order, bringing the self-correct capability by bidirectional attention with faster and more diverse generations. After integrating Gibbs- BERT with the CLIP that is used to evaluate the similar- ity between image and text, our proposed framework can perform zero-shot IC. By further introducing a task-specific discriminator for control signal into our framework, our proposed framework can perform controllable zero-shot IC.The main contributions of this paper are: • We propose to solve the controllable zero-shot IC task in a polishing way. By combining Gibbs sampling with a MLM, we can randomly initialize the caption and then polish every word based on the full context (bidirectional information) in the caption. • ConZIC is free of parameter updates, achieving about 5 × faster generation speed than the SOTA method, ZeroCap. • Equipped with Gibbs-BERT, ConZIC can perform flexi- ble searching, thus generating sentences with higher di- versity, as shown in Table. 1. • To the best of our knowledge, ConZIC is the first control- lable zero-shot IC method. Four classes of controllable signals, including length, infilling, styles, and parts-of- speech, are evaluated in our experiments.
Zhang_Revisiting_Rotation_Averaging_Uncertainties_and_Robust_Losses_CVPR_2023	Abstract In this paper, we revisit the rotation averaging problem applied in global Structure-from-Motion pipelines. We ar- gue that the main problem of current methods is the mini- mized cost function that is only weakly connected with the input data via the estimated epipolar geometries. We pro- pose to better model the underlying noise distributions by directly propagating the uncertainty from the point corre- spondences into the rotation averaging. Such uncertain- ties are obtained for free by considering the Jacobians of two-view refinements. Moreover, we explore integrat- ing a variant of the MAGSAC loss into the rotation av- eraging problem, instead of using classical robust losses employed in current frameworks. The proposed method leads to results superior to baselines, in terms of accu- racy, on large-scale public benchmarks. The code is public. https://github.com/zhangganlin/GlobalSfMpy	1. Introduction Building large 3D reconstructions from unordered im- age collections is an essential component in any system that relies on crowd-sourced mapping. The current paradigm is to perform this reconstruction via Structure-from-Motion [29] which jointly estimates the camera parameters and the scene geometry represented with a 3D point cloud. Meth-ods for Structure-from-Motion can generally be categorized into two classes; Incremental methods [29, 32, 33, 38] that sequentially grows a seed reconstruction by alternating tri- angulation and registering new images, and Global meth- ods[8, 22, 24, 27] which first estimate pairwise geometries and then aggregate them in a bottom-up approach. Histor- ically, incremental methods are more robust and accurate, but the need for frequent bundle adjustment [35] comes with significant computational cost which limits their scalabil- ity. In contrast, global (or non-sequential) methods require much lower computational effort and can in principle scale to larger image collections. However, in practice, current methods are held back by the lack of accuracy and have not enjoyed the same level of success as incremental methods. Global methods work by first estimating a set of pair- wise epipolar geometries between co-visible images. Next, viarotation averaging , a set of globally consistent rotations are estimated by ensuring they agree with the pairwise rel- ative rotations. Once the rotations are known, the camera positions and 3D structure are estimated, and refined jointly in a single final bundle adjustment. Rotation averaging has a long history in computer vision (seee.g. [16,22] for early works) and is a well-studied prob- lem. Most methods formulate it as an optimization problem, finding the rotation assignment that minimizes some energy. A common choice is the chordal distance , measuring 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. 17215 discrepancy in the rotation matrices in the L2-sense min {Ri}N i=1NX i=1X j∈N(i)∥ˆRijRi−Rj∥2 F (1) where ˆRijis the relative rotation estimated between image iandj. There are also other choices such as using angle- axis [34] or quaternion [16] as rotation representation, or optimising over a Lie algebra [17], however the overall idea (measuring some consistency with the relative estimates) remains the same. Many works have focused on the opti- mization problem itself, both theoretically [11, 36] and by providing new algorithms [9], but did not consider whether the cost itself is suitable for the task. In (1), each relative ro- tation measurement is given the same weight. However in practice, the quality of the epipolar geometries varies sig- nificantly. Figure 1 shows two images with wildly different uncertainties (and errors) in the rotation estimate. To ad- dress this problem, there is a line of work [6, 18, 31] which augment the cost in (1) with robust loss functions that give a lower weight to large residuals. However, the same loss function is generally applied to each residual, independent of the measurement uncertainty. In this paper we revisit the rotation averaging problem. We argue that the main problem in current methods is that the cost functions that are minimized are only weakly con- nected with the input data via the estimated epipolar ge- ometries. We propose to better model the underlying noise distributions (coming from the keypoint detection noise and spatial distribution) by directly propagating the uncertainty from the point correspondences into the rotation averaging problem, as shown in Figure 2. While the idea itself is sim- ple, we show that this allows us to get significantly more accurate estimates of the absolute rotations; reducing the gap between incremental and global methods. Note that the uncertainties we leverage are essentially obtained for free by considering the Jacobians of the two-view refinement. As a second contribution, we explore integrating a vari- ant of the MAGSAC [3] loss into the rotation averaging problem, instead of using the classical robust losses em- ployed in current frameworks. MAGSAC [3] was originally proposed as a threshold-free estimator for two-view epipo- lar geometry, where the idea is to marginalize over an in- terval of acceptable thresholds, i.e., noise range. We show that this fits well into the context of rotation-averaging, as it is not obvious how to set the threshold for deciding on in- lier/outlier relative rotation measurements, especially in the uncertainty-reweighted cost that we propose.
Yang_Learning_Event_Guided_High_Dynamic_Range_Video_Reconstruction_CVPR_2023	Abstract Limited by the trade-off between frame rate and exposure time when capturing moving scenes with conventional cam- eras, frame based HDR video reconstruction suffers fromscene-dependent exposure ratio balancing and ghosting ar-tifacts. Event cameras provide an alternative visual repre-sentation with a much higher dynamic range and temporalresolution free from the above issues, which could be an effective guidance for HDR imaging from LDR videos. Inthis paper , we propose a multimodal learning framework for event guided HDR video reconstruction. In order tobetter leverage the knowledge of the same scene from thetwo modalities of visual signals, a multimodal representa-tion alignment strategy to learn a shared latent space and a fusion module tailored to complementing two types of sig- nals for different dynamic ranges in different regions areproposed. Temporal correlations are utilized recurrentlyto suppress the ﬂickering effects in the reconstructed HDRvideo. The proposed HDRev-Net demonstrates state-of-the- art performance quantitatively and qualitatively for both synthetic and real-world data.	1. Introduction The dynamic range of the real world usually exceeds what a conventional camera and 8-bit image can record by a large margin. High dynamic range (HDR) imaging, whichexpands the luminance range limited by low dynamic range(LDR) images or videos, is a broadly used technique with extensive applications in photography/videography, video games, and high-end display. Most HDR imaging methods for conventional cameras rely on capturing and merging multiple snapshots with dif-ferent exposure times [ 9,49], which is challenging for cap- turing videos. There have been enduring efforts for sophisti-cated modiﬁcation on conventional frame based cameras tocapture multi-exposure sequences (nearly) simultaneously,e.g., beam splitting with three or more sensors [ 69,70], tem- porally [ 7,30,32] or spatially [ 1,8,22,28,53,54] varying exposure. Nevertheless, their abilities for HDR video re-construction are limited by the trade-off between a higherframe rate (for a smooth viewing experience) and a higherdynamic range (for capturing details in dark regions with ∗Corresponding author Project page: https://yixinyang-00.github.io/HDRev/ 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. 13924 prolonged exposure time). Moreover, the optimal expo- sure ratio between LDR frame sequences with different ex-posure settings is scene-dependent and temporally-varying,whose balancing is difﬁcult for diverse scenes captured invideos. Even worse, moving objects or camera shakingduring video capture can lead to ghosting effects in framesgenerated by long exposure shots. An HDR video couldalso be hallucinated from LDR inputs in a frame-by-framemanner by leveraging prior knowledge of tone-mapping op-erators [ 61] or data modeling powers of deep learning [ 11]. However, due to the highly ill-posed nature of the hallucina-tion process, it inevitably leads to severe ﬂickering effects. In recent years, the event camera [ 16] has drawn increas- ing attention of researchers, due to its advantages over con-ventional frame based ones in sensing fast motions and ex-tended dynamic ranges ( e.g.,120 dB for DA VIS346). Unlike using multi-exposure frames, events recorded along with an LDR video encode HDR irradiance changes without sacri-ﬁcing the frame rate/exposure time of the LDR video, whichavoids ghosting artifacts as well, and is very promising asguidance for HDR video reconstruction. However, integrating events with LDR video for HDR video reconstruction is challenging due to inconsistency be-tween events and frames in three aspects: 1) Modality mis- alignment : Frames and events are completely different rep- resentations of visual information , and “fusing” them by ﬁrst translating events into intensity values [ 60,80] like [ 24] often includes artifacts from solving the ill-posed event in-tegration problem. 2) Dynamic range gap : Performing im- age/video reconstruction under the guidance of events [ 75], i.e., doubly integrating events as intensity changes within the exposure time [ 56,57], ignores the dynamic range clip- ping in the capturing process of LDR frames, which leads touncertainties in under/over-exposed regions .3 ) Tex- ture mismatching : Regions with smooth textures and slow motion hardly produce effective event observations, whichresults in inconsistent textures among consecutive eventstacks and ﬂickering effects in the reconstructed videos. We propose HDRev-Net , a multimodal learning frame- work for event guided HDR video reconstruction to tacklethe challenges by the following strategies: 1) To achievemultimodal representation alignment for the two modalities of the same scene, we propose a learning strategy to pro-gressively project them onto a shared representation space . 2) To reliably complement information from the two modal- ities in over/under-exposed regions, the representations pro-duced by the two modality-speciﬁc encoders are fused foran expressive joint representation using a conﬁdence guided multimodal fusion module . 3) To effectively suppress the ﬂickering effects , we utilize the temporal redundant infor- mation between consecutive frames and events via the pro-posed recurrent convolutional encoders . As shown in Fig. 1, HDRev-Net can successfully fuseLDR frames and events to obtain HDR frames with more details and less ﬂickering effects. It demonstrates state-of-the-art HDR video reconstruction performance on both syn-thetic and real data by making the following contributions: • We design a multimodal alignment strategy to bridge the gap between events and frames by aligning theirrepresentation in a shared latent space. • We develop a conﬁdence guided fusion module to complement HDR information from events and ﬁnerdetails from well-exposed regions in LDR frames. • We utilize the temporal correlation from consecutive events and LDR frames in a recurrent fashion to alle-viate the ﬂickering effects for recovered HDR videos.
Yang_Diffusion_Probabilistic_Model_Made_Slim_CVPR_2023	Abstract Despite the recent visually-pleasing results achieved, the massive computational cost has been a long-standing flaw for diffusion probabilistic models (DPMs), which, in turn, greatly limits their applications on resource-limited plat- forms. Prior methods towards efficient DPM, however, have largely focused on accelerating the testing yet overlooked their huge complexity and sizes. In this paper, we make a dedicated attempt to lighten DPM while striving to pre- serve its favourable performance. We start by training a small-sized latent diffusion model (LDM) from scratch, but observe a significant fidelity drop in the synthetic images. Through a thorough assessment, we find that DPM is in- trinsically biased against high-frequency generation, and learns to recover different frequency components at differ- ent time-steps. These properties make compact networks unable to represent frequency dynamics with accurate high- frequency estimation. Towards this end, we introduce a customized design for slim DPM, which we term as Spec- tral Diffusion ( SD), for light-weight image synthesis. SD incorporates wavelet gating in its architecture to enable frequency dynamic feature extraction at every reverse step, and conducts spectrum-aware distillation to promote high- frequency recovery by inverse weighting the objective based on spectrum magnitude. Experimental results demonstrate that,SDachieves 8-18 ×computational complexity reduc- tion as compared to the latent diffusion models on a series of conditional and unconditional image generation tasks while retaining competitive image fidelity.	1. Introduction Diffusion Probabilistic Models (DPMs) [18,57,59] have recently emerged as a powerful tool for generative mod- eling, and have demonstrated impressive results in image synthesis [8, 45, 48], video generation [17, 20, 77] and 3D editing [43]. Nevertheless, the gratifying results come with a price: DPMs suffer from massive model sizes. In fact, Corresponding author DPMLiteDPM274.1M#Param96.1GMACs22.4M#Param7.9GMACsSpectralDPM (Ours)21.1M#Param6.7G MACs Smaller #Params(M)BetterFID ① ②(Ours)Figure 1. (1) Visualization of the frequency gap among generated images with the DPM [48], Lite DPM and our SDon FFHQ [27] dataset. Lite-DPM is unable to recover fine-grained textures, while SDcan produce realistic patterns. (2) Model size, Multiply-Add cumulation (MACs) and FID score on ImageNet [7]. Our model achieves compelling visual quality with minimal computational cost.∗indicates our re-implemented version. state-of-the-art DPMs requires billions of parameters, with hundreds or even thousands of inference steps per image. For example, DALL·E 2[45], which is composed of 4 sep- arate diffusion models, requires 5.5B parameters and 356 sampling steps in total. such an enormous model size, in turn, makes DPMs extremely cumbersome to be employed in resource-limited platforms. However, existing efforts towards efficient DPMs have focused on model acceleration, but largely overlooked light- ening of the model. For example, the approaches of [1, 32, 37, 38, 40, 52, 56] strive for faster sampling, while those of [13, 19, 48, 62] rely on reducing the input size. Admit- tedly, all of these methods give rise to shortened training or inference time, yet still, the large sizes prevent them from many real-world application scenarios. In this paper, we make a dedicated efforts towards build- ing compact DPMs. To start with, we train a lite version of the popular latent diffusion model (LDM) [48] by re- 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. 22552 ducing the channel size. We show the image generated by the original and and lite DPM in Figure 1. While the lite LDM sketches the overall structure of the faces, the high- frequency components, such as the skin and hair textures, are unfortunately poorly recovered. This phenomenon can be in fact revealed by the Discrete Fourier Transform (DFT) coefficient shown on the right column, indicating that the conventional design for DPMs leads to high-frequency de- ficiency when the model is made slim. We then take an in-depth analysis on the DPMs through the lens of frequency, which results in two key obser- vations. (1) Frequency Evolution. Under mild assump- tions, we mathematically prove that DPMs learn different functionalities at different stages of the denoising process. Specifically, we show that the optimal denoiser in fact boils down to a cascade of wiener filters [66] with growing band- widths. After recovering the low-frequency components, high-frequency features are added gradually in the later de- noising stages. This evolution property, as a consequence, small DPMs fails to learn dynamic bandwidths with limited parameters. (2) Frequency Bias. DPM is biased towards dominant frequency components of the data distribution. It is most obvious when the noise amplitude is small, lead- ing to inaccurate noise prediction at the end of the reverse process. As such, small DPMs struggle to recover the high- frequency band and image details. Motivated by these observations, we propose a novel Spectral Diffusion ( SD) model, tailored for light-weight im- age synthesis. Our core idea is to introduce the frequency dynamics and priors into the architecture design and train- ing objective of the small DPM, so as to explicitly preserve the high-frequency details. The proposed solution consists of two parts, each accounting for one aforementioned obser- vations. For the frequency evolution, we propose a wavelet gating operation, which enables the network to dynamically adapt to the spectrum response at different time-steps. In the upsample and downsample stage, the input feature is first decomposed through wavelet transforms and the coef- ficients are re-weighted through a learnable gating function. It significantly lowers the parameter requirements to repre- sent the frequency evolution in the reverse process. To compensate for the frequency bias for small DPMs, we distill high-frequency knowledge from a teacher DPM to a compact network. This is achieved by inversely weight- ing the distillation loss based on the magnitudes of the fre- quency spectrum. In particular, we give more weight to the frequency bands with small magnitudes, which strength- ens the recovery of high-frequency details for the student model. By integrating both designs seamlessly, we build a slim latent diffusion model, called SD, which largely pre- serves the performance of LDM. Notably, SDinherits the advantages of DPMs, including superior sample diversity, training stability, and tractable parameterization. As shownin Figure 1, our model is 8∼18×times smaller and runs 2∼5×times faster than the original LDM, while achieving competitive image fidelity. The contributions of this study are threefold: 1. This study investigates the task of diffusion model slimming, which remains largely unexplored before. 2. We identify that the key challenge lies in its unrealistic recovery for the high-frequency components. By prob- ing DPMs from a frequency perspective, we show that there exists a spectrum evolution over different denois- ing steps, and the rare frequencies cannot be accurately estimated by small models. 3. We propose SD, a slim DPM that effectively restores imagery textures by enhancing high-frequency genera- tion performance. SDachieves gratifying performance on image generation tasks at a low cost.
Zhang_Modeling_Video_As_Stochastic_Processes_for_Fine-Grained_Video_Representation_Learning_CVPR_2023	Abstract A meaningful video is semantically coherent and changes smoothly. However, most existingne-grained video representation learning methods learn frame-wise features by aligning frames across videos or exploring rel- evance between multiple views, neglecting the inherent dy- namic process of each video. In this paper, we propose to learn video representations by modeling Video as Stochas- tic Processes (VSP) via a novel process-based contrastive learning framework, which aims to discriminate between video processes and simultaneously capture the temporal dynamics in the processes. Specically, we enforce the em- beddings of the frame sequence of interest to approximate a goal-oriented stochastic process, i.e., Brownian bridge, in the latent space via a process-based contrastive loss. To construct the Brownian bridge, we adapt specialized sam- pling strategies under different annotations for both self- supervised and weakly-supervised learning. Experimental results on four datasets show that VSP stands as a state-of- the-art method for various video understanding tasks, in- cluding phase progression, phase classication, and frame retrieval. Code is available at https://github.com/ hengRUC/VSP .	1. Introduction Fine-grained video representation learning [ 11] is one of the fundamental problems in computer vision, which has great practical value in various real-world applications such as action phase classication [ 11,40], phase boundary de- tection [ 26], and video object segmentation [ 7,9,22,33]. The way to model videos, especially the temporal dynam- ics, is the core problem of video representation learning and is highly relevant to available data annotations. Pioneer *Equal contributions. †Corresponding author.works [ 4,29] directly model video as 3D data where tempo- ral is one dimension, and they require large-scale human- generated annotations for representation learning. How- ever, it is labor-intensive and time-consuming to collect those annotations. Besides, human-generated annotations hinder domain generalization to multiple downstream tasks. To alleviate the requirement on labeled data, some re- cent works [ 11–13] model the video alignment (Figure 1(a)) across the temporal dimensions by the cycle-consistency loss [ 11] or temporal alignment loss [ 13]. Their basic as- sumption is that two videos of the same action can be aligned over temporal ordering in the embedding space, and the latent correspondences across sequence pairs can be re- garded as a supervisory signal. However, these methods es- sentially work in a weakly-supervised manner that requires video-level annotations to construct video pairs, impeding their application in the real-world scene where the semantic labels are absent. As an alternative, self-supervised video representation learning [ 5,26] explores the view relevance (Figure 1(b)) between two augmented views of one video. By model- ing video as a sequence along the temporal dimensions, they elaborately construct two views through a series of spatio-temporal data augmentations. The training objec- tive is to encourage the relevance of two augmented views to conform to their assumptions,e.g., spatio-temporal con- trast [ 26] or similarity distribution [ 5]. However, those methods are sensitive to complex hand-craft view augmen- tation thus suffering from sub-optimal performance. As crucial and intrinsic cues, the dynamics of videos impose temporal correlations among successive frames. Therefore, the evolution process of the correspondingne- grained representations should follow coherent constraints, which can be modeled as a stochastic process. To this end, we propose a new perspective that considers Video as Stochastic Processes (VSP) to explicitly capture the tem- poral dynamics of videos by exploring process agreement 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. 2225 Figure 1. The evolution ofne-grained video representation learning.(a)Video alignment(e.g., TCC [ 11], LA V [ 13]) enforces two videos from the same action aligned temporally.(b)View relevance(e.g., TCN [ 26], CARL [ 5]) enforces the relevance of two augmented views conform to specic assumptions.(c)The proposedprocess agreementmodels video as stochastic process and enforces an arbitrary frame to agree with a time-variant Gaussian distribution conditioned on the start and end frames. (Figure 1(c)). The basic assumption is that a video phase is coherent and smoothly changes from the start to the end, which is essentially a goal-oriented stochastic process that neighboring points are similar to each other and their co- herent changes abide by the direction of the endpoint. For example, a baseball pitching video demonstrates a series of continuous movements as the ballies out of the hand. Specically, we model a video phase as a goal-oriented stochastic process,i.e., the Brownian bridge [ 3,25], where the frame representations in the latent embedding space are expected to be smooth temporal dynamics conditioned on thexed start and end frames. With this intuitive assump- tion, an arbitrary frame is enforced to be like a noisy lin- ear interpolation between the start and end frames with un- certainty in a latent space,i.e., agree with a time-variant Gaussian distribution. By modeling video as stochastic pro- cesses, the proposed method captures the dynamics of each action and establishes dependencies between video frames as well as the semantic consistency of the whole video. Compared with video alignment which assumes pairing videos can be temporally aligned or view relevance which assumes two augmented views are relevant, VSP only re- quests process agreement that assumes the internal frames agree with the start and end frames, discarding the expen- sive annotated video pairs or hand-crafted view pairs. The implementation of VSP follows a process-based contrastive learning framework where each sample is a frame triplet (start, internal, end). The start and end frames of each sample are identied as the beginning and end of the Brownian Bridge. The positive samples are the frame inside the Brownian bridge while the negatives are outside ones. The training objective is to enforce the positive samples conform to the distribution of the target Brownian bridges process while the negative samples stay away from it. Ben- eting from the tunability of the start and end points of the Brownian bridge, VSP is versatile for various annotationsituations. For the most generic situations where human an- notations are not accessible, VSP works in a self-supervised manner by randomly sampling the triplets with an empirical length as Brownian bridges. With the phase-level annota- tions, VSP gains more powerful representations by taking each phase as a Brownian bridge in a weakly-supervised manner. As for the frame-level annotations, the proposed process-based contrastive objective serves as the regulariza- tion term of conventional contrastive losses. The main contributions are summarized as follows: • We propose a novelne-grained video representation learning framework that models Video as Stochastic Processes (VSP) by enforcing frame sequences to con- form to Brownian bridge distributions via a process- based contrastive loss. • We adopt specialized sampling strategies for differ- ent types of annotated data by adjusting the Brownian bridge and therefore acquire favorable video represen- tations in both self-supervised and weakly-supervised manners. • To the best of our knowledge, we are therst to model video as a stochastic process and achieve state-of-the- art performance on variousne-grained video under- standing tasks across four widely-used datasets.
Yan_PlenVDB_Memory_Efficient_VDB-Based_Radiance_Fields_for_Fast_Training_and_CVPR_2023	Abstract In this paper, we present a new representation for neural radiance fields that accelerates both the training and the inference processes with VDB, a hierarchical data struc- ture for sparse volumes. VDB takes both the advantages of sparse and dense volumes for compact data representa- tion and efficient data access, being a promising data struc- ture for NeRF data interpolation and ray marching. Our method, Plenoptic VDB (PlenVDB), directly learns the VDB data structure from a set of posed images by means of a novel training strategy and then uses it for real-time ren- dering. Experimental results demonstrate the effectiveness and the efficiency of our method over previous arts: First, it converges faster in the training process. Second, it delivers a more compact data format for NeRF data presentation. Finally, it renders more efficiently on commodity graphics hardware. Our mobile PlenVDB demo achieves 30+ FPS, 1280×720 resolution on an iPhone12 mobile phone. Check plenvdb.github.io for details. *Work done while the author was an intern at ByteDance. †Corresponding authors.	1. Introduction With the recent advancement of Neural Radiance Fields (NeRF) [17], high-quality Novel View Synthesis from a sparse set of input images can be achieved. It has many ap- plications in multimedia, AR/VR, gaming, etc. On the other hand, new content creation paradigms have been proposed based on NeRF, such as Dreamfusion [25], which enable the possibility of general text-to-3D synthesis. Despite the promising results, one shortage of NeRF is the expensive computation of training and rendering, which prohibits real-time applications and effective scene creation. There have been many efforts to accelerate NeRF rendering by pre-computing and storing the results or intermediate re- sults into a 3D grid. Thus, the computation cost for ren- dering will be reduced by several orders of magnitude. Al- though the methods that exploit 3D dense grid [7,27,30] can achieve real-time rendering and fast training, they usually introduce more storage overhead, which limits the applica- tion on mobile devices. On the other hand, for the methods that utilize 3D sparsity [4, 5, 9, 10, 33], real-time rendering and small storage overhead can be achieved, but the training time is usually getting worse since many of them will first 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. 88 train a Vanilla NeRF or a dense grid, and then convert it to the sparse representation. In this paper, we propose an efficient sparse neural vol- ume representation, which we call Plenoptic VDB (Plen- VDB). The VDB [19] is an industry-proven efficient hierar- chical data structure being used in high-performance anima- tion and simulation for years. We adopt its design principle and use VDB to represent NeRF. VDB takes both advan- tages of sparse and dense volumes for compact data repre- sentation and efficient data access, being a promising data structure for NeRF data interpolation and ray casting. In addition, we propose a novel training approach to directly learn the VDB data without additional conversion steps, so that our model is neat and compact. We show that our model represents high-resolution details of a scene with a lower volume size for fast training and rendering over the state of the arts. Moreover, the trained VDB model can be ex- ported into the NanoVDB [20] format and be used in graph- ics shaders, such as the GLSL fragment shader, that enables rendering a NeRF model on mobile devices in real-time. In our experiment, the mobile PlenVDB achieves 30+ FPS, 1280×720 resolution on an iPhone12 mobile phone. In summary, our approach has two main contributions: • We first use VDB as the sparse volume data structure for NeRF acceleration, and achieve fast rendering even on mobile devices. • We propose a strategy that learns the VDB directly and achieves fast training and occupies small storage.
Zemni_OCTET_Object-Aware_Counterfactual_Explanations_CVPR_2023	Abstract Nowadays, deep vision models are being widely de- ployed in safety-critical applications, e.g., autonomous driving, and explainability of such models is becoming a pressing concern. Among explanation methods, counter- factual explanations aim to find minimal andinterpretable changes to the input image that would also change the output of the model to be explained. Such explanations point end-users at the main factors that impact the deci- sion of the model. However, previous methods struggle to explain decision models trained on images with many ob- jects, e.g., urban scenes, which are more difficult to work with but also arguably more critical to explain. In this work, we propose to tackle this issue with an object-centric framework for counterfactual explanation generation. Our method, inspired by recent generative modeling works, en- codes the query image into a latent space that is structured in a way to ease object-level manipulations. Doing so, it provides the end-user with control over which search di- rections (e.g., spatial displacement of objects, style mod- ification, etc.) are to be explored during the counterfac- tual generation. We conduct a set of experiments on coun- terfactual explanation benchmarks for driving scenes, and we show that our method can be adapted beyond classifi- cation, e.g., to explain semantic segmentation models. To complete our analysis, we design and run a user study that measures the usefulness of counterfactual explanations in understanding a decision model. Code is available at https://github.com/valeoai/OCTET .	1. Introduction Deep learning models are now being widely deployed, notably in safety-critical applications such as autonomous driving. In such contexts, their black-box nature is a major concern, and explainability methods have been developed to improve their trustworthiness. Among them, counterfactual explanations have recently emerged to provide insights into a model’s decision [7, 52, 55]. Given a decision model and an input query, a counterfactual explanation is a data point that differs minimally butmeaningfully from the query in Query image Left Left Left Left Counterfactual Explanations w/ OCTET targeting the r oad targeting the yellow carFigure 1. Counterfactual explanations generated by OCTET. Given a classifier that predicts whether or not it is possible to go left, and a query image ( top left ), OCTET produces a counterfac- tual explanation where the most influential features that led to the decision are changed ( top right ). On the bottom row, we show that OCTET can also operate under different settings that result in dif- ferent focused explanations. We report the prediction made by the decision model at the top left of each image. a way that changes the output decision of the model. By looking at the differences between the query and the expla- nation, a user is able to infer — by contrast — which ele- ments were essential for the model to come to its decision. However, most counterfactual methods have only shown re- sults for explaining classifiers trained on single-object im- ages such as face portraits [4, 28, 29, 42, 46]. Aside from the technical difficulties of scaling up the resolution of the images, explaining decision models trained on scenes com- posed of many objects also present the challenge that those decisions are often multi-factorial. In autonomous driving, for example, most decisions have to take into account the position of all other road users, as well as the layout of the road and its markings, the traffic light and signs, the overall visibility, and many other factors. In this paper, we present a new framework, dubbed OCTET for Object-aware CounTerfactual ExplanaTions, to generate counterfactual examples for autonomous driving. We leverage recent advances in unsupervised compositional generative modeling [14] to provide a flexible explanation method. Exploiting such a model as our backbone, we can 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. 15062 assess the contribution of each object of the scene indepen- dently and look for explanations in relation to their posi- tions, their styles, or combinations of both. To validate our claims, extensive experiments are con- ducted for self-driving action decision models trained with the BDD100k dataset [60] and its BDD-OIA extension [59]. Fig. 1 shows counterfactual explanations found by OCTET: given a query image and a visual decision system trained to assess if the ego vehicle is allowed to go left or not, OCTET proposes a counterfactual example with cars parked on the left side that are moved closer. When inspecting specific items (bottom row of the figure), OCTET finds that moving the yellow car to the left, or adding a double line marking on the road, are ways to change the model’s decision. These explanations highlight that the absence of such elements on the left side of the road heavily influenced the model. To sum up, our contributions are as follows: • We tackle the problem of building counterfactual expla- nations for visual decision models operating on com- plex compositional scenes. We specifically target au- tonomous driving visual scenarios. • Our method is also a tool to investigate the role of spe- cific objects in a model’s decision, empowering the user with control over which type of explanation to look for. • We thoroughly evaluate the realism of our counter- factual images, the minimality and meaningfulness of changes, and compare against previous reference strate- gies. Beyond explaining classifiers, we also demon- strate the versatility of our method by addressing ex- planations for a segmentation network. • Finally, we conduct a user-centered study to assess the usefulness of our explanations in a practical case. As standard evaluation benchmarks for counterfactual ex- planations are lacking a concrete way to measure the interpretability of the explanations, our user-centered study is a key element to validate the presented pipeline.
Zhang_PRISE_Demystifying_Deep_Lucas-Kanade_With_Strongly_Star-Convex_Constraints_for_Multimodel_CVPR_2023	Abstract The Lucas-Kanade (LK) method is a classic iterative ho- mography estimation algorithm for image alignment, but often suffers from poor local optimality especially when im- age pairs have large distortions. To address this challenge, in this paper we propose a novel DeepStar-Convexif ied Luca s-Kanad e(PRISE) method for multimodel image align- ment by introducing strongly star-convex constraints into the optimization problem. Our basic idea is to enforce the neural network to approximately learn a star-convex loss landscape around the ground truth give any data to facili- tate the convergence of the LK method to the ground truth through the high dimensional space defined by the network. This leads to a minimax learning problem, with contrastive (hinge) losses due to the definition of strong star-convexity that are appended to the original loss for training. We also provide an efficient sampling based algorithm to leverage the training cost, as well as some analysis on the quality of the solutions from PRISE. We further evaluate our approach on benchmark datasets such as MSCOCO, GoogleEarth, and GoogleMap, and demonstrate state-of-the-art results, espe- cially for small pixel errors. Code can be downloaded from https://github.com/Zhang-VISLab .	1. Introduction Deep learning networks have achieved great success in homography estimation by directly predicting the transfor- mation matrix in various scenarios. However, the existing classic algorithms still take the place for showing more ex- plainability compared with the deep learning architectures. Such algorithms are often rooted from well-studied theoret- ical and empirical grounding. Current works often focus on combining the robustness of deep learning with explain- ability of classical algorithms to handle multimodel image alignment such as image modality and satellite modality. However, due to the high nonconvexity in homography esti-mation, such methods often suffer from poor local optimality. Recently Zhao et al. [77] proposed DeepLK for multi- model image alignment, i.e.,estimating the homography be- tween two planar projections of the same view but across dif- ferent modalities such as map and satellite images (see Sec. 3.1.1 for formal definition), based on the LK method [46]. This method consists of two novel components: •A new deep neural network was proposed to map im- ages from different modalities into the same feature space where the LK method can align them. •A new training algorithm was proposed as well by enforc- ing the local change on the loss landscape should be no less than a quadratic shape centered at the ground truth for any image pair, with no specific reason. Surprisingly, when we evaluate DeepLK based on the public code1, the proposed network cannot work well without the proposed training algorithm. This strongly motivate us to discover the mysteries in the DeepLK training algorithm. Deep Reparametrization. Our first insight from DeepLK is that the deep neural network essentially maps the align- ment problem into a much higher dimensional space by introducing a large amount of parameters. The high dimen- sional space provides the feasibility to reshape the loss land- scape of the LK method. Such deep reparametrization has been used as a means of reformulating some problems such as shape analysis [11], super-resolution and denoising [8], while preserving the properties and constraints in the original problems. This insight at test time can be interpreted as min ω∈Ωℓ(ω;x)reparametrization= = = = = = = = = ⇒ via deep learningmin ω∈Ωℓf(ω;x, θ∗), (1) where x∈ X denotes the input data, ℓdenotes a nonconvex differentiable function ( e.g., the LK loss) parametrized by ω∈Ω,f:X × Θ→ X denotes an auxiliary function presented by a neural network with learned weights θ∗∈Θ (e.g., the proposed network in DeepLK), and ℓfdenotes the 1https://github.com/placeforyiming/CVPR21-Deep- Lucas-Kanade-Homography 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. 13187 loss with deep reparametrization ( e.g., the DeepLK loss). In this way, the learning problem is how to train the network so that the optimal solutions can be located using gradient descent (GD) given data. Convex-like Loss Landscapes. Our second insight from DeepLK is that the learned loss landscape from their training algorithm tends to be convex-like (see their experimental results). This is an interesting observation, as it is evidenced in [39] that empirical more convex-like loss landscapes often return better performance. However, we cannot find any ex- plicit explanation through the paper about the reason, which raises the following questions that we aim to address: • Does the convex-like shape hold for any image pair? • If so, why? Is there any guarantee on solutions? Our Approach: Dee pStar-Convexif ied Luca s-Kanad e (PRISE). To mitigate the issue of poor local optimality in homography estimation, in this paper we propose a novel approach, namely PRISE, to enforce deep neural networks to approximately learn star-convex loss landscapes for the downstream tasks. Recently star-convexity [49] in noncon- vex optimization has been attracting more and more atten- tion [27, 30, 35, 38] because of its capability of finding near- optimal solutions based on GD with theoretical guarantees. Star-convex functions refer to a particular class of (typically) non-convex functions whose global optimum is visible from every point in a downhill direction. From this view, con- vexity is a special case of star-convexity. In the literature, however, most of the works focus on optimizing and analyz- ing star-convex functions, while learning such functions is hardly explored. In contrast, our PRISE imposes additional hinge losses, derived from the definition of star-convexity, on the learning objective during training. At test time, the nice convergence properties of star-convexity help find provably near-optimal solutions for the tasks using the LK method. We further show that DeepLK is a simplified and approxi- mate algorithm of PRISE, and thus shares some properties with ours, but with worse performance. Recently [78] have shown that stochastic gradient descent (SGD) will converge to global minimum in deep learning if the assumption of star-convexity in the loss landscapes hold. They validated this assumption (in a major part of training processes) empirically using relatively shallow networks and small-scale datasets by showing the classification training losses can converge to zeros. Nevertheless, we argue that this assumption may be too strong to hold in complex networks for challenging tasks, if without any additional imposition on learning. In our experiments we show that even we attempt to learn star-convex loss landscapes, the outputs at both training and test time are hardly perfect for complicated tasks. Contributions. Our key contributions are listed as follows: •We propose a novel PRISE method for multimodel im- age alignment by introducing (strongly) star-convex con-straints into the network training, which is rarely explored in the literature of deep learning. •We provide some analysis on the quality of the solutions from PRISE through star-convex loss landscapes. •We demonstrate the state-of-the-art results on some bench- mark datasets for multimodel image alignment with much better accuracy, especially when the pixel errors are small.
Zhang_Aligning_Step-by-Step_Instructional_Diagrams_to_Video_Demonstrations_CVPR_2023	Abstract Multimodal alignment facilitates the retrieval of in- stances from one modality when queried using another. In this paper, we consider a novel setting where such an align- ment is between (i) instruction steps that are depicted as assembly diagrams (commonly seen in Ikea assembly man- uals) and (ii) segments from in-the-wild videos; these videos comprising an enactment of the assembly actions in the real world. We introduce a supervised contrastive learning ap- proach that learns to align videos with the subtle details of assembly diagrams, guided by a set of novel losses. To study this problem and evaluate the effectiveness of our method, we introduce a new dataset: IAW—for Ikea assembly in the wild—consisting of 183 hours of videos from diverse fur- niture assembly collections and nearly 8,300 illustrations from their associated instruction manuals and annotated for their ground truth alignments. We define two tasks on this dataset: First, nearest neighbor retrieval between video segments and illustrations, and, second, alignment of in- struction steps and the segments for each video. Extensive experiments on IAW demonstrate superior performance of our approach against alternatives.	1. Introduction The rise of Do-It-Yourself (DIY) videos on the web has made it possible even for an unskilled person (or a skilled robot) to imitate and follow instructions to complete com- plex real world tasks [4, 23, 31]. One such task that is of- ten cumbersome to infer from instruction descriptions yet easy to imitate from a video is the task of assembling fur- niture from its parts. Often times the instruction steps in- volved in such a task are depicted in pictorial form, so that *Supported by an ANU-MERL PhD scholarship agreement. †Supported by Marie Sklodowska-Curie grant agreement No. 893465. ‡Supported by an ARC Future Fellowship No. FT200100421. … … … … … … Figure 1. An illustration of video-diagram alignment between a YouTube video (top) He0pCeCTJQM and an Ikea furniture man- ual (bottom) s49069795. they are comprehensible beyond the boundaries of language (e.g., Ikea assembly manuals). However, such instructional diagrams can sometimes be ambiguous, unclear, or may not match the furniture parts at disposal due to product variabil- ity. Having access to video sequences that demonstrate the precise assembly process could be very useful in such cases. Unfortunately, most DIY videos on the web are created by amateurs and often involve content that is not necessarily related to the task at hand. For example, such videos may include commentary about the furniture being assembled, or personal assembly preferences that are not captured in the instruction manual. Further, there could be large collections of videos on the web that demonstrate the assembly pro- cess for the same furniture but in diverse assembly settings; watching them could consume significant time from the as- sembly process. Thus, it is important to have a mechanism that can effectively align relevant video segments against the instructions steps illustrated in a manual. In this paper, we consider this novel task as a multimodal alignment problem [25,27], specifically for aligning in-the- wild web videos of furniture assembly and the respective diagrams in the instruction manuals as shown in Fig. 1. In contrast to prior approaches for such multimodal alignment, which usually uses audio, visual, and language modalities, 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. 2483 our task of aligning images with video sequences brings in several unique challenges. First, instructional diagrams can be significantly more abstract compared to text and audio descriptions. Second, illustrations of the assembly process can vary subtly from step-to-step (e.g., a rectangle placed on another rectangle could mean placing a furniture part on top of another). Third, the assembly actions, while depicted in a form that is easy for humans to understand, can be incom- prehensible for a machine. And last, there need not be com- mon standard or visual language followed when creating such manuals (e.g., a furniture piece could be represented as a rectangle based on its aspect ratio, or could be marked with an identifier, such as a part number). These issues make automated reasoning of instruction manuals against their video enactments extremely challenging. In order to tackle the above challenges, we propose a novel contrastive learning framework for aligning videos and instructional diagrams, which better suits the specifics of our task. We utilize two important priors—a video only needs to align with its own manual and adjacent steps in a manual share common semantics—that we encode as terms in our loss function with multimodal features computed from video and image encoder networks. To study the task in a realistic setting, we introduce a new dataset as part of this paper, dubbed IAW for Ikea as- sembly in the wild. Our dataset consists of nearly 8,300 il- lustrative diagrams from 420 unique furniture types scraped from the web and 1,005 videos capturing real-world furni- ture assembly in a variety of settings. We used the Ama- zon Mechanical Turk to obtain ground truth alignments of the videos to their instruction manuals. The videos involve significant camera motions, diverse viewpoints, changes in lighting conditions, human poses, assembly actions, and tool use. Such in-the-wild videos offer a compelling setting for studying our alignment task within its full generality and brings with it a novel research direction for exploring the multimodal alignment problem with exciting real-world ap- plications, e.g., robotic imitation learning, guiding human assembly, etc. To evaluate the performance of our learned alignment, we propose two tasks on our dataset: (i) nearest neighbor retrieval between videos and instructional diagrams, and (ii) alignment of the set of instruction steps from the manual to clips from an associated video sequence. Our experimental results show that our proposed approach leads to promising results against a compelling alternative, CLIP [27], demon- strating 9.68% improvement on the retrieval task and 12% improvement on the video-to-diagram alignment task.
Yang_Towards_Bridging_the_Performance_Gaps_of_Joint_Energy-Based_Models_CVPR_2023	Abstract Can we train a hybrid discriminative-generative model with a single network? This question has recently been answered in the affirmative, introducing the field of Joint Energy-based Model (JEM) [17, 48], which achieves high classification accuracy and image generation quality si- multaneously. Despite recent advances, there remain two performance gaps: the accuracy gap to the standard soft- max classifier, and the generation quality gap to state-of- the-art generative models. In this paper, we introduce a variety of training techniques to bridge the accuracy gap and the generation quality gap of JEM. 1) We incorporate a recently proposed sharpness-aware minimization (SAM) framework to train JEM, which promotes the energy land- scape smoothness and the generalization of JEM. 2) We exclude data augmentation from the maximum likelihood estimate pipeline of JEM, and mitigate the negative im- pact of data augmentation to image generation quality. Ex- tensive experiments on multiple datasets demonstrate our SADA-JEM achieves state-of-the-art performances and out- performs JEM in image classification, image generation, calibration, out-of-distribution detection and adversarial robustness by a notable margin. Our code is available at https://github.com/sndnyang/SADAJEM .	1. Introduction Deep neural networks (DNNs) have achieved state-of- the-art performances in a wide range of learning tasks, in- cluding image classification, image generation, object de- tection, and language understanding [21,30]. Among them, energy-based models (EBMs) have seen a flurry of inter- est recently, partially inspired by the impressive results of IGEBM [10] and JEM [17], which exhibit the capability of training generative models within a discriminative frame- work. Specifically, JEM [17] reinterprets the standard soft- max classifier as an EBM and achieves impressive perfor- mances in image classification and generation simultane- ously. Furthermore, these EBMs enjoy improved perfor- mance on out-of-distribution detection, calibration, and ad-versarial robustness. The follow-up works (e.g., [18, 48]) further improve the training in terms of speed, stability and accuracy. Despite the recent advances and the appealing property of training a single network for hybrid modeling, training JEM is still challenging on complex high-dimensional data since it requires an expensive MCMC sampling. Further- more, models produced by JEM still have an accuracy gap to the standard softmax classifier and a generation quality gap to the GAN-based approaches. In this paper, we introduce a few simple yet effective training techniques to bridge the accuracy gap and gener- ation quality gap of JEM. Our hypothesis is that both per- formance gaps are the symptoms of lack of generalization of JEM trained models. We therefore analyze the trained models under the lens of loss geometry. Figure 1 visu- alizes the energy landscapes of different models by the technique introduced in [34]. Since different models are trained with different loss functions, visualizing their loss functions is meaningless for the purpose of comparison. Therefore, the LSE energy functions (i.e., Eq. 4) of dif- ferent models are visualized. Comparing Figure 1(a) and (b), we find that JEM converges to extremely sharp lo- cal maxima of the energy landscape as manifested by the significantly large y-axis scale. By incorporating the re- cently proposed sharpness-aware minimization (SAM) [12] to JEM, the energy landscape of trained model (JEM+SAM) becomes much smoother as shown in Figure 1(c). This also substantially improves the image classification accu- racy and generation quality. To further improve the en- ergy landscape smoothness, we exclude data augmentation from the maximum likelihood estimate pipeline of JEM, and visualize the energy landscape of SADA-JEM in Fig- ure 1(d), which achieves the smoothest landscape among all the models considered. This further improves image gen- eration quality dramatically while retaining or sometimes improving classification accuracy. Since our method im- proves the performance of JEM primarily in the framework of sharpness-aware optimization, we refer it as SADA-JEM, a Sharpness-Aware Joint Energy-based Model with single branched Data Augmentation. 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. 15732 1.00 0.75 0.50 0.25 0.000.250.500.751.001.00 0.75 0.50 0.25 0.000.250.500.751.004.55.05.56.06.57.07.58.0(a) Softmax Classifier 1.00 0.75 0.50 0.25 0.000.250.500.751.001.00 0.75 0.50 0.25 0.000.250.500.751.00200 150 100 50 0 (b) JEM 1.00 0.75 0.50 0.25 0.000.250.500.751.001.00 0.75 0.50 0.25 0.000.250.500.751.0017.5 15.0 12.5 10.0 7.5 5.0 2.5 0.0 (c) JEM+SAM 1.00 0.75 0.50 0.25 0.000.250.500.751.001.00 0.75 0.50 0.25 0.000.250.500.751.003.0 2.5 2.0 1.5 1.0 0.5 0.00.5 (d) SADA-JEM Figure 1. Visualizing the energy landscapes [34] of different models trained on CIFAR10. Note the dramatic scale differences of the y-axes, indicating SADA-JEM identifies the smoothest local optimum among all the methods considered. Our main contributions are summarized as follows: 1. We investigate the energy landscapes of different mod- els and find that JEM leads to the sharpest one, which potentially undermines the generalization of trained models. 2. We incorporate the sharpness-aware minimization (SAM) framework to JEM to promote the energy land- scape smoothness, and thus model generalization. 3. We recognize the negative impact of data augmenta- tion in the training pipeline of JEM, and introduce two data loaders for image classification and image gen- eration separately, which improves image generation quality significantly. 4. Extensive experiments on multiple datasets show that SADA-JEM achieves the state-of-the-art discrimina- tive and generative performances, while outperforming JEM in calibration, out-of-distribution detection and adversarial robustness by a notable margin.
Zhang_Analyzing_Physical_Impacts_Using_Transient_Surface_Wave_Imaging_CVPR_2023	Abstract The subtle vibrations on an object’s surface contain in- formation about the object’s physical properties and its in- teraction with the environment. Prior works imaged sur- face vibration to recover the object’s material properties via modal analysis, which discards the transient vibra- tions propagating immediately after the object is disturbed. Conversely, prior works that captured transient vibrations focused on recovering localized signals ( e.g., recording nearby sound sources), neglecting the spatiotemporal re- lationship between vibrations at different object points. In this paper, we extract information from the transient surface vibrations simultaneously measured at a sparse set of object points using the dual-shutter camera described by Sheinin et al. [37]. We model the geometry of an elastic wave gen- erated at the moment an object’s surface is disturbed ( e.g., a knock or a footstep) and use the model to localize the dis- turbance source for various materials ( e.g., wood, plastic, tile). We also show that transient object vibrations contain additional cues about the impact force and the impacting object’s material properties. We demonstrate our approach in applications like localizing the strikes of a ping-pong ball on a table mid-play and recovering the footsteps’ locations by imaging the floor vibrations they create.	1. Introduction Our environment is teeming with vibrations created by the interaction of physical objects. Some vibrations, like a knock on the door or the sound of a ball bouncing off the ground, can be perceived by humans because they are transmitted from the vibrating object’s surface via the air. However, many vibrations that fill our world are too subtle for auditory-based remote sensing. Moreover, much like ripples in a pond, the transient spatial shapes such vibrations create on object surfaces are a visual cue that can disclose the source of the disturbance and other object properties. Object vibrations can be divided into two main types: transient and modal. For example, consider the vibrations of a tuning fork. When struck, the impulse creates tran- sient waves propagating from the impact source until they [37] Figure 1. When physical objects interact, like a ping pong ball bouncing off the table, they create minute vibrations that propa- gate through the objects’ surfaces and interiors. The transient vi- brations that occur immediately on impact, exaggerated here for visualization, carry information about the impact source location. We image the surface vibrations at a sparse set of locations using the imaging system of Sheinin et al. [37]. We model the elastic wave propagation and recover the impact source locations with- out a direct line-of-sight on the impacted surface. Visit the project page for videos of results [1]. reach and vibrate the fork’s entire body. After a short time interval, the transient vibrations die down, leaving the fork to vibrate at its resonant modal frequencies. Modal analysis, which aims to measure these resonant frequen- cies [11, 13, 42], can reveal the tuning fork’s designed tone (e.g., 440 Hz for the Atone) and can also be used to analyze the fork’s material properties [9, 14, 18]. While extremely useful, modal analysis ignores the tran- sient vibrations that occur at the moment of impact. Such transient vibrations contain valuable cues about the dis- turbance’s origin, its magnitude, and the properties of the object causing the disturbance ( e.g., a falling basketball vs. a falling rock). Prior works that did sense transient vibrations primarily focused on localized low-dimensional signals such as heartbeats [42, 44, 48], music and speech [8, 15, 37, 45, 46, 48], and musical instruments [37]. These works disregard the spatiotemporal relationship between transient vibrations at different object points. This paper focuses on recovering the physical loca- tion of an impacting object from transient surface vibra- 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. 4339 (a)setup (b)pre-reflections (c)post-reflections Figure 2. Elastic wave propagation in isotropic objects. (a)An electronic knocker creates repeated short knocks on a whiteboard. For each knock, a laser Doppler vibrometer (LDV) sensor is used to optically measure the temporal vertical displacement at a single point. Aggregating and synchronizing measurements from multi- ple board points generates a video showing the surface displace- ment with time. (b)Displacement 1ms after impact. Observe the circular shape of the outgoing wave. (c)Displacement 3.1ms after impact. Here, the outgoing wave has reflected from the board’s boundaries. tions measured simultaneously at multiple surface points using the dual-shutter camera of Sheinin et al. [37]. This task opens the door to potential applications like localizing sound sources in walls ( e.g., pipe bursting), localizing bullet or bird impacts on airplanes mid-flight, or impacts on ship hulls from dockside, tugs, or other debris, localizing shell- ground impacts on battlefields, localizing people in building fires or hostage situations by observing external vibrations on ceilings or side walls, and more. While, in general, object shape and material determine its vibration profile, we show that immediately after impact, there exists a short time interval ( ∼1.5 ms long) where the surface vibrations can be modeled as an outwardly propa- gating elastic wave. We derive an approximate model of the wave’s geometry for both isotropic and anisotropic materi- als and develop a backprojection-based algorithm to local- ize the impact sources using the vibrations within this time interval. Unlike prior works that merely visualize acoustic wave propagation [36], we explicitly model its transient be- havior and show that only a sparse set of points is required to determine the wave’s source. We verified our approach on various materials, including wood, plastic, glass, porcelain, and gypsum. In our exper- iments, we localized impact sources with an average error between 1.1 cm and 2.9 cm for 40 cm ×40 cm and 90 cm ×90 cm surfaces, respectively. We also show applications like localizing ping-pong ball strikes on the table mid-play and localizing footsteps through floor vibrations beyond a camera’s line of sight. Beyond impact localization, we show that the transient surface vibrations can convey more information about the impacting object and the impacted surface. For surfaces of unknown material, we estimate the material anisotropy by measuring vibrations at known surface points and fit- ting a material-specific wave propagation model parameter. Our preliminary experiments suggest that the transient vi-brations’ amplitudes relate to the force applied to disturb the object [20, 28, 31], and that the vibrations’ frequency content depends on the stiffness and shape on the impacting object. We thus believe our work can inspire a new class of transient vibration imaging approaches that opens the door for novel vision tasks.
Yu_DyLiN_Making_Light_Field_Networks_Dynamic_CVPR_2023	Abstract Light Field Networks, the re-formulations of radiance fields to oriented rays, are magnitudes faster than their co- ordinate network counterparts, and provide higher fidelity with respect to representing 3D structures from 2D obser- vations. They would be well suited for generic scene rep- resentation and manipulation, but suffer from one problem: they are limited to holistic and static scenes. In this pa- per, we propose the Dynamic Light Field Network (DyLiN) method that can handle non-rigid deformations, including topological changes. We learn a deformation field from in- put rays to canonical rays, and lift them into a higher di- mensional space to handle discontinuities. We further in- troduce CoDyLiN, which augments DyLiN with controllable attribute inputs. We train both models via knowledge distil- lation from pretrained dynamic radiance fields. We eval- uated DyLiN using both synthetic and real world datasets that include various non-rigid deformations. DyLiN qual- itatively outperformed and quantitatively matched state-of- the-art methods in terms of visual fidelity, while being 25− 71×computationally faster. We also tested CoDyLiN on at- tribute annotated data and it surpassed its teacher model. Project page: https://dylin2023.github.io .	1. Introduction Machine vision has made tremendous progress with re- spect to reasoning about 3D structure using 2D observa-tions. Much of this progress can be attributed to the emer- gence of coordinate networks [6,21,26], such as Neural Ra- diance Fields (NeRF) [23] and its variants [2, 20, 22, 39]. They provide an object agnostic representation for 3D scenes and can be used for high-fidelity synthesis for unseen views. While NeRFs mainly focus on static scenes, a series of works [10,27,29,34] extend the idea to dynamic cases via additional components that map the observed deformations to a canonical space, supporting moving and shape-evolving objects. It was further shown that by lifting this canonical space to higher dimensions the method can handle changes in scene topology as well [28]. However, the applicability of NeRF models is consid- erably limited by their computational complexities. From each pixel, one typically casts a ray from that pixel, and nu- merically integrates the radiance and color densities com- puted by a Multi-Layer Perceptron (MLP) across the ray, approximating the pixel color. Specifically, the numeri- cal integration involves sampling hundreds of points across the ray, and evaluating the MLP at all of those locations. Several works have been proposed for speeding up static NeRFs. These include employing a compact 3D represen- tation structure [9, 18, 43], breaking up the MLP into multi- ple smaller networks [30,31], leveraging depth information [7, 24], and using fewer sampling points [17, 24, 42]. Yet, these methods still rely on integration and suffer from sam- pling many points, making them prohibitively slow for real- time applications. Recently, Light Field Networks (LFNs) [32] proposed replacing integration with a direct ray-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. 12397 color regressor, trained using the same sparse set of images, requiring only a single forward pass. R2L [36] extended LFNs to use a very deep residual architecture, trained by distillation from a NeRF teacher model to avoid overfit- ting. In contrast to static NeRF acceleration, speeding up dynamic NeRFs is a much less discussed problem in the literature. This is potentially due to the much increased dif- ficulty of the task, as one also has to deal with the high variability of motion. In this direction, [8, 38] greatly re- duce the training time by using well-designed data struc- tures, but their solutions still rely on integration. LFNs are clearly better suited for acceleration, yet, to the best of our knowledge, no works have attempted extending LFNs to the dynamic scenario. In this paper, we propose 2 schemes extending LFNs to dynamic scene deformations, topological changes and controllability. First, we introduce DyLiN, by incorpo- rating a deformation field and a hyperspace representa- tion to deal with non-rigid transformations, while distilling knowledge from a pretrained dynamic NeRF. Afterwards, we also propose CoDyLiN, via adding controllable input attributes, trained with synthetic training data generated by a pretrained Controllable NeRF (CoNeRF) [13] teacher model. To test the efficiencies of our proposed schemes, we perform empirical experiments on both synthetic and real datasets. We show that our DyLiN achieves better image quality and an order of magnitude faster rendering speed than its original dynamic NeRF teacher model and the state- of-the-art TiNeuV ox [8] method. Similarly, we also show that CoDyLiN outperforms its CoNeRF teacher. We further execute ablation studies to verify the individual effective- ness of different components of our model. Our methods can be also understood as accelerated versions of their re- spective teacher models, and we are not aware of any prior works that attempt speeding up CoNeRF. Our contributions can be summarized as follows: • We propose DyLiN, an extension of LFNs that can handle dynamic scenes with topological changes. DyLiN achieves this through non-bending ray defor- mations, hyperspace lifting for whole rays, and knowl- edge distillation from dynamic NeRFs. • We show that DyLiN achieves state-of-the-art results on both synthetic and real-world scenes, while being an order of magnitude faster than the competition. We also include an ablation study to analyze the contribu- tions of our model components. • We introduce CoDyLiN, further extending our DyLiN to handle controllable input attributes.
Yang_Modeling_Entities_As_Semantic_Points_for_Visual_Information_Extraction_in_CVPR_2023	Abstract Recently, Visual Information Extraction (VIE) has been becoming increasingly important in both the academia and industry, due to the wide range of real-world applications. Previously, numerous works have been proposed to tackle this problem. However, the benchmarks used to assess these methods are relatively plain, i.e., scenarios with real-world complexity are not fully represented in these benchmarks. As the first contribution of this work, we curate and re- lease a new dataset for VIE, in which the document im- ages are much more challenging in that they are taken from real applications, and difficulties such as blur, partial oc- clusion, and printing shift are quite common. All these fac- tors may lead to failures in information extraction. There- fore, as the second contribution, we explore an alternative approach to precisely and robustly extract key information from document images under such tough conditions. Specif- ically, in contrast to previous methods, which usually ei- ther incorporate visual information into a multi-modal ar- chitecture or train text spotting and information extraction in an end-to-end fashion, we explicitly model entities as se- mantic points, i.e., center points of entities are enriched with semantic information describing the attributes and re- lationships of different entities, which could largely bene- fit entity labeling and linking. Extensive experiments on standard benchmarks in this field as well as the proposed *Equal Contribution. †Correspondence Author.dataset demonstrate that the proposed method can achieve significantly enhanced performance on entity labeling and linking, compared with previous state-of-the-art models. Dataset is available at https://www.modelscope. cn/datasets/damo/SIBR/summary .	1. Introduction Visually Rich Documents (VRDs) are ubiquitous in daily, industrial, and commercial activities, such as receipts of shopping, reports of physical examination, product man- uals, and bills of entry. Visual Information Extraction (VIE) aims to automatically extract key information from these VRDs, which can significantly facilitate subsequent pro- cessing and analysis. Due to its broad applications and grand technical challenges, VIE has recently attracted con- siderable attention from both the Computer Vision com- munity [33, 34, 39] and the Natural Language Processing community [16,35,37]. Typical techniques for tackling this challenging problem include essential electronic conversion of image (OCR) [28, 30, 41], intermediate procedure of structure analysis [25] and high-level understanding of con- tents [35], among which entities play an important role as an aggregation of vision, structure, and language. Though substantial progresses [11, 19, 35] have been made, it is still challenging to precisely and reliably extract key information from document images in unconstrained conditions. As shown in Fig. 1, in real-world scenarios doc- uments may have various formats, be captured casually with 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. 15358 a mobile phone, or exist occlusion or shift in printing, all of which would pose difficulties for VIE algorithms. To highlight the challenges in real applications and pro- mote the development of research in VIE, we establish a new dataset called Structurally-rich Invoices, Bills and Receipts in the Wild ( SIBR for short), which contains 1,000 images with 71,227 annotated entity instances and 39,004 entity links. The challenges of SIBR lie in: (1) The docu- ments are from different real-world scenarios, so their for- mats and structures might be complicated and varying; (2) The image quality may be very poor, i.e., blur, noise, and uneven illumination are frequently seen; (3) The printing process is imperfect that shift and rotation might happen. To deal with these difficulties, we explore an novel ap- proach for information extraction from VRDs. Different from previous methods, which usually employ a sequen- tial pipeline that first uses an off-the-shelf OCR engine to detect and read textual information (location and content) and then fuses such information with visual cues for follow- up entity labeling ( a.k.a. entity extraction) and linking in a multi-modal architecture (mostly a Transformer) [19, 37], the proposed method adopts a unified framework that all components, including text detection, text recognition, en- tity extraction and linking, are jointly modeled and trained in an integrated way. This means that in our method a sep- arate OCR engine is no longer necessary. The benefits are two-fold: (1) The accuracy of entity labeling and linking will not be limited by the capacity of the OCR engine; (2) The running speed of the whole pipeline could be boosted. Drawing inspirations from general object detection [4, 15, 40, 42] and vision-language joint learning [11, 14, 31], we put forward to model entities as semantic points ( ESP for short). Specifically, as shown in Fig. 3, entities are rep- resented using their center points, which are enriched with semantics, such as geometric and linguistic information, to perform entity labeling and linking. To better learn a joint vision-language representation, we also devise three train- ing tasks that are well integrated into the paradigm. The entity-image text matching ( EITM ) task, which is only used in the pre-training stage, learns to align entity-level vision vectors and language vectors (encoded with off-the-shell BERT) with a contrastive learning paradigm. Entity ex- traction ( EE) and Entity linking ( EL), the main tasks for VIE, are used in the pre-training, fine-tuning, and inference stages. In these two modules, region features and position embedding (from ground truth or detection branch) are en- coded with transformer layers and then decoded to entity classes and relations. Owing to the joint vision-language representation, text recognition is no longer a necessary module in our framework, and we will discuss the impact of the text recognition branch in Sec. 5.4. Extensive experiments have been conducted on stan- dard benchmarks for VIE (such as FUNSD, XFUND, andCORD) as well as the proposed SIBR dataset. We found that compared with previous state-of-the-art methods, the proposed ESP algorithm can achieve highly competitive performance. Especially, it shows an advantage in the task of entity linking. Our main contributions can be summa- rized as follows: (1) We curate and release a new dataset for VIE, in which the document images are with real-world complexity and difficulties. (2) We devise a unified frame- work for spotting, labeling and linking entities, where a sep- arate OCR engine is unnecessary. (3) We adopt three vision- language joint modeling tasks for learning informative rep- resentation for VIE. (4) Extensive experiments demonstrate the effectiveness and advantage of our approach.
Yu_Block_Selection_Method_for_Using_Feature_Norm_in_Out-of-Distribution_Detection_CVPR_2023	Abstract Detecting out-of-distribution (OOD) inputs during the inference stage is crucial for deploying neural networks in the real world. Previous methods typically relied on the highly activated feature map outputted by the network. In this study, we revealed that the norm of the feature map obtained from a block other than the last block can serve as a better indicator for OOD detection. To leverage this insight, we propose a simple framework that comprises two metrics: FeatureNorm , which computes the norm of the feature map, and NormRatio , which calculates the ra- tio of FeatureNorm for ID and OOD samples to evaluate the OOD detection performance of each block. To iden- tify the block that provides the largest difference between FeatureNorm of ID and FeatureNorm of OOD, we cre- ate jigsaw puzzles as pseudo OOD from ID training sam- ples and compute NormRatio, selecting the block with the highest value. After identifying the suitable block, OOD detection using FeatureNorm outperforms other methods by reducing FPR95 by up to 52.77% on CIFAR10 bench- mark and up to 48.53% on ImageNet benchmark. We demonstrate that our framework can generalize to vari- ous architectures and highlight the significance of block se- lection, which can also improve previous OOD detection methods. Our code is available at https://github.com/gist- ailab/block-selection-for-OOD-detection.	1. Introduction Neural networks have widely been utilized in the real world, such as in autonomous cars [9, 21] and medical di- agnoses [7,38]. In the real world, neural networks often en- counter previously unseen input that are different from the training data. If the system fails to recognize those input as unknown input, there can be a dangerous consequence. For example, a medical diagnosis system may recognize an un- seen disease image as one of the known diseases. This gives rise to the importance of the out-of-distribution (OOD) de- tection, which makes users operate a neural network system more safely in the real world. Figure 1. Histogram of norm of the feature map produced by con- volutional blocks of ResNet18. In last block (a), the norm of ID (black) is hard to separate from OOD (blue, orange) compared to the one from the penultimate block (b). In practice, various outputs of the network can be used as an indicator to separate the in-distribution (ID) and out- of-distribution (OOD) data. For instance, output probabil- ity [15], calibrated output probability [28], and output en- ergy [30] are used as an indicator. The output of a neural network is commonly calculated using a feature vector of the feature extractor and a weight vector of the classifica- tion layer. It is known that the norm of the feature vector can be an indicator of input image quality [23, 37, 40] or level of awareness [46]. Thus, we ask the following ques- tion: Can we use the norm of the feature as an indicator to separate ID and OOD? In this paper, we first reveal the key observation con- cerning the last block of neural networks sometimes deteri- orating owing to the overconfidence issue [10, 11]. Empir- ically, we show that OOD images highly activate filters of the last block (i.e., large norm; see Figure 1, left) on a net- work trained with CIFAR10 while lowly activate filters of the penultimate block (i.e., small norm; see Figure 1, right). As a result, OOD detection methods that consider overacti- vated feature [42] and overconfident output [28] have been successful. However, we find that the norm of the feature map for the OOD and ID is quite separable in the penulti- mate block compared to the last block. This motivates a simple and effective OOD detection 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. 15701 Figure 2. Illustration of our proposed out-of-distribution detection framework. FeatureNorm refers to a norm calculation for the given feature map produced by the block. We use NormRatio of ID and pseudo OOD (i.e., jigsaw puzzles) to find which block is suitable for OOD detection (a). During inference time, for a given input image, the OOD score is calculated by FeatureNorm on the selected block (b). IfFeatureNorm for a given input is smaller than the threshold, the given
Zhang_Extracting_Motion_and_Appearance_via_Inter-Frame_Attention_for_Efficient_Video_CVPR_2023	Abstract Effectively extracting inter-frame motion and appear- ance information is important for video frame interpolation (VFI). Previous works either extract both types of informa- tion in a mixed way or devise separate modules for each type of information, which lead to representation ambiguity and low efficiency. In this paper, we propose a new mod- ule to explicitly extract motion and appearance information via a unified operation. Specifically, we rethink the infor- mation process in inter-frame attention and reuse its at- tention map for both appearance feature enhancement and motion information extraction. Furthermore, for efficient VFI, our proposed module could be seamlessly integrated into a hybrid CNN and Transformer architecture. This hy- brid pipeline can alleviate the computational complexity of inter-frame attention as well as preserve detailed low- level structure information. Experimental results demon- strate that, for both fixed- and arbitrary-timestep interpo- lation, our method achieves state-of-the-art performance on various datasets. Meanwhile, our approach enjoys a lighter computation overhead over models with close per- formance. The source code and models are available at https://github.com/MCG-NJU/EMA-VFI .	1. Introduction As a fundamental low-level vision task, the goal of video frame interpolation (VFI) is to generate intermediate frames given a pair of consecutive frames [17, 33]. It has a wide range of real-life applications, such as video com- pression [53], novel-view rending [13,47], and slow-motion video creation [19]. In general, VFI can be seen as the pro- cess of capturing the motion between consecutive frames and then blending the corresponding appearance to synthe- size the intermediate frames. From this perspective, the mo- tion and appearance information between input frames is essential for achieving excellent performance in VFI tasks. *: Corresponding author ([email protected]). Figure 1. Illustration of various approaches in video frame inter- polation for acquiring motion and appearance information. Concerning the extraction paradigm of motion and ap- pearance information, the current VFI approaches can be divided into two categories. The first is to handle both ap- pearance and motion information in a mixed way [2,11,14, 17, 20, 21, 30, 33, 37, 38, 44], as shown in Fig. 1(a). The two neighboring frames are directly concatenated and fed into a backbone composed of stacked similar modules to generate features with mixed motion and appearance infor- mation. Though simple, this approach requires an elabo- rate design and high capacity in the extractor module, as it needs to deal with both motion and appearance information jointly. The absence of explicit motion information also re- sults in limitations for arbitrary-timestep interpolation. The second category, as shown in Fig. 1(b), is to design sep- arate modules for motion and appearance information ex- traction [9, 18, 35, 40–42, 45, 56]. This approach requires additional modules, such as cost volume [18, 40, 41], to ex- tract motion information, which often imposes a high com- putational overhead. Also, only extracting appearance fea- 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. 5682 tures from a single frame fails to capture the correspondence of appearance information of the same regions between frames, which is an effective cue for the VFI task [18]. To address the issues of the above two extraction paradigms, in this paper, we propose to explicitly extract both motion and appearance information via a unified op- eration of inter-frame attention. With a single inter-frame attention, as shown in Fig. 1(c), we are able to enhance the appearance features between consecutive frames and ac- quire motion features at the same time by reusing the atten- tion maps. This basic processing unit could be stacked to obtain the hierarchical motion and appearance information. Specifically, for any patch in the current frame, we take it as the query and its temporal neighbors as keys and values to derive an attention map representing their temporal cor- relation. After that, the attention map is leveraged to aggre- gate the appearance features of neighbors to contextualize the current region representation. In addition, the attention map is also used to weight the displacement of neighbors to get an approximate motion vector of the patch from the current frame to the neighbor frame. Finally, the obtained features are utilized with light networks for motion estima- tion and appearance refinement to synthesize intermediate frames. Compared with previous works, our design enjoys three advantages. (1) The appearance features of each frame can be enhanced with each other yet not be mixed with mo- tion features to preserve the detailed static structure infor- mation. (2) The obtained motion features can be scaled by time and then used as cues to guide the generation of frames at any moment between input frames. (3) We only need to control the complexity and the number of modules to bal- ance the overall performance and the inference speed. Directly using inter-frame attention on original reso- lution results in huge memory usage and computational overhead. Inspired by some recent works [8, 12, 26, 49, 54, 55, 58], which combines Convolutional Neural Net- work (CNN) [23] with Transformer [48] to improve the model learning ability and robustness, we adopt a sim- ple but effective architecture: first utilize CNN to extract high-resolution low-level features and then use Transformer blocks equipped with inter-frame attention to extracting low-resolution motion features and inter-frame appearance features. Our proposed module could be seamlessly inte- grated into this hybrid pipeline to extract motion and ap- pearance features efficiently without losing fine-grained in- formation. Our contributions are summarized as follows: • We propose to utilize inter-frame attention to extract both motion and appearance information simultane- ously for video frame interpolation. • An hybrid CNN and Transformer design is adopted to overcome the overhead bottleneck of the inter- frame attention at high-resolution input while preserv-ing fine-grained information. • Our model achieves state-of-the-art performance on various datasets while being efficient compared to models with similar performance.
Zhang_Federated_Domain_Generalization_With_Generalization_Adjustment_CVPR_2023	Abstract Federated Domain Generalization (FedDG) attempts to learn a global model in a privacy-preserving manner that generalizes well to new clients possibly with domain shift. Recent exploration mainly focuses on designing an unbi- ased training strategy within each individual domain. How- ever, without the support of multi-domain data jointly in the mini-batch training, almost all methods cannot guar- antee the generalization under domain shift. To overcome this problem, we propose a novel global objective incorpo- rating a new variance reduction regularizer to encourage fairness. A novel FL-friendly method named Generaliza- tion Adjustment (GA) is proposed to optimize the above ob- jective by dynamically calibrating the aggregation weights. The theoretical analysis of GA demonstrates the possibility to achieve a tighter generalization bound with an explicit re-weighted aggregation, substituting the implicit multi- domain data sharing that is only applicable to the con- ventional DG settings. Besides, the proposed algorithm is generic and can be combined with any local client training- based methods. Extensive experiments on several bench- mark datasets have shown the effectiveness of the proposed method, with consistent improvements over several FedDG algorithms when used in combination. The source code is released at https://github.com/MediaBrain- SJTU/FedDG-GA	1. Introduction Federated Learning (FL) has recently emerged as a prevalent privacy-preserving paradigm for collaborative learning on distributed data [32]. Existing studies mainly investigate the problem of how to improve the conver- gence and performance of the source clients’ data distribu- tion [18, 27, 44]. A more practical problem, how to make models trained on sites of heterogeneous distributions gen- Aggregation on client modelsMini-Batch Training Unseen T est DomainSource Domains for DG Source Domains for FedDG Generalization modelClient 1 Client 2 Client 3 DG FedDGSOTA DG requires access multi-domains to capture the invariant patternsFigure 1. The difference between DG and FedDG is whether the domains are isolated in training. Specifically, previous SOTA DG methods that require access to multiple domains in the mini-batch training are inapplicable to FedDG. eralize to target clients of unknown distributions, i.e.Feder- ated Domain Generalization (FedDG) [30], remains under- explored. While label distribution shift has been considered in traditional FL, FedDG focuses on the domain shift among clients and considers each client as an individual domain. The challenge lies in the domain shift [19] both among the training clients and from training to testing clients. While FedDG shares a similar goal as standard Do- main Generalization (DG) [4,12,40], i.e.,generalizing from multi-source domains to unseen domains, it disallows di- rect data sharing among clients, as shown in Figure 1, which makes most existing DG methods hardly applica- ble. Current methods for FedDG focus on unbiased lo- cal training within each isolated domain. As the first at- tempt, Liu et al. [30] propose a meta-learning framework 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. 3954 with Fourier-based augmentation during the local training for better generalization. Jiang et al . [17] further pro- pose constraining local models’ flatness on top of a simi- lar Fourier-based normalization method. However, only fo- cusing on an improved local training strategy cannot guar- antee that the global model is generalizable enough to un- seen domains. Instead, a common practice for aggregat- ing local models into a global model is by fixed weights as in FedAvg [32], assuming that each client constantly con- tributes to the global model. Even the subsequent improve- ments from the federated optimization perspective, e.g., FedNova [44], are mainly designed for the statistical het- erogeneity of the same domain, not for the setting of treat- ing each client as an individual domain. Yuan et al. [50] have suggested that domains tend to contribute non-equally to the global model and ignoring their differences may sig- nificantly reduce the model’s generalizability. As one has no clue regarding to the distribution of unseen domains, it is reasonable to assume that a global model with fair performance among all clients may lead to better gener- alization performance. We thus introduce a new fairness ob- jective measured by the variance of the generalization gaps among different source domains. The data privacy issue in the FL setting has prevented direct optimization of the pro- posed objective. We thus design a novel privacy-preserving method named Generalization Adjustment to optimize the objective. At the high level, GA leverages the domain flat- ness constraint, a surrogate of the intractable domain di- vergence constraint, to approximately explore the optimal domain weights. Technically, we use a momentum mech- anism to dynamically compute a weight for each isolated domain by tracing the domain generalization gap, which is then involved in the aggregation of FedDG to enhance the generalization ability. Because the gap information does not contain any domain information of each client, GA will not cause additional risk of privacy leakage. Meanwhile, the theoretical analysis of our method shows that a tighter generalization bound is achieved by setting the aggregation weights inversely proportional to the generalization gaps, which leads to reduced variance in generalization gaps. The contribution of our paper is summarized as follows: • We introduce a novel optimization objective for FedDG with a new variance reduction regularizer, which can con- strain the fairness of the global model. • We design an FL-friendly method named Generalization Adjustment to tackle the aforementioned novel objective. Our theoretical analysis has revealed that GA leads to a tighter generalization bound for FedDG. • Extensive experiments on a range of benchmark datasets have shown consistent improvement when combining GA with different federated learning algorithms.
Yu_OSRT_Omnidirectional_Image_Super-Resolution_With_Distortion-Aware_Transformer_CVPR_2023	Abstract Omnidirectional images (ODIs) have obtained lots of re- search interest for immersive experiences. Although ODIsrequire extremely high resolution to capture details of theentire scene, the resolutions of most ODIs are insufﬁcient. Previous methods attempt to solve this issue by image super-resolution (SR) on equirectangular projection (ERP)images. However , they omit geometric properties of ERP inthe degradation process, and their models can hardly gener-alize to real ERP images. In this paper , we propose Fisheye downsampling, which mimics the real-world imaging pro- cess and synthesizes more realistic low-resolution samples.Then we design a distortion-aware Transformer (OSRT) tomodulate ERP distortions continuously and self-adaptively.Without a cumbersome process, OSRT outperforms previ-ous methods by about 0.2dB on PSNR. Moreover , we pro- pose a convenient data augmentation strategy, which syn-thesizes pseudo ERP images from plain images. This simplestrategy can alleviate the over-ﬁtting problem of large net-works and signiﬁcantly boost the performance of ODISR. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT.	1. Introduction In pursuit of the realistic visual experience, omnidi- rectional images (ODIs), also known as 360◦images or panoramic images, have obtained lots of research interest inthe computer vision community. In reality, we usually viewODIs with a narrow ﬁeld-of-view (FOV), e.g., viewing in a headset. To capture details of the entire scene, ODIs requireextremely high resolution, e.g.,4 K×8K [ 1]. However, due to the high industrial cost of camera sensors with high pre-cision, the resolutions of most ODIs are insufﬁcient. Recently, some attempts have been made to solve this problem by image super-resolution (SR) [ 12,15,28,39,40]. *Equal contribution †Corresponding author (e-mail: [email protected]) Unseen LR LAU-Net [ 12] w/o Fisheye OSRT w/o Fisheye OSRT w/ Fisheye Figure 1. Visual comparisons of ×8 SR results on LR images1with unknown degradations. Fisheye denotes that the downsamplingprocess in training stages is under Fisheye images. As most of the ODIs are stored and transmitted in the equirectangular projection (ERP) type, the SR process isusually performed on the ERP images. To generate high-/low-resolution training pairs, existing ODISR methods[12,15,28,39,40] directly apply uniform bicubic down- sampling on the original ERP images (called ERP down- sampling), which is identical to general image SR settings [24,43]. While omitting geometric properties of ERP in the degradation process, their models can hardly general-ize to real ERP images. We can observe missing struc-tures and blur textures in Fig. 1. Therefore, we need a more appropriate degradation model before studying SR al- gorithms. In practice, ODIs are acquired by the ﬁsheyelens and stored in ERP . Given that the low-resolution is-sue in real-world scenarios is caused by insufﬁcient sensorprecision and density, the downsampling process should beapplied to original-formatted images before converting intoother storage types. Thus, to be conformed with real-worldimaging processes, we propose to apply uniform bicubic 1Photoed by Peter Leth on Flickr, with CC license . 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. 13283 downsampling on Fisheye images, which are the original format of ODIs. The new downsampling process (calledFisheye downsampling) applies uniform bicubic downsam-pling on Fisheye images before converting them to ERP im-ages. Our Fisheye downsampling is more conducive to ex-ploring the geometric property of ODIs. The key issue of ODISR algorithm design is to utilize the geometric properties of ERP images, which is also thefocus of previous methods. For example, Nishiyama et al. [28] add a distortion-related condition as an additional in- put. LAU-Net [ 12] splits the whole ERP image into patches by latitude band and learns upscaling processes separately.However, the separated learning process will lead to infor-mation disconnection between adjacent patches. SphereSR[40] learns different upscaling functions on various projec- tion types, but will inevitably introduce multiple-time com- putation costs. To push the performance upper bound, wepropose the ﬁrst Transformer for Omnidirectional image Super-Resolution (OSRT), and incorporate geometric prop-erties in a distortion-aware manner. Speciﬁcally, to mod-ulate distorted feature maps, we implement feature-levelwarping, in which offsets are learned from latitude condi- tions. In OSRT, we introduce two dedicated blocks to adaptlatitude-related distortion: distortion-aware attention block(DAAB), and distortion-aware convolution block (DACB).DAAB and DACB are designed to perform distortion mod-ulation in arbitrary Transformers and ConvNets. Thesetwo blocks can directly replace the multi-head self-attentionblock and convolution layer, respectively. The beneﬁt ofDAAB and DACB can be further improved when being in-serted into the same backbone network. OSRT outperformsprevious methods by about 0.2dB on PSNR (Tab. 2). However, the increase of network capacity will also en- large the overﬁtting problem of ODISR, which is rarelymentioned before. The largest ODIs dataset [ 12] contains only 1K images, which cannot provide enough diversity fortraining Transformers. Given that acquiring ODIs requiresexpensive equipment and tedious work, we propose to gen-erate distorted ERP samples from plain images for data aug-mentation. In practice, we regard a plain image as a sampledperspective, and project it back to the ERP format. Then wecan introduce 146K additional training patches, 6 times of the previous dataset. This simple strategy can signiﬁcantlyboost the performance of ODISR (Tab. 4) and alleviate the over-ﬁtting problem of large networks (Fig. 9). A similar data augmentation method is also applied in Nishiyama et al.[28], but shows marginal improvement on small models under ERP downsampling settings. Our contributions are threefold. 1)F o r problem formula- tion : To generate more realistic ERP low-resolution images, we propose Fisheye downsampling, which mimics the real-world imaging process. 2)F o r method : Combined with the geometric properties of ERP , we design a distortion-awareTransformer, which modulates distortions continuously and self-adaptively without cumbersome process. 3)F o r data : To reduce overﬁtting, we propose a convenient data aug-mentation strategy, which synthesizes pseudo ERP imagesfrom plain images. Extensive experiments have demon-strated the state-of-the-art performance of our OSRT 2.
Yang_Resource-Efficient_RGBD_Aerial_Tracking_CVPR_2023	Abstract Aerial robots are now able to fly in complex environ- ments, and drone-captured data gains lots of attention in object tracking. However, current research on aerial per- ception has mainly focused on limited categories, such as pedestrian or vehicle, and most scenes are captured in ur- ban environments from a birds-eye view. Recently, UAVs equipped with depth cameras have been also deployed for more complex applications, while RGBD aerial tracking is still unexplored. Compared with traditional RGB ob- ject tracking, adding depth information can more effec- tively deal with more challenging scenes such as target and background interference. To this end, in this paper, we explore RGBD aerial tracking in an overhead space, which can greatly enlarge the development of drone-based visual perception. To boost the research, we first propose a large-scale benchmark for RGBD aerial tracking, contain- ing 1,000 drone-captured RGBD videos with dense annota- tions. Then, as drone-based applications require for real- time processing with limited computational resources, we also propose an efficient RGBD tracker named EMT. Our tracker runs at over 100 fps on GPU, and 25 fps on the edge platform of NVidia Jetson NX Xavier, benefiting from its ef- ficient multimodal fusion and feature matching. Extensive experiments show that our EMT achieves promising track- ing performance. All resources are available at https:// github.com/yjybuaa/RGBDAerialTracking .	1. Introduction Aerial robots have been widely used in complex mis- sions. For example, Unmanned Aerial Vehicles (UA Vs) equipped with cameras are able to perceive and understand unknown environments and have wide applications on agri- culture and surveillance [11, 43]. Specifically, color-based visual tracking with drones has been rapidly developed, thanks to large-scale datasets [27, 43] and dedicated algo- †Equal contribution. ∗Corresponding author.rithms [2–4, 9, 10, 12, 17, 24, 35]. However, these UA Vs merely equipped with color-based sensors generally fail to deal with the challenges in complex environments, such as background clutters and dark scenes, which break the visi- bility and illumination limitations in color-only domain. For example, current drones have difficulties on tracking a per- son in dark scenes. While, RGBD tracking is effective to tackle such kinds of tracking failures. However, for a long time, depth sensors are only incor- porated with UA Vs to enable aerial autonomy and collision avoidance [14]. Visual perception like RGBD tracking with drones is unexplored due to the multiple limitations. For example, commercial RGBD sensors are strictly limited by application scenarios and depth measurement range. On the other hand, we notice that current UA V tracking datasets record video sequences in the manner of aerial photogra- phy [8, 43]. The captured objects mainly focus on pedes- trians and vehicles, and the captured scenes are in urban environments from a birds-eye view. In this work, we explore RGBD aerial tracking from a more practical viewpoint. Different from existing UA V tracking works, we focus on the unexplored overhead space (2 - 5 meters above the ground), aiming to save the ground space greatly with drone-based visual perception. Instead of mainly focusing on people and vehicles, our research can include more generic objects of different categories, such as hands, cups, or balls. Thus, multimodal aerial platforms in this space are very important, as flying robots with short- range perception capabilities can potentially be used in a wider range of scenarios, such as human-robot interaction. Notably, the new task brings challenges in drone-based visual perception, which can be concluded as follows: Complex real-world circumstances. The real-world flight comes with complicated and changeable natural en- vironments. On the one hand, the high mobility of drones brings intense pose changes, resulting in huge variations of target scale and considerable motion blurs. Except for the common challenges in visible situations, drone vision also suffers from other problems like low illumination, similar objects and background clutter. 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. 13374 Limited onboard computational resources. In practi- cal applications, flying platforms generally require higher efficiency on edge platforms with limited resources, while state-of-the-art trackers can only run on powerful GPUs. Especially for multimodal trackers, the model efficiency is always the least valued in model design. Real-time practical applications. Real time is a basic requirement in aerial tracking. Moving platforms require real-time responses and real-world applications also require trackers to function in real-time speed. However, most of current state-of-the-art trackers even cannot achieve real- time speed on powerful GPUs, not to mention their real- world applications. Therefore, to achieve UA V visual tracking with depth, we first build a novel RGBD aerial platform to collect videos. The platform is particularly designed to simulate the environments in real-world applicaitions. The captured videos can comprehensively reflect those challenges to be tackled. Using this aerial platform, a large-scale dataset forDrone-based RGB Daerial tracking, named D2Cube , is built. Some examples in our dataset are given in Fig. 1. In total, 1,000 sequences are provided with dense bound- ing box annotations. The settings of captured videos cover diverse scenarios in daily life. Furthermore, we propose an efficient tracker named EMT to facilitate the development of on-board RGBD tracking. The proposed EMT can be treated as a strong baseline for on-board multimodal tracking to simultane- ously tackle above three issues. Thanks to the efficient mul- timodal fusion and feature matching, our proposed tracker can successfully balance the tight computational budget and tracking accuracy. We perform extensive experiments in di- verse scenarios and various platforms to validate the effec- tiveness of our EMT. Competitive tracking performance is observed in comparison with state-of-the-art RGB-only and multimodal trackers, in which EMT runs at a high frame rate of over 100 FPS. Practical application tests are given onNVIDIA Jetson NX Xavier , where our EMT can run at a frame rate of over 25 FPS. To conclude, our dataset covers complex aerial tracking scenarios and our method shows a promising balance of accuracy, resources and speed. The contributions are summarised below: •New Problem: We propose a new task of RGBD air tracking for newly defined overhead space (2m - 5m). Unlike previous aerial tracking, this task is more rele- vant to human life and has wider applications. •New Benchmark: We construct a large-scale high- diversity benchmark for RGBD aerial tracking. The advantage is that much more categories (34 classes) can be considered than existing aerial tracking datasets. As far as we know, this is the first dataset that can test multimodal aerial tracking models.•New Baseline: An efficient tracking baseline is pro- posed for RGBD aerial tracking, which is the first real- time tracker for efficient on-board multimodal track- ing. It performs better than classical UA V trackers and maintains comparable efficiency.
Zhang_Two-Stage_Co-Segmentation_Network_Based_on_Discriminative_Representation_for_Recovering_Human_CVPR_2023	Abstract Recovering 3D human mesh from videos has recently made signiﬁcant progress. However , most of the existingmethods focus on the temporal consistency of videos, while ignoring the spatial representation in complex scenes, thusfailing to recover a reasonable and smooth human meshsequence under extreme illumination and chaotic back- grounds. To alleviate this problem, we propose a two- stage co-segmentation network based on discriminative rep- resentation for recovering human body meshes from videos. Speciﬁcally, the ﬁrst stage of the network segments the videospatial domain to spotlight spatially ﬁne-grained informa-tion, and then learns and enhances the intra-frame discrimi- native representation through a dual-excitation mechanism and a frequency domain enhancement module, while sup- *represents corresponding author,†represents the equal contribution.pressing irrelevant information (e.g., background). The sec-ond stage focuses on temporal context by segmenting the video temporal domain, and models inter-frame discrimina- tive representation via a dynamic integration strategy. Fur-ther , to efﬁciently generate reasonable human discrimina-tive actions, we carefully elaborate a landmark anchor area loss to constrain the variation of the human motion area. Extensive experimental results on large publicly available datasets indicate superiority in comparison with most state-of-the-art. The Code will be made public.	1. Introduction 3D human mesh recovery from images and videos has been widely concerned in recent years. Existing methodsfor estimating human pose and shape from a single im-age are based on parametric human models such as SMPL [17] etc, which takes a set of model parameters as input 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. 5662 and ﬁnally outputs a human body mesh. These methods capture the statistical information on human body shapeand provide human body mesh for various applications.While these methods recover body mesh from a single im-age [ 3,12,15,16] can accurately predict human pose, they may be jittery and intermittent when applied to videos. The reason for this problem is that the body pose is in-consistent over successive frames and does not reﬂect thebody’s motion in the rapidly changing complex scenes ofthe video. This thus leads to temporal non-smoothness andspatial non-accuracy. Several approaches [ 5,7,12,14,22] have been proposed to efﬁciently extend single image-basedmethods to video. They utilize different temporal encoders to learn the temporal representation directly from videos tobetter capture temporal information. However, these meth- ods only encode spatial features, ignoring the effective uti- lization of spatial ﬁne-grained features and human motion discriminative features. Therefore, it fails to recover a rea-sonable and smooth human sequence in chaotic and extreme illumination scenes. For example, TCMR [ 5] recovers the unsatisfactory motion on the left arm of the actor in Figure1in complex scenes. The background and the human in spatial features have a complex relationship. When spatial features are input tothe network, it is difﬁcult for the network to distinguish be-tween the human body and the background. At the same time, this relationship is not conducive to our discovery ofﬁne-grained and discriminable features. Speciﬁcally, in ex- treme illumination and chaotic scenes, messy background severely interferes with human details and movement infor- mation, thus the network cannot reason about accurate hu-man detail features in complex scenes and lacks the ability to discriminate reasonable human movements. We considerboth intra-frame and inter-frame multi-level spatial repre- sentations are ideal cues to efﬁciently reason about spa-tial ﬁne-grained information and temporal contextual dis- criminative information. In addition, learning to repre- sent features at different stages is expected to strengthenthe model to strip away the complex background and ﬁndhuman-separable motion features, thereby further improv- ing human-speciﬁc discriminative capabilities. Based on the above perspectives, we propose a two-stage co-segmentation network based on discriminative represen- tation for recovering human mesh from videos. In contrast to previous approaches using common spatial features for encoding temporal features, we attempt to segment spatialfeatures into distinct hierarchical of spatial representations and process them separately in different stages. Speciﬁ-cally, the network learns and models intra-frame and inter- frame multi-level discriminative representations by seg- menting spatial features along feature channels and tempo- ral dimensions in two stages. In the ﬁrst stage of the intra-frame discriminative representation, we design a dual exci-tation mechanism that combines self-excitation and channel excitation mechanism to simulate and activate human mo-tion while attenuating the interferences of complex back-grounds. In addition, we design a frequency domain en- hancement module to capture motion information that canhighlight motion features in the frequency domain. In the second stage of inter-frame discriminative representation,we offer a new discriminative representation: the superposi-tion of fragments, which enhances the spatio-temporal rep- resentation of past and future frames by a dynamic integra- tion strategy, while modeling the discriminative represen-tation of the temporal context. Furthermore, to ensure the integrity and plausibility of discriminative motion represen- tation in consecutive frames, we also carefully design a newlandmark anchor area loss to optimize the network, thereby further helping the model to reconstruct accurate 3D human actions and poses. The core contributions of our work are as follows: • We present a co-segmentation network based on dis- criminative representation for recovering human meshfrom videos. Our method motivates and learns spatio- temporal discriminative features at different stages. • In Stage 1, our proposed dual excitation mechanism and frequency domain enhancement effectively en-hance human motion features and mitigate background interference. In Stage 2, we develop a dynamic inte-gration strategy to integrate the discriminative repre-sentations of distinct stages. We also carefully design a landmark anchor area loss to constrain the generationof the reasonable pose. • Both the quantitative and qualitative results of our method show the effectiveness of the proposed method on widely evaluated benchmark datasets in comparisonwith state-of-the-arts.
Xu_MV-JAR_Masked_Voxel_Jigsaw_and_Reconstruction_for_LiDAR-Based_Self-Supervised_Pre-Training_CVPR_2023	Abstract This paper introduces the Masked Voxel Jigsaw and Reconstruction (MV-JAR) method for LiDAR-based self- supervised pre-training and a carefully designed data- efﬁcient 3D object detection benchmark on the Waymo dataset. Inspired by the scene-voxel-point hierarchy in downstream 3D object detectors, we design masking and re- construction strategies accounting for voxel distributions in the scene and local point distributions within the voxel. We employ a Reversed-Furthest-Voxel-Sampling strategy to ad- dress the uneven distribution of LiDAR points and propose MV-JAR, which combines two techniques for modeling the aforementioned distributions, resulting in superior perfor- mance. Our experiments reveal limitations in previous data- efﬁcient experiments, which uniformly sample ﬁne-tuning splits with varying data proportions from each LiDAR se- quence, leading to similar data diversity across splits. To address this, we propose a new benchmark that samples scene sequences for diverse ﬁne-tuning splits, ensuring ad- equate model convergence and providing a more accu- rate evaluation of pre-training methods. Experiments on our Waymo benchmark and the KITTI dataset demonstrate that MV-JAR consistently and signiﬁcantly improves 3D detection performance across various data scales, achiev- ing up to a 6.3% increase in mAPH compared to training from scratch. Codes and the benchmark are available at https://github.com/SmartBot-PJLab/MV-JAR .	1. Introduction Self-supervised pre-training has gained considerable at- tention, owing to its exceptional performance in visual rep- resentation learning. Recent advancements in contrastive Corresponding author.0 5 10 15 20 25 Epoch596061626364656667L2 mAPHAccelerating Convergence MV-JAR Random 0 10 20 30 40 50 Data (%)40444852566064L2 mAPHData-Efficient Results MV-JAR Random Figure 1. 3D object detection results on the Waymo dataset. Our MV-JAR pre-training accelerates model convergence and greatly improves the performance with limited ﬁne-tuning data. learning [ 4,6,8,16,45] and masked autoencoders [ 1,7,15, 36,47] for images have sparked interest among researchers and facilitated progress in modalities such as point clouds. However, LiDAR point clouds differ from images and dense point clouds obtained by reconstruction as they are naturally sparse, unorganized, and irregularly distributed. Developing effective self-supervised proxy tasks for these unique properties remains an open challenge. Construct- ing matching pairs for contrastive learning in geometry- dominant scenes is more difﬁcult [ 20,40], as points or re- gions with similar geometry may be assigned as negative samples, leading to ambiguity during training. To address this, our study explores masked voxel modeling paradigms for effective LiDAR-based self-supervised pre-training. Downstream LiDAR-based 3D object detectors [ 12,19, 33,38,41,50] typically quantize the 3D space into vox- els and encode point features within them. Unlike pix- els, which are represented by RGB values, the 3D space presents a scene-voxel-point hierarchy, introducing new 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. 13445 challenges for masked modeling. Inspired by this, we de- sign masking and reconstruction strategies that consider voxel distributions in the scene and local point distributions in the voxel. Our proposed method, Masked V oxel Jigsaw And Reconstruction (MV-JAR), harnesses the strengths of both voxel and point distributions to improve performance. To account for the uneven distribution of LiDAR points, we ﬁrst employ a Reversed-Furthest-V oxel-Sampling (R- FVS) strategy that samples voxels to mask based on their sparseness. This approach prevents masking the furthest distributed voxels, thereby avoiding information loss in re- gions with sparse points. To model voxel distributions, we propose Masked V oxel Jigsaw (MVJ), which masks the voxel coordinates while preserving the local shape of each voxel, enabling scene reconstruction akin to solving a jigsaw puzzle. For modeling local point distributions, we introduce Masked V oxel Reconstruction (MVR), which masks all coordinates of points within the voxel but retains one point as a hint for reconstruction. Combining these two methods enhances masked voxel modeling. Our experiments indicate that existing data-efﬁcient ex- periments [ 20,40] inadequately evaluate the effectiveness of various pre-training methods. The current benchmarks, which uniformly sample frames from each data sequence to create diverse ﬁne-tuning splits, exhibit similar data di- versity due to the proximity of neighboring frames in a sequence [ 3,14,30]. Moreover, these experiments train models for the same number of epochs across different ﬁne-tuning splits, potentially leading to incomplete conver- gence. As a result, the beneﬁts of pre-trained representa- tions become indistinguishable across splits once the ob- ject detector is sufﬁciently trained on the ﬁne-tuning data. To address these shortcomings, we propose sampling scene sequences to form diverse ﬁne-tuning splits and establish a new data-efﬁcient 3D object detection benchmark on the Waymo [ 30] dataset, ensuring sufﬁcient model convergence for a more accurate evaluation. We employ the Transformer-based SST [ 12] as our de- tector and pre-train its backbone for downstream detec- tion tasks. Comprehensive experiments on the Waymo and KITTI [ 14] datasets demonstrate that our pre-training method signiﬁcantly enhances the model’s performance and convergence speed in downstream tasks. Notably, it im- proves detection performance by 6.3% mAPH when using only 5% of scenes for ﬁne-tuning and reduces training time by half when utilizing the entire dataset (Fig. 1). With the representation pre-trained by MV-JAR, the 3D object de- tectors pre-trained on Waymo also exhibit generalizability when transferred to KITTI.
Zhang_Learning_Debiased_Representations_via_Conditional_Attribute_Interpolation_CVPR_2023	Abstract An image is usually described by more than one attribute like “shape” and “color”. When a dataset is biased, i.e., most samples have attributes spuriously correlated with the target label, a Deep Neural Network (DNN) is prone to make predictions by the “unintended” attribute, especially if it is easier to learn. To improve the generalization abil- ity when training on such a biased dataset, we propose a χ2-model to learn debiased representations. First, we de- sign a χ-shape pattern to match the training dynamics of a DNN and find Intermediate Attribute Samples (IASs) — samples near the attribute decision boundaries, which in- dicate how the value of an attribute changes from one ex- treme to another. Then we rectify the representation with a χ-structured metric learning objective. Conditional interpo- lation among IASs eliminates the negative effect of periph- eral attributes and facilitates retaining the intra-class com- pactness. Experiments show that χ2-model learns debiased representation effectively and achieves remarkable improve- ments on various datasets. Code is available at: https: //github.com/ZhangYikaii/chi-square	1. Introduction Deep neural networks (DNNs) have emerged as an epoch- making technology in various machine learning tasks with impressive performance [5,26]. In some real applications, an object may possess multiple attributes, and some of them are only spuriously correlated to the target label. For example, in Figure 1, the intrinsic attribute of an image annotated by “lifeboats” is its shape . Although there are many lifeboats colored orange, a learner can not make predictions through thecolor ,i.e., there is a misleading correlation from attribute asone containing “orange” color is the target “lifeboats” . When the major training samples can be well discerned by such peripheral attribute, especially learning on it is easier than on the intrinsic one, a DNN is prone to bias towards that “unintended” bias attribute [6, 11, 21, 43, 47, 48, 51], like recognizing a “cyclist” wearing orange as a “lifeboat”. Similar spurious attribute also exists in various applications 97.8% lifeboat(a) Orange lifeboat救生艇_0.5703651905059814 -串联自行车 _0.1379992961883545 -玩具店_0.08981618285179138 -山地自行车_0.04711327701807022 -橄榄球头盔 _0.03929492458701134 57.0% lifeboat 13.8% bicycle -built-for-two 8.98% toyshop 47.1% amphibian 28.1% lifeboat 18.1% speedboat独木舟_0.2981044054031372 -快艇 _0.19624267518520355 -水陆两用车 _0.17111073434352875 -船桨 _0.13951753079891205- 湖边 _0.08463700115680695救生艇_0.9784454107284546 -灯塔 _0.004665153566747904 -集装箱船_0.003798476653173566 -码头 _0.003606501966714859- 消防船 _0.0030068345367908478 (b) Orange cyclists (b) Green lifeboat 0.5% beacon 0.4% container ship57.0% lifeboat 13.8% bicycle -built-for-two 9.0% toyshopFigure 1. Classification of a standard ResNet-50 of (a)an orange lifeboat in the training set (with both color andshape attributes), and(b)an orange cyclist for the test (aligned with color attribute but conflicting with the shape one). Most of the lifeboats in the training set are orange. The biased model is prone to predict via the “unintended” color attribute rather than the intrinsic shape . such as recommendation system [8, 35, 53, 59] and neural language processing [13, 14, 33, 41, 56]. Given such a biased training dataset, how to get rid of the negative effect of the misleading correlations? One in- tuitive solution is to perform special operations on those samples highly correlated to the bias attributes, which re- quires additional supervision, such as the pre-defined bias type [1, 4, 11, 12, 22, 30, 34, 46, 50]. Since prior knowledge of the dataset bias requires expensive manual annotations and is naturally missing in some applications, learning a debiased model without additional supervision about bias is in demand. Nam et al. [36] identify samples with intrin- sic attributes based on the observation that malignant bias attributes are often easier-to-learn than others. Then the valu- able samples for a debiasing scheme could be dynamically reweighted or augmented [11,27,34]. However, the restricted number of such samples
Yang_Global_Vision_Transformer_Pruning_With_Hessian-Aware_Saliency_CVPR_2023	Abstract Transformers yield state-of-the-art results across many tasks. However, their heuristically designed architecture impose huge computational costs during inference. This work aims on challenging the common design philosophy of the Vision Transformer (ViT) model with uniform dimension across all the stacked blocks in a model stage, where we redistribute the parameters both across transformer blocks and between different structures within the block via the first systematic attempt on global structural pruning. Deal- ing with diverse ViT structural components, we derive a novel Hessian-based structural pruning criteria comparable across all layers and structures, with latency-aware regu- larization for direct latency reduction. Performing iterative pruning on the DeiT-Base model leads to a new architec- ture family called NViT (Novel ViT), with a novel parameter redistribution that utilizes parameters more efficiently. On ImageNet-1K, NViT-Base achieves a 2.6×FLOPs reduction, 5.1×parameter reduction, and 1.9×run-time speedup over the DeiT-Base model in a near lossless manner. Smaller NViT variants achieve more than 1%accuracy gain at the same throughput of the DeiT Small/Tiny variants, as well as a lossless 3.3×parameter reduction over the SWIN-Small model. These results outperform prior art by a large margin. Further analysis is provided on the parameter redistribution insight of NViT, where we show the high prunability of ViT models, distinct sensitivity within ViT block, and unique parameter distribution trend across stacked ViT blocks. Our insights provide viability for a simple yet effective parameter redistribution rule towards more efficient ViTs for off-the- shelf performance boost.	1. Introduction Transformer models demonstrate high model capacity, easy scalability, and superior ability in capturing long-range dependency [1, 9, 19, 30, 38]. Vision Transformer, i.e., the ViT [12], shows that embedding image patches into tokens and passing them through a sequence of transformer blocks *Work done during an internship at NVIDIA.can lead to higher accuracy compared to state-of-the-art CNNs. DeiT [35] further presents a data-efficient training method such that acceptable accuracy can be achieved with- out extensive pretraining. Offering competitive performance to CNNs under similar training regimes, transformers now point to the appealing perspective of solving both NLP and vision tasks with the same architecture [18, 20, 49]. Unlike CNNs built with convolutional layers that con- tain few dimensions like the kernel size and the number of filters, the ViT has multiple distinct components, i.e., QKV projection, multi-head attention, multi-layer percep- tron, etc. [38], each defined by independent dimensions. As a result, the dimension of each component in each ViT block needs to be carefully designed to achieve a decent trade-off between efficiency and accuracy. However, this is typically not the case for state-of-the-art models. Models such as ViT [12] and DeiT [35] mainly inherit the design heuristics from NLP tasks, e.g., use MLP expansion ratio 4 ,fix QKV per head ,all the blocks having the same dimensions , etc., which may not be optimal for computer vision [4], caus- ing significant redundancy in the base model and a worse efficiency-accuracy trade-off upon scaling, as extensively shown empirically. New developments in ViT architectures incorporate additional design tricks like multi-stage archi- tecture [41], more complicated attention schemes [23], and additional convolutional layers [13] etc., yet no attempt has been made on understanding the potential of redistributing parameters within the stacked vision transformer blocks. This work targets efficient ViTs by exploring parameter redistribution within ViT blocks and across multiple layers of cascading ViT blocks. To this end, we start with the straight- forward DeiT design space, with only ViT blocks. We ana- lyze the importance and redundancy of different components in the DeiT model via latency-aware global structural prun- ing, leveraging the insights to redistribute parameters for enhanced accuracy-efficiency trade-off. Our approach, as visualized in Fig. 1, starts from analyzing the blocks in the computation graph of ViT to identify all the dimensions that can be independently controlled. We apply global structural pruning over all the components in all blocks. This offers complete flexibility to explore their combinations towards 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. 18547 MLP headC classes QQQPrunable weights QQKQQVPROJPROJProj.FC1 FC2 ExQKxHExQKxHE x VVxExHExMMxEVision T ransformer Prunable Components Analysis NViTInput tokensNViT Block NViT Block NViT BlockNViT BlockNViT BlockNViT BlockMLP headC classes Pruned weights FC10.7Ex0.6QK x0.5H0.7Ex0.9Vx0.5H0.9Vx0.7Ex0.5H0.7Ex0.6M0.6M x0.7E Global Importance RankingBlocks / LayersImportanceThres. FC1FC2 PROJProj. QQ KK VVParameter RedistributionBlocks LayersDimension Pruned modelTrendInput tokensH heads Ex(VxH)Q K VEx(QKxH) khqh vhkhqh vhkhqh vh LayerNorm Input tokens NxEProj. (VxH)xEMultihead Self Attention (MSA) Output tokens NxEConcat splitEx(QKxH) Transformer Blocks x12 ExM MxEMulti-layer Perceptron (MLP)FC 2 LayerNormFC 1 0.7Ex0.6QK x0.5H Global Structural PruningTowards Efficient InferenceFigure 1. Towards efficient vision transformer models. Starting form ViT, specifically DeiT, we identify the design space of pruning (i) embedding size E, (ii) number of head H, (iii) query/key size QK, (iv) value size V and (v) MLP hidden dimension M in Sec. 3.1. Then we utilize a global ranking of latency-aware importance score to perform iterative global structural pruning in Sec. 3.2, achieving pruned NViT models. Finally we analyze the parameter redistribution trend of all the components in the NViT model, as in Sec. 5.1. an optimal architecture in a complicated design space. Per- forming global pruning on ViT is significantly challenging, given the diverse structural components and significant mag- nitude differences. Previous methods only attempts on per- component pruning with the same pruning ratio [5], which cannot lead to parameter redistribution across components and blocks. We derive a new importance score based on the Hessian matrix norm of the loss for global structural pruning, for the first time offering comparability among all prunable components. Furthermore, we incorporate the estimated la- tency reduction into the importance score. This guides the final pruned architecture to be faster on target devices. The iterative structural pruning of the DeiT-Base model enables a family of efficient ViT models: NViT. On the ImageNet-1K benchmark [33], NViT enables a nearly loss- less 5.14 ×parameter reduction, 2.57 ×FLOPs reduction and 1.86 ×speed up on V100 GPU over the DeiT-Base model. An 1% and 1.7% accuracy gain is observed over DeiT-Small and DeiT-Tiny models when we scale down the NViT to a similar latency. NViT achieves a further 1.8 × FLOPs reduction and an 1.5 ×speedup over NAS-based Aut- oFormer [4] (ICCV’21) and the SOTA structural pruning method S2ViTE [5] (NeurIPS’21).The efficiency and perfor- mance benefit of NViT trained on ImageNet also transfers to downstream classification and segmentation tasks. Using structural pruning for architectural guidance, we further make an important observation that the popular uni- form distribution of parameters across all layers is, in fact, not optimal. To this end, we present further empirical andtheoretical analysis on the new parameter distribution rule of efficient ViT architectures, which provides a new angle on understanding the learning dynamic of vision transformer model. We believe our findings would inspire future design of efficient ViT architectures. Our main contributions are as follows: •Propose NViT, a novel hardware-friendly global struc- tural pruning algorithm enabled by a latency-aware , Hessian-based importance-based criteria and tailored towards the ViT architecture, achieving a nearly loss- less 1.9 ×speedup, significantly outperforms SOTA ViT compression methods and efficient ViT designs; •Provide a systematic analysis on the prunable compo- nents in the ViT model. We perform structural pruning on the embedding dimension, number of heads, MLP hidden dimension, QK dimension and V dimension of each head separately; •Explore hardware-friendly parameter redistribution of ViT, finding high prunability of ViT models, distinct sensitivity within ViT block, and unique parameter distribution trend across stacked ViT blocks.
Zhang_Complete-to-Partial_4D_Distillation_for_Self-Supervised_Point_Cloud_Sequence_Representation_Learning_CVPR_2023	Abstract Recent work on 4D point cloud sequences has attracted a lot of attention. However, obtaining exhaustively labeled 4D datasets is often very expensive and laborious, so it is especially important to investigate how to utilize raw unla- beled data. However, most existing self-supervised point cloud representation learning methods only consider ge- ometry from a static snapshot omitting the fact that se- quential observations of dynamic scenes could reveal more comprehensive geometric details. To overcome such is- sues, this paper proposes a new 4D self-supervised pre- training method called Complete-to-Partial 4D Distillation. Our key idea is to formulate 4D self-supervised represen- tation learning as a teacher-student knowledge distilla- tion framework and let the student learn useful 4D repre- sentations with the guidance of the teacher. Experiments show that this approach significantly outperforms previ- ous pre-training approaches on a wide range of 4D point cloud sequence understanding tasks. Code is available at: https://github.com/dongyh20/C2P.	1. Introduction Recently, there is a surge of interest in understanding point cloud sequences in 4D (3D space + 1D time) [7, 8, 11, 21, 30]. As the direct sensor input for a wide range of applications including robotics and augmented reality, point cloud sequences faithfully depict a dynamic environ- ment regarding its geometric content and object movements in the context of the camera ego-motion. Though widely accessible, such 4D data is prohibitively expensive to an- notate in large scale with fine details. As a result, there is a strong need for leveraging the colossal amount of un- labeled sequences. Among the possible solutions, self- supervised representation learning has shown its effective- ness in a wide range of fields including images [2, 12, 13], videos [6, 9, 15, 22, 28] and point clouds [16, 24, 31, 33, 34]. *Equal contribution with the order determined by rolling dice. †Corresponding author. Figure 1. We propose a complete-to-partial 4D distillation (C2P) approach. Our key idea is to formulate 4D self-supervised rep- resentation learning as a teacher-student knowledge distillation framework in which students learn useful 4D representations un- der the guidance of a teacher. The learned features can be trans- ferred to a range of 4D downstream tasks. We therefore aim to fill in the absence of self-supervised point cloud sequence representation learning in this work. Learning 4D representations in a self-supervised man- ner seems to be a straightforward extension of 3D cases. However, a second thought reveals its challenging nature since such representations need to unify the geometry and motion information in a synergetic manner. From the ge- ometry aspect, a 4D representation learner needs to under- stand 3D geometry in a dynamic context. However, most existing self-supervised point cloud representation learn- ing methods [16, 24, 34] only consider geometry from a static snapshot, omitting the fact that sequential observa- tions of dynamic scenes could reveal more comprehensive geometric details. From the motion aspect, a 4D represen- tation learner needs to understand motion in the 3D space, which requires an accurate cross-time geometric associa- tion. Nevertheless, existing video representation learning frameworks [6, 9, 22] mostly model motion as image space flows so geometric-aware motion cues are rarely encoded. Due to such challenges, 4D has been rarely discussed in the self-supervised representation learning literature, with only a few works [4, 23] designing learning objectives in 4D to facilitate static 3D scene understanding. To address the above challenges, we examine the na- ture of 4D dynamic point cloud sequences, and draw two 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. 17661 main observations. First, most of a point cloud sequence depicts the same underlying 3D content with an optional dynamic motion. Motion understanding could help aggre- gate temporal observations to form a more comprehensive geometric description of the scene. Second, geometric cor- respondences across time could help estimate the relative motion between two frames. Therefore better geometric un- derstanding should facilitate a better motion estimation. At the core are the synergetic nature of geometry and motion. To facilitate the synergy of geometry and motion, we de- velop a Complete-to-Partial 4D Distillation (C2P) method. Our key idea is to formulate 4D self-supervised represen- tation learning as a teacher-student knowledge distillation framework and let the student learn useful 4D representa- tions with the guidance of the teacher. And we present a unified solution to the following three questions: How to teach the student to aggregate sequential geometry for more complete geometric understanding leveraging motion cues? How to teach the student to predict motion based upon bet- ter geometric understanding? How to form a stable and high-quality teacher? In particular, our C2P method consists of three key de- signs. First, we design a partial-view 4D sequence genera- tion method to convert an input point cloud sequence which is already captured partially into an even more partial se- quence. This is achieved by conducting view projection of input frames following a generated camera trajectory. The generated partial 4D sequence allows bootstrapping multi- frame geometry completion. This is achieved by feeding the input sequence and the generated partial-view sequence to teacher and student networks respectively and distill the teacher knowledge to a 4D student network. Second, the student network not only needs to learn completion by mim- icking the corresponding frames of the teacher network, but also needs to predict the teacher features of other frames within a time window, to achieve so-called 4D distillation. Notice the teacher feature corresponds to more complete ge- ometry, which also encourages the student to exploit the benefit of geometry completion in motion prediction. Fi- nally, we design an asymmetric teacher-student distillation framework for stable training and high-quality representa- tion, i.e., the teacher network has weaker expressivity com- pared with the student but is easier to optimize. We evaluate our method on three downstream tasks including indoor and outdoor scenarios: 4D action seg- mentation on HOI4D [21], 4D semantic segmentation on HOI4D [21], 4D semantic segmentation on Synthia 4D [27] and 3D action recognition on MSR-action3D [17]. We demonstrate significant improvements over the previous method( +2.5%accuracy on HOI4D action segmentation, +1.0%mIoU on HOI4D semantic segmentation, +1.0% mIoU on Synthia 4D semantic segmentation and +2.1%ac- curacy on MSR-Action3D).The contributions of this paper are fourfold: First, we propose a new 4D self-supervised representation learning method named Complete-to-Partial 4D Distillation which facilitates the synergy of geometry and motion learning. Second, we propose a natural and effective way to generate partial-view 4D sequences and demonstrate that it can work well as learning material for knowledge distillation. Third, we find that asymmetric design is crucial in the complete- to-partial knowledge distillation process and we propose a new asymmetric distillation architecture. Fourth, extensive experiments on three tasks show that our method outper- forms previous state-of-the-art methods by a large margin.
Yu_Boost_Vision_Transformer_With_GPU-Friendly_Sparsity_and_Quantization_CVPR_2023	Abstract The transformer extends its success from the language to the vision domain. Because of the stacked self-attention and cross-attention blocks, the acceleration deployment of vi- sion transformer on GPU hardware is challenging and also rarely studied. This paper thoroughly designs a compres- sion scheme to maximally utilize the GPU-friendly 2:4 fine- grained structured sparsity and quantization . Specially, an original large model with dense weight parameters is first pruned into a sparse one by 2:4 structured pruning, which considers the GPU’s acceleration of 2:4 structured sparse pattern with FP16 data type, then the floating-point sparse model is further quantized into a fixed-point one by sparse- distillation-aware quantization aware training, which con- siders GPU can provide an extra speedup of 2:4 sparse cal- culation with integer tensors. A mixed-strategy knowledge distillation is used during the pruning and quantization pro- cess. The proposed compression scheme is flexible to sup- port supervised and unsupervised learning styles. Exper- iment results show GPUSQ-ViT scheme achieves state-of- the-art compression by reducing vision transformer models 6.4-12.7 ×on model size and 30.3-62 ×on FLOPs with neg- ligible accuracy degradation on ImageNet classification, COCO detection and ADE20K segmentation benchmarking tasks. Moreover, GPUSQ-ViT can boost actual deployment performance by 1.39-1.79 ×and3.22-3.43 ×of latency and throughput on A100 GPU, and 1.57-1.69 ×and2.11-2.51 × improvement of latency and throughput on AGX Orin.	1. Introduction Transformer-based neural models [48] have garnered im- mense interest recently due to their effectiveness and gen- eralization across various applications. Equipped with the attention mechanism [52] as the core of its architecture, transformer-based models specialize in handling long-range dependencies, which are also good at extracting non-local *Tao Chen and Zhongxue Gan are corresponding authors.features [9] [5] in the computer vision domain. With com- parable and even superior accuracy than the traditional con- volution neural networks (CNN) [12] [49], more vision transformer models are invented and gradually replace the CNN with state-of-the-art performance on image classifi- cation [27] [26], object detection [70] [59], and segmenta- tion [58] [68] tasks. Due to the vision transformer models having a generally weaker local visual inductive bias [9] in- herent in CNN counterparts, many transformer blocks are stacked for compensation. Moreover, the attention module in the transformer block contains several matrix-to-matrix calculations between key, query, and value parts [52]. Such designs give the naive vision transformers more parameters and higher memory and computational resource require- ments, causing high latency and energy consuming during the inference stage. It is challenging for actual acceleration deployment in GPU hardware . Model compression techniques to transfer the large-scale vision transformer models to a lightweight version can bring benefits to more efficient computation with less on- device memory and energy consumption. There are some previous studies to inherit CNN compression methods, in- cluding pruning [43] [15], quantization [28] [23], distilla- tion [61], and architecture search [6] on vision transformers. However, there are some drawbacks in previous studies: • Most of these common methods aim to reduce the theoretical model size and Floating Point Operations (FLOPs). But it has been proved [33] [37] that smaller model sizes and FLOPs are not directly proportional to better efficiency on deployed hardware. • The compression patterns do not match hardware char- acteristics. For example, pruned [43] or searched [6] vision transformer models have the unstructured sparse pattern in weight parameters, i.e., the distribution of non-zero elements is random. So deployed hardware can not provide actual speedup due to lacking the char- acteristics support for unstructured sparsity [35]. • How to keep the accuracy to the best with multiple compression methods and how to generalize on mul- tiple vision tasks lack systematical investigation. 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. 22658 Sparse M✕N✕K GEMMDense M✕N✕K GEMMKA matrix (Dense)☓Accumulator (result)NDense operation on Tensor Core MB matrix (Dense) C matrix (Dense)MK K/2A matrix (Sparse)Non-zero data values2-bits indicesK/2☓Accumulator (result)Sparse operation on Tensor CoreSelectB matrix (Dense) C matrix (Dense)NChoose matching K/2 elements out of K elements MMKFigure 1. Comparison of computing a M×N×KGEMM onto a Tensor Core. Dense matrix A of size M×Kinleft side becomes M×K 2inright side after compressing with 2:4 fine-grained structured sparse pattern . Sparse Tensor Core automatically picks only the elements from B according to the nonzero elements in A. Comparing the dense and sparse GEMMs, B and C are the same dense K×NandM×Nmatrices, respectively. By skipping the unnecessary multiplications of redundant zeros, sparse GEMM ac- celerate the dense GEMM with 2×. General Matrix Multiplication (GEMM) is the funda- mental implementation inside the common parts of vision transformers, such as convolution, linear projection, and transformer blocks. A specific acceleration unit called Tensor Core [39] is firstly introduced in NVIDIA V olta GPU [34] to accelerate these GEMM instructions and further enhanced to support sparse GEMM in Ampere GPU [35]. To make the GPU hardware efficient for sparse GEMM, a constraint named 2:4 fine-grained structured sparsity [33] is imposed on the allowed sparsity pattern, i.e., two values from every four contiguous elements on rows must be zero. Due to the 2:4 sparsity support on GPU Tensor Core hardware, sparse GEMM can reduce memory storage and bandwidth by almost 2×and provide 2×math throughput compared to dense GEMM by skipping the re- dundant zero-value computation, as shown in Figure 1. Am- pere GPU supports various numeric precision for 2:4 spar- sity, including FP32, FP16, INT8, and INT4, etc. Inspired by GPU’s acceleration characteristic for 2:4 fine-grained structured sparse pattern with various low- precision operators, we thoroughly design the compres- sion scheme GPUSQ-ViT by utilizing the GPU -friendly Sparsity and Quantization to boost deployment efficacy for Vision Transformer models, especially on GPU platforms. GPUSQ-ViT contains two main workflows. Firstly, 2:4 sparse pruning with knowledge distillation [14] (KD) is pro- posed to compress the specific structures in vision trans- former architecture, e.g., transformer block, patch embed- ding, to be GPU-friendly. Secondly, we further quantize the sparse model through sparse-distillation-aware Quanti- zation Aware Training [30] (QAT). To measure the influence of quantization errors, we use the feature-based distillation loss in the sparse pruning workflow as the weight factor. The feature-based KD utilizes the scale factor in the quan- tization compression workflow, which can best compensatefor the final compressed model’s accuracy. We demonstrate thatGPUSQ-ViT can generally apply to vision transformer models and benchmarking tasks, with state-of-the-art the- oretical metrics on model size and FLOPs. Moreover, as GPUSQ-ViT compresses with GPU-friendly patterns, the compressed models can achieve state-of-the-art deployment efficacy on GPU platforms. Our main contributions include: • Unlike previous compression methods only aiming at reducing theoretical metrics, we propose GPUSQ-ViT from the perspective of GPU-friendly 2:4 sparse pat- tern with low-precision quantization for the first time, achieving GPU acceleration of 4 times than prior arts. •GPUSQ-ViT combines feature-based KD with sparse pruning and QAT, which can best compensate for sparse and quantized models’ accuracy. •GPUSQ-ViT can apply to various vision transformer models and benchmarking tasks, with proven state-of- the-art efficacy on model size, FLOPs, and actual de- ployment performance on multiple GPUs. Moreover, GPUSQ-ViT can work without ground truth label an- notations in an unsupervised learning style.
Yang_IDGI_A_Framework_To_Eliminate_Explanation_Noise_From_Integrated_Gradients_CVPR_2023	Abstract Integrated Gradients (IG) as well as its variants are well- known techniques for interpreting the decisions of deep neu- ral networks. While IG-based approaches attain state-of- the-art performance, they often integrate noise into their ex- planation saliency maps, which reduce their interpretabil- ity. To minimize the noise, we examine the source of the noise analytically and propose a new approach to re- duce the explanation noise based on our analytical find- ings. We propose the Important Direction Gradient Integra- tion (IDGI) framework, which can be easily incorporated into any IG-based method that uses the Reimann Integra- tion for integrated gradient computation. Extensive exper- iments with three IG-based methods show that IDGI im- proves them drastically on numerous interpretability met- rics. The source code for IDGI is available at https: //github.com/yangruo1226/IDGI .	1. Introduction With the deployment of deep neural network (DNN) models for safety-critical applications such as autonomous driving [5–7] and medical diagnostics [10, 24], explaining the decisions of DNNs has become a critical concern. For humans to trust the decision of DNNs, not only the model must perform well on the specified task, it also must gen- erate explanations that are easy to interpret. A series of ex- planation methods (e.g., gradient-based saliency/attribution map approaches [21, 22, 29, 33, 36, 38, 43, 46] as well as many that are not based on gradients [4, 11, 13, 19, 25, 27, 30, 32, 35, 39, 40, 42, 47, 48]) have been developed to con- nect a DNN’s prediction to its input. Among them, In- tegrated Gradients (IG) [43], a well-known gradient-based explanation method, and its variants [22, 46] have attracted significant interest due to their state-of-the-art explanation performance and desirable theoretical properties. However, we observe that explanation noise exists in the attribution generated by these IG methods (please see Fig. 1). In this research, we investigate IG-based methods, study the expla- nation noise introduced by these methods, propose a frame- Figure 1. Saliency/attribution map of the existing IG-based meth- ods and those with our method on explaining the prediction from InceptionV3 model. Our method significantly reduces the noise in the saliency map created by these IG-based methods. work to remove the explanation noise, and empirically val- idate the effectiveness of our approach. A few recent IG-based methods (e.g., [38] [22], [46], [41]) have been proposed to address the noise issue. Kapishnikov et al. [22] provide the following main reasons1 that could generate the noise: 1) DNN model’s shape often has a high curvature; and 2) The choice of the reference point impacts explanation. They propose Guided Integrated Gradients (GIG) [22], which tackles point #1 by iteratively finding the integration path that tries to avoid the high cur- vature points in the space. Blur Integral Gradients [46], on the other hand, shows that the noise could be reduced by finding the integration path through the frequency domain instead of the original image domain. Formally, it finds the path by successively blurring the input via a Gaussian blur filter. Sturmfels et al. [41] tackle point #2 by performing the integration from multiple reference points, while Smilkov et al. [38] aggregate the attribution with respect to multiple Gaussian noisy inputs to reduce the noise. Nevertheless, all IG-based methods share a common point in that they compute the integration of gradients via the Riemann inte- gral. We highlight that, the integration calculation by the existing methods fundamentally introduces the explanation noise. To this end, we offer a general solution that elimi- 1[22] mentions the accuracy of integration is also a reason to generate the noise, but this is not the focus of existing IG methods and this paper. 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. 23725 nates the noise by examining the integration directions from the explanation perspective. Specifically, we investigate each computation step in the Riemann Integration and then theorize about the noises’ ori- gin. Each Riemann integration calculation integrates the gradient in the original direction —it first computes the gra- dient with respect to the starting point of the current path segment and then multiplies the gradient by the path seg- ment. We show that the original direction can be divided into an important direction and a noise direction . We theo- retically demonstrate that the true gradient is orthogonal to thenoise direction , resulting in the gradient’s multiplication along the noise direction having no effect on the attribution. Based on this observation, we design a framework, termed Important Direction Gradient Integration (IDGI), that can eliminate the explanation noise in each step of the compu- tation in any existing IG method. Extensive investigations reveal that IDGI reduces noise significantly when evaluated using state-of-the-art IG-based methods. In summary, our main contributions are as follows: • We propose the Important Direction Gradient Integra- tion (IDGI), a general framework to eliminate the ex- planation noise in IG-based methods, and investigate its theoretical properties. • We propose a novel measurement for assessing the at- tribution techniques’ quality, i.e., AIC and SIC using MS-SSIM. We show that this metric offers a more pre- cise measurement than the original AIC and SIC. • Our extensive evaluations on 11 image classifiers with 3 existing and 1 proposed attribution assessment tech- niques indicate that IDGI significantly improves the at- tribution quality over the existing IG-based methods.
Zhang_3D_Registration_With_Maximal_Cliques_CVPR_2023	Abstract As a fundamental problem in computer vision, 3D point cloud registration (PCR) aims to seek the optimal pose to align a point cloud pair. In this paper, we present a 3D reg- istration method with maximal cliques (MAC). The key in- sight is to loosen the previous maximum clique constraint, and mine more local consensus information in a graph for accurate pose hypotheses generation: 1) A compatibility graph is constructed to render the affinity relationship be- tween initial correspondences. 2) We search for maximal cliques in the graph, each of which represents a consensus set. We perform node-guided clique selection then, where each node corresponds to the maximal clique with the great- est graph weight. 3) Transformation hypotheses are com- puted for the selected cliques by the SVD algorithm and the best hypothesis is used to perform registration. Ex- tensive experiments on U3M, 3DMatch, 3DLoMatch and KITTI demonstrate that MAC effectively increases registra- tion accuracy, outperforms various state-of-the-art meth- ods and boosts the performance of deep-learned methods. MAC combined with deep-learned methods achieves state- of-the-art registration recall of 95.7% / 78.9% on 3DMatch / 3DLoMatch.	1. Introduction Point cloud registration (PCR) is an important and fun- damental problem in 3D computer vision and has a wide range of applications in localization [13], 3D object detec- tion [17] and 3D reconstruction [25]. Given two 3D scans of the same object (or scene), the goal of PCR is to estimate a six-degree-of-freedom (6-DoF) pose transformation that accurately aligns the two input point clouds. Using point- to-point feature correspondences is a popular and robust so- lution to the PCR problem. However, due to the limita- tions of existing 3D keypoint detectors & descriptors, the limited overlap between point clouds and data noise, corre- *Corresponding author. Code will be available at https://github.com/zhangxy0517/ 3D-Registration-with-Maximal-Cliques . (a) #corr: 5182, inlier ratio: 8.97%(b) #corr in maximal clique: 4, inlier ratio: 100%(c) RE=4.12° , TE=12.88cm (b) #corr in maximum clique: 6, inlier ratio: 0%(c) RE=12.59° , TE=63.04cm Success FailFigure 1. Comparison of maximal andmaximum cliques on a low overlapping point cloud pair. Maximal cliques (MAC) effec- tively choose the optimal 6-DoF transformation hypothesis with low rotation error (RE) and translation error (TE) for two point clouds with a low inlier ratio, while the maximum clique fails in this case. spondences generated by feature matching usually contain outliers, resulting in great challenges to accurate 3D regis- tration. The problem of 3D registration by handling correspon- dences with outliers has been studied for decades. We classify them into geometric-only and deep-learned meth- ods. For geometric-only methods [5, 6, 21, 30, 31, 38–41], random sample consensus (RANSAC) and its variants per- form an iterative sampling strategy for registration. Al- though RANSAC-based methods are simple and efficient, their performance is highly vulnerable when the outlier rate increases, and it requires a large number of iterations to ob- tain acceptable results. Also, a series of global registra- tion methods based on branch-and-bound (BnB) are pro- posed to search the 6D parameter space and obtain the op- timal global solution. The main weakness of these meth- ods is the high computational complexity, especially when the correspondence set is of a large magnitude and has an extremely high outlier rate. For deep-learned methods, some [1–4, 9, 10, 14, 16, 18, 19, 27, 35] focus on improving 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. 17745 one module in the registration process, such as investigating more discriminate keypoint feature descriptors or more ef- fective correspondence selection techniques, while the oth- ers [22, 29, 43] focus on registration in an end-to-end man- ner. However, deep-learned based methods require a large amount of data for training and usually lack generalization on different datasets. At present, it is still very challenging to achieve accurate registrations in the presence of heavy outliers and in cross-dataset conditions. In this paper, we propose a geometric-only 3D registra- tion method based on maximal cliques (MAC). The key in- sight is to loosen the previous maximum clique constraint, and mine more local consensus information in a graph to generate accurate pose hypotheses. We first model the ini- tial correspondence set as a compatibility graph, where each node represents a single correspondence and each edge be- tween two nodes indicates a pair of compatible correspon- dences. Second, we search for maximal cliques in the graph and then use node-guided clique filtering to match each graph node with the appropriate maximal clique containing it.Compared with the maximum clique, MAC is a looser constraint and is able to mine more local information in a graph. This helps us to achieve plenty of correct hypothe- ses from a graph. Finally, transformation hypotheses are computed for the selected cliques by the SVD algorithm. The best hypothesis is selected to perform registration us- ing popular hypothesis evaluation metrics in the RANSAC family. To summarize, our main contributions are as fol- lows: • We introduce a hypothesis generation method named MAC. Our MAC method is able to mine more local information in a graph, compared with the previous maximum clique constraint. We demonstrate that hy- potheses generated by MAC are of high accuracy even in the presence of heavy outliers. • Based on MAC, we present a novel PCR method, which achieves state-of-the-art performance on U3M, 3DMatch, 3DLoMatch and KITTI datasets. Notably, our geometric-only MAC method outperforms several state-of-the-art deep learning methods [3, 9, 19, 27]. MAC can also be inserted as a module into multiple deep-learned frameworks [1, 10, 18, 29, 43] to boost their performance. MAC combined with GeoTrans- former achieves the state-of-the-art registration recall of95.7% / 78.9% on 3DMatch / 3DLoMatch.
Zeng_CLIP2_Contrastive_Language-Image-Point_Pretraining_From_Real-World_Point_Cloud_Data_CVPR_2023	Abstract Contrastive Language-Image Pre-training, benefiting from large-scale unlabeled text-image pairs, has demon- strated great performance in open-world vision understand- ing tasks. However, due to the limited Text-3D data pairs, adapting the success of 2D Vision-Language Models (VLM) to the 3D space remains an open problem. Existing works that leverage VLM for 3D understanding generally resort to constructing intermediate 2D representations for the 3D data, but at the cost of losing 3D geometry information. To take a step toward open-world 3D vision understand- ing, we propose Contrastive Language- Image-Point Cloud Pretraining (CLIP2) to directly learn the transferable 3D point cloud representation in realistic scenarios with a novel proxy alignment mechanism. Specifically, we exploit naturally-existed correspondences in 2D and 3D scenarios, and build well-aligned and instance-based text-image-point proxies from those complex scenarios. On top of that, we propose a cross-modal contrastive objective to learn se- mantic and instance-level aligned point cloud representa- ∗Equal contribution.1Huawei Noah’s Ark Lab2Hong Kong University of Science and Technology3The Chinese University of Hong Kong4Sun Yat-san University†Corresponding Author: xu.hang@ huawei.comtion. Experimental results on both indoor and outdoor sce- narios show that our learned 3D representation has great transfer ability in downstream tasks, including zero-shot and few-shot 3D recognition, which boosts the state-of-the- art methods by large margins. Furthermore, we provide analyses of the capability of different representations in real scenarios and present the optional ensemble scheme.	1. Introduction Powerful 3D point cloud representation plays a cru- cial role in various real-world applications, e.g., 3D object recognition and detection [10, 20, 31, 40, 44]. Compared to 2D images, 3D point cloud provides specific informa- tion like accurate geometry that is robust to illumination changes. However, current methods [25, 40] that learn 3D representations generally rely on the predefined number of object categories and require plenty of labor-intensive an- notations. Those learned 3D representations are insufficient for safety-critical scenarios like self-driving which includes a long-tail class distribution far beyond the predefined tax- onomy. Therefore, it is highly demanded to learn a transfer- able 3D representation equipped with zero-shot recognition ability in vocabulary scalable real-world scenes. Figure 1 shows an open-world recognition example by our CLIP2in 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. 15244 outdoor and indoor scenes, where the 3D objects can be classified with the correlation alignment between 3D repre- sentations and open-world vocabularies. The critical ingredient of open-world understanding is that the models learn sufficient knowledge to obtain general representations. To achieve this, recent Vision-Language Models (VLM) [14, 27, 38] leverage Internet-scale text- image pairs to conduct vision-language pretraining, which facilitates transferable 2D representation and demonstrates promising performance in 2D open-vocabulary tasks. How- ever, 3D vision-language pretraining remains unexplored due to the limitation of existing 3D datasets in diversity and scale compared to the massive data sources in 2D counter- parts [14,15,27,38]. Though some recent works [12,13,43] try to avoid this problem by transferring the pretrained 2D VLM into the intermediate representation including pro- jected image patches [12, 18] or depth maps [1, 43], those representations suffer from the loss of 3D geometric infor- mation and limited viewpoints under realistic scenarios. Es- pecially the camera images are only sometimes available due to the sensor failure in 3D scenes. We believe the 3D representation based on original point cloud data retains most information and is the optimal solution for 3D real world understanding, which requires a rethink of learning the transferable 3D representation under realistic scenarios. To this end, we propose a Contrastive Language- Image- Point cloud Pretraining framework, short for CLIP2, which directly aligns 3D space with broader raw text and advances the 3D representation learning into an open-world era. Our learning process can be decomposed into two stages: Firstly, we introduce a Triplet Proxy Collection to alleviate the limitation of accessible pretraining data by construct- ing language-image-point triplets from real-world scenes. Since the large-scale realistic 3D datasets for outdoor driv- ing [2,19] and indoor scenarios [9,32] are collected in open- world, it contains huge amounts of realistic objects that vary in semantics and diversity. Thus we consider them as potential pretraining data sources without extra human supervision. Specifically, we propose “Proxy” instances as the bridges between language descriptions, 2D images and 3D point clouds. Enabled by a well-aligned VLM, a scal- able caption list and the geometry transformation between 2D and 3D, we automatically create more than 1 million triplets to facilitate pretraining. Secondly, we further pro- pose a Cross-Modal Pretraining scheme to jointly optimize the feature space alignments of three modalities, i.e.point cloud, language and image. It contains both the contrastive learning objective of semantic-level text-3D correlation and instance-level image-3D correlation, which contributes to better transferability of learned 3D representation. We study the transferable capability of CLIP2by bench- marking the zero-shot recognition performance on four pop- ular indoor and outdoor real-world datasets, and find a sig- nificant improvement over current methods, achieving Top1 accuracy 61.3% on SunRGBD [32], 43.8% on ScanNet [9]),28.8% on nuScenes [2] and 56.0% on ONCE [19]. For a fair comparison with existing methods [1, 13, 36, 43], we conduct zero-shot and few-shot classification on single ob- ject dataset ScanObjectNN [34] and find consistent dom- inance, 16.1% relative improvement on zero-shot classifi- cation over previous state-of-the-art method [13]. To vali- date the vocabulary-increasing ability of CLIP2, we report the quantity results and visualizations to show the improved discovery of the long-tail categories. Moreover, we make ablations and analisis on different representations, and in- vestigate ensembling alternatives to merge complementary knowledge of all available representations in realistic appli- cations. Our contributions can be summarized as follows: • We propose a novel CLIP2framework that aligns 3D space with open-world language representation, facili- tating zero-shot transfer in realistic scenarios. • We present a Triplet Proxies Collection scheme in real- world scenes, which alleviates the shortage of text-3D data sources and facilitates the pretraining methods. • CLIP2jointly optimizes the correlation alignment be- tween point cloud, language and image by proposed cross-modal pretraining mechanism, which enhances the transferability of learned 3D representation. • Our CLIP2achieves the state-of-the-art zero-shot transfer performance on 5 datasets (indoor/outdoor scenes and single-object) and shows quality results on vocabulary-increasing discovery in real world.
Yu_Foundation_Model_Drives_Weakly_Incremental_Learning_for_Semantic_Segmentation_CVPR_2023	Abstract Modern incremental learning for semantic segmentation methods usually learn new categories based on dense anno- tations. Although achieve promising results, pixel-by-pixel labeling is costly and time-consuming. Weakly incremen- tal learning for semantic segmentation (WILSS) is a novel and attractive task, which aims at learning to segment new classes from cheap and widely available image-level labels. Despite the comparable results, the image-level labels can not provide details to locate each segment, which limits the performance of WILSS. This inspires us to think how to im- prove and effectively utilize the supervision of new classes given image-level labels while avoiding forgetting old ones. In this work, we propose a novel and data-efficient frame- work for WILSS, named FMWISS. Specifically, we propose pre-training based co-segmentation to distill the knowl- edge of complementary foundation models for generating dense pseudo labels. We further optimize the noisy pseudo masks with a teacher-student architecture, where a plug- in teacher is optimized with a proposed dense contrastive loss. Moreover, we introduce memory-based copy-paste augmentation to improve the catastrophic forgetting prob- lem of old classes. Extensive experiments on Pascal VOC and COCO datasets demonstrate the superior performance of our framework, e.g., FMWISS achieves 70.7% and 73.3% in the 15-5 VOC setting, outperforming the state-of-the-art method by 3.4% and 6.1%, respectively.	1. Introduction Semantic segmentation is a fundamental task in com- puter vision and has witnessed great progress using deep learning in the past few years. It aims at assigning each pixel a category label. Modern supervised semantic seg- mentation methods [12, 14] are usually based on published large-scale segmentation datasets with pixel annotations. Despite the promising results, one model pre-trained on one 框架区别 Image“Horse”“a photo of Horse”FoundationModels(a) WILSS(b) FMWISS (Ours)Image-level label ofStep t<latexit sha1_base64="6EAb1KoGHnaQ8cLUAdO5OOMiX8M=">AAAC0HicjVHLSsNAFD2Nr1pfVZdugkVwVVIRdVnsxqWKVaGtMhmnOjh5mEzEUoq49Qfc6leJf6B/4Z0xBR+ITkhy5tx7zsy914+VTLXnvRSckdGx8YniZGlqemZ2rjy/cJhGWcJFk0cqSo59lgolQ9HUUitxHCeCBb4SR/5lw8SPrkWSyig80L1YdAJ2Hsqu5EwT1WkHTF9wpvqNwYk+LVe8qmeX+xPUclBBvnaj8jPaOEMEjgwBBEJowgoMKT0t1OAhJq6DPnEJIWnjAgOUSJtRlqAMRuwlfc9p18rZkPbGM7VqTqcoehNSulghTUR5CWFzmmvjmXU27G/efetp7tajv597BcRqXBD7l26Y+V+dqUWjiy1bg6SaYsuY6njuktmumJu7n6rS5BATZ/AZxRPC3CqHfXatJrW1m94yG3+1mYY1e57nZngzt6QB176P8yc4XKvWNqrre+uV+nY+6iKWsIxVmucm6tjBLprkfYUHPOLJ2XdunFvn7iPVKeSaRXxZzv07tbKUyQ==</latexit>Ct <latexit sha1_base64="xvJmIIEAPIcDqbL6LREB13LnQvA=">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</latexit>Yt1[CtCls.Loss ImageStep tSeg.LossFigure 1. Illustration of the major difference of pipeline between previous WILSS work and FMWISS. Given a model pre-trained on old classes with pixel-level labels ( Yt−1), previous work [8] learn new classes ( e.g.,horse ) via image-level labels ( Ct), while FMWISS improves and effectively utilizes the supervision from complementary foundation models. dataset is prone to easily forget learned knowledge when be- ing retrained on another dataset with new classes. This phe- nomenon is known as catastrophic forgetting [37], which is caused by large changes of model parameters to model new samples with novel categories without accessing old sam- ples. A promising approach to solve such catastrophic for- getting problem is called incremental learning. Many methods have been proposed to solve image classification task [7, 10, 17, 25, 28, 33, 41, 44, 46, 49, 50]. Recently, a few methods have been presented to address incremental learning for semantic segmentation (ILSS) task, where only new classes of training samples of the current step are la- beled with pixel annotations and old classes of the previ- ous step are labeled as background. Modern ILSS methods can be classified into two categories: regularization-based and replay-based. Regularization-based methods [9, 18, 39] focus on distilling knowledge, e.g., output probability, in- termedia features, from pre-trained model of previous step. Replay-based methods [36] propose to store the information of previous old classes or web-crawled images and replay for new training steps. However, a key barrier to further de- velop these methods is the requirement for pixel-level an- notations for new classes. Very recently, WILSON [8] first proposes a new task, weakly incremental learning for se- mantic segmentation (WILSS), to incrementally update the model from image-level labels for new classes. Despite the comparable results, the image-level labels can not provide 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. 23685 details to accurately locate each segment, which limits the performance and development of WILSS. In this work, we explore to improve and more effec- tively utilize the supervision of new classes given image- level labels while preserving the knowledge of old ones. We propose a Foundation Model drives Weakly Incremental learning for Semantic Segmentation framework, dubbed FMWISS. Firstly, as shown in Figure 1, we are the first attempt to leverage pre-trained foundation models to improve the su- pervision given image-level labels for WILSS in a training- free manner. To be specific, we propose pre-training based co-segmentation to distill the knowledge of vision-language pre-training models ( e.g., CLIP [42]) and self-supervised pre-training models ( e.g., iBOT [52]), which can be com- plementary to each other. However, it is not trivial to apply the pre-trained models. We first adapt CLIP for category- aware dense mask generation. Based on the initial mask for each new class, we then propose to extract compact category-agnostic attention maps with seeds guidance us- ing self-supervised models. We finally refine the pseudo masks via mask fusion. We further propose to optimize the still noisy pseudo masks with a teacher-student archi- tecture, where the plug-in teacher is optimized with the pro- posed dense contrastive loss. Thus we can more effectively utilize the pseudo dense supervision. Finally, we present memory-based copy-paste augmentation to remedy the for- getting problem of old classes and can further improve the performance. The contributions of this paper are as follows: • We present a novel and data-efficient WILSS frame- work, called FMWISS, which is the first attempt to uti- lize complementary foundation models to improve and more effectively use the supervision given only image- level labels. • We propose pre-training based co-segmentation to generate dense masks by distilling both category- aware and category-agnostic knowledge from pre- trained foundation models, which provides dense su- pervision against original image labels. • To effectively utilize pseudo labels, we use a teacher- student architecture with a proposed dense contrastive loss to dynamically optimize the noisy pseudo labels. • We further introduce memory-based copy-paste aug- mentation to remedy the forgetting problem of old classes and can also improve performance. • Extensive experiments on Pascal VOC and COCO datasets demonstrate the significant efficacy of our FMWISS framework.
Yang_Object_Pose_Estimation_With_Statistical_Guarantees_Conformal_Keypoint_Detection_and_CVPR_2023	Abstract The two-stage object pose estimation paradigm ﬁrst de- tects semantic keypoints on the image and then estimates the 6D pose by minimizing reprojection errors. Despite per- forming well on standard benchmarks, existing techniques offer no provable guarantees on the quality and uncertainty of the estimation. In this paper, we inject two fundamental changes, namely conformal keypoint detection andgeomet- ric uncertainty propagation , into the two-stage paradigm and propose the ﬁrst pose estimator that endows an estima- tion with provable and computable worst-case error bounds . On one hand, conformal keypoint detection applies the sta- tistical machinery of inductive conformal prediction to con- vert heuristic keypoint detections into circular or elliptical prediction sets that cover the groundtruth keypoints with a user-speciﬁed marginal probability ( e.g.,90%). Geometric uncertainty propagation, on the other, propagates the ge- ometric constraints on the keypoints to the 6D object pose, leading to a Pose UnceRtainty SEt ( PURSE )that guarantees coverage of the groundtruth pose with the same probabil- ity. The PURSE , however, is a nonconvex set that does not directly lead to estimated poses and uncertainties. There- fore, we develop RANdom SAmple averaGing ( RANSAG ) to compute an average pose and apply semideﬁnite relax- ation to upper bound the worst-case errors between the average pose and the groundtruth. On the LineMOD Oc- clusion dataset we demonstrate: (i) the PURSE covers the groundtruth with valid probabilities; (ii) the worst-case er- ror bounds provide correct uncertainty quantiﬁcation; and (iii) the average pose achieves better or similar accuracy as representative methods based on sparse keypoints.	1. Introduction Estimating object poses from images is a fundamental problem in computer vision and ﬁnds extensive applications in augmented reality [ 42], autonomous driving [ 80], robotic manipulation [ 60], and space robotics [ 19]. One of the most popular paradigms for object pose estimation is a two-stage pipeline [ 20,71,72,79,81,85,89,101], where the ﬁrst stage detects (semantic) keypoints of the objects on the image, and the second stage computes the object pose by solving an optimization known as Perspective- n-Points (PnP) that minimizes reprojection errors of the detected keypoints. Safety-critical applications call for provably correct computer vision algorithms. Existing algorithms in the two- stage paradigm (reviewed in Section 2), however, provide few performance guarantees on the quality of the estimated poses, due to three challenges. (C1) It is difﬁcult to en- sure the detected keypoints (typically from neural networks) are close to the groundtruth keypoints. In practice, the ﬁrst stage often outputs keypoints that are arbitrarily wrong, known as outliers . (C2) Robust estimation is employed in the second stage to reject outliers, leading to nonconvex op- timizations. Fast heuristics such as RANSAC [26] are widely adopted to ﬁnd an approximate solution but they cannot guarantee global optimality and often fail without notice. (C3) There is no provably correct uncertainty quantiﬁcation of the estimation, notably, a formal worst-case error bound between the estimation and the groundtruth. Though recent work [ 98] proposed convex relaxations to certify global op- timality of RANSAC and addressed (C2), it cannot ensure correct estimation as the optimal pose may be far away from 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. 8947 the correct pose when the keypoints are unreliable. Contributions . We propose a two-stage object pose es- timation framework with statistical guarantees , illustrated in Fig. 1. Given an input image, we assume a neural net- work [ 71] is available to generate heatmap predictions of the object keypoints (Fig. 1(a)). Our framework then pro- ceeds in two stages, namely conformal keypoint detection (Section 4) and geometric uncertainty propagation (Sec- tion5). We ﬁrst apply the statistical machinery of induc- tive conformal prediction (introduced in Section 3), with nonconformity functions inspired by the design of resid- ual functions in classical geometric vision [ 39], to con- formalize the heatmaps into circular or elliptical predic- tion sets –one for each keypoint– that guarantee coverage of the groundtruth keypoints with a user-speciﬁed marginal probability (Fig. 1(b)). This provides a simple and general methodology to bound the keypoint prediction errors ( i.e., addressing (C1)). Given the keypoint prediction sets, we reformulate the constraints (enforced by the prediction sets) on the keypoints as constraints on the object pose, leading to aPose UnceRtainty SEt (PURSE ) that guarantees coverage of the groundtruth pose with the same probability. Fig. 1(c) plots the boundary of an example PURSE (roll, pitch, raw angles for the rotation, and Euclidean coordinates for the translation). The PURSE , however, is an abstract nonconvex set that does not directly admit estimated poses and uncer- tainty. Therefore, we develop RANdom SAmple averaGing (RANSAG ) to compute an average pose (Fig. 1(d)) and em- ploy semideﬁnite relaxations to upper bound the worst-case rotation and translation errors between the average pose and the groundtruth (Fig. 1(e)). This gives rise to the ﬁrst kind of computable worst-case probabilistic error bounds for object pose estimation ( i.e., addressing (C3)). Our PURSE method- ology has connections to the framework of unknown-but- bounded noise estimation in control theory [ 63], with spe- cial provisions to derive the bounds in a statistically princi- pled way and enable efﬁcient computation. We test our framework on the LineMOD Occlusion ( LM- O) dataset [ 11] to verify the correctness of the theory (Sec- tion 6). First, we empirically show that the PURSE in- deed contains the groundtruth pose according to the user- speciﬁed probability. Second, we demonstrate the correct- ness of the worst-case error bounds: when the PURSE con- tains the groundtruth, our bounds are always larger than, and in many cases close to, the actual errors between the av- erage pose and the groundtruth pose. Third, we benchmark the accuracy of the average pose (coming from RANSAG ) with representative two-stage pipelines based on sparse key- points ( e.g., PVNet [ 72]) and show that the average pose achieves better or similar accuracy. Limitations . A drawback of our approach, and confor- mal prediction in general, is that the size of the prediction sets depends on the nonconformity function (whose designcan be an art) and may be conservative. Our experiments suggest the bounds are loose when the keypoint prediction sets are large ( e.g., giving 180rotation bound). We discuss challenges and opportunities in tightening the bounds.
Yan_Long-Term_Visual_Localization_With_Mobile_Sensors_CVPR_2023	Abstract Despite the remarkable advances in image matching and pose estimation, image-based localization of a camera in a temporally-varying outdoor environment is still a chal- lenging problem due to huge appearance disparity between query and reference images caused by illumination, seasonal and structural changes. In this work, we propose to leverage additional sensors on a mobile phone, mainly GPS, compass, and gravity sensor, to solve this challenging problem. We show that these mobile sensors provide decent initial poses and effective constraints to reduce the searching space in image matching and final pose estimation. With the initial pose, we are also able to devise a direct 2D-3D matching network to efficiently establish 2D-3D correspondences in- stead of tedious 2D-2D matching in existing systems. As no public dataset exists for the studied problem, we collect a new dataset that provides a variety of mobile sensor data and significant scene appearance variations, and develop a system to acquire ground-truth poses for query images. We benchmark our method as well as several state-of-the-art baselines and demonstrate the effectiveness of the proposed approach. Our code and dataset are available on the project page: https://zju3dv.github.io/sensloc/ .	1. Introduction Visual localization aims at estimating the camera trans- lation and orientation for a given image relative to a known scene. Solving this problem is crucial for many applications such as autonomous driving [11], robot navigation [37] and augmented and virtual reality [7, 41]. State-of-the-art approaches to visual localization typi- cally involve matching 3D points in a pre-built map and 2D pixels in a query image [4, 22, 42, 50 –52, 67, 69]. An intermediate image retrieval step [2, 21, 23, 46] is often ap- plied to determine which parts of the scene are likely visible The authors from Zhejiang University are affiliated with the State Key Lab of CAD&CG and ZJU-SenseTime Joint Lab of 3D Vision. †Corresponding author: Xiaowei Zhou. AA Reference Query Illumination changesDark nightRainy weatherCross seasonsNew construction Figure 1. Visual localization under extremely challenging condi- tions. The proposed benchmark dataset SensLoc exhibits long-term appearance changes due to illumination, weather, season, day-night alternation, and new constructions. in the query image, in order to handle large-scale scenes. The resulting camera poses are estimated using a standard Perspective-n-Point (PnP) solver [24, 32] inside a robust RANSAC [3, 12, 13, 18] loop. However, in real-world out- door scenarios, obtaining such correspondences and further recovering the 6-DoF pose are difficult, since the outdoor scenes can experience large appearance variations caused by illumination (e.g., day and night), seasonal (e.g., summer and winter) and structure changes. Visual localization under such challenging conditions remains an unsolved problem, as reported by recent benchmarks [48, 53, 76]. Fortunately, nowadays, with the popularity of smart de- vices that come equipped with various sensors such as In- ertial Measurement Unit (IMU), gravity, compass, GPS, or radio signals (like WiFi and Bluetooth), new possibilities arise for mobile phone pose estimation exploiting these addi- tional multi-modality sensors. Nevertheless, previous works 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. 17245 only take independent sensors into consideration. For ex- ample, some methods utilize GPS as a prior to bound the Visual-Inertial Odometry (VIO) drift [31, 45, 56, 75] or sim- plify the image retrieval process [35, 72, 73], while others focus on employing the gravity direction as a reliable prior to benefit the PnP solver [1, 20, 33, 52, 63–65]. In this paper, we introduce a novel framework, named SensLoc, that localizes an image by tightly coupling visual perception with complementary mobile sensor information for robust localization under extreme conditions. In the first stage, our approach leverages GPS and compass to constrain the search space for image retrieval, which not only reduces the risk of misrepresentation of global features but also speedups the retrieval procedure with fewer database candidates. In the second stage, inspired by the recent CAD- free object pose estimation method, OnePose++ [26], we design a transformer-based network to directly match 3D points of the retrieved sub-map to dense 2D pixels of the query photo in a coarse-to-fine manner. Compared to the modern visual localization pipeline, which establishes 2D- 3D correspondences by repeatedly matching 2D-2D local features between query and retrieve images, our solution shows a significant acceleration and a better performance especially under challenging appearance changes. In the last stage, we implement a simple yet effective gravity validation algorithm and integrate it into the RANSAC loop to filter the wrong pose hypotheses, leveraging the precise roll and pitch angles from mobile gravity sensors. The gravity validation leads to an improvement of RANSAC in terms of efficiency and accuracy, as false hypotheses can be removed in advance. To the best of our knowledge, there is no public dataset for multi-sensor localization under strong visual changes. To facilitate the research of this area, we created a new bench- mark dataset SensLoc , as shown in Fig. 1. Specifically, we first used a consumer-grade panoramic camera (Insta360) and a handheld Real-time Kinematic (RTK) recorder, to cap- ture and reconstruct a large-scale reference map. Half a year later, we collected query sequences with large scene appear- ance changes through a mobile phone bounded with a RTK and recorded all available built-in sensor data. As direct registration between the query sequences and the reference map is difficult, we rebuilt an auxiliary map at the same time as acquiring the query sequences using Insta360 and RTK and aligned the auxiliary map with the reference map through ICP. Thus, we only needed to register the query im- ages with the auxiliary map, which was easier as they were captured at the same time. To achieve this, we developed a pseudo-ground-truth (GT) generation algorithm to accu- rately register each query sequence against the auxiliary map by incorporating feature matching, visual-inertial-odometry, RTK positions and gravity directions. The GT generation algorithm does not ask for any manual intervention or extra setup in the environment, enabling scalable pose labeling.We evaluate several state-of-the-art image retrieval and localization methods on our proposed dataset. We show that the performance of the existing methods can be drastically improved by considering sensor priors available in mobile devices, such as GPS, compass, and gravity direction. The experiments also demonstrate that our method outperforms the state-of-the-art approach HLoc [50,51] by a large margin in challenging night-time environments, while taking only 66msto find 2D-3D correspondences on a GPU and 8ms for PnP RANSAC. In summary, our main contributions include: •A novel outdoor visual localization framework with multi-sensor prior for robust and accurate localization under extreme visual changes. •A new dataset for multi-sensor visual localization with seasonal and illumination variations. •Benchmarking existing methods and demonstrating the effectiveness of the proposed approach.
Zhang_CLAMP_Prompt-Based_Contrastive_Learning_for_Connecting_Language_and_Animal_Pose_CVPR_2023	Abstract Animal pose estimation is challenging for existing image-based methods because of limited training data and large intra- and inter-species variances. Motivated by the progress of visual-language research, we propose that pre-trained language models ( e.g., CLIP) can fa- cilitate animal pose estimation by providing rich prior knowledge for describing animal keypoints in text. How- ever, we found that building effective connections between pre-trained language models and visual animal keypoints is non-trivial since the gap between text-based descrip- tions and keypoint-based visual features about animal pose can be significant. To address this issue, we introduce a novel prompt-based Contrastive learning scheme for connecting Language and AniMalPose (CLAMP) effec- tively. The CLAMP attempts to bridge the gap by adapt- ing the text prompts to the animal keypoints during net- work training. The adaptation is decomposed into spatial- aware and feature-aware processes, and two novel con- trastive losses are devised correspondingly. In practice, the CLAMP enables the first cross-modal animal pose es- timation paradigm. Experimental results show that our method achieves state-of-the-art performance under the su- pervised, few-shot, and zero-shot settings, outperforming image-based methods by a large margin. The code is avail- able at https://github.com/xuzhang1199/CLAMP .	1. Introduction Animal pose estimation aims to locate and identify a se- ries of animal body keypoints from an input image. It plays a key role in animal behavior understanding, zoology, and wildlife conservation which can help study and protect an- imals better. Although the animal pose estimation task is analogous to human pose estimation [2] to some extent, we argue that the two tasks are very different. For example, ani- mal pose estimation involves multiple animal species, while Spatial-aware Adaptation Text Prompt: ‘left front paw’ Effective Animal Pose EstimationCLAMP Feature-aware Adaptation Spatial-aware AdaptationCLAMP Feature-aware Adaptation Figure 1. Conceptualized visualization of our CLAMP method. Regarding the animal pose estimation task, we proposed to exploit rich language information from texts to facilitate the visual iden- tification of animal keypoints. To better connect texts and animal images, we devise the CLAMP to adapt pre-trained language mod- els via a spatial-aware and a feature-aware process. As a result, the CLAMP helps deliver better animal pose estimation performance. human pose estimation only focuses on one category. Be- sides, it is much more difficult to collect and annotate ani- mal pose data covering different animal species, thus exist- ing animal pose datasets are several times smaller than the human pose datasets [20] regarding the number of samples per species. Recently, Yu et al. [38] attempted to alleviate this problem by presenting the largest animal pose estima- tion dataset, i.e., AP-10K, which contains 10K images from 23 animal families and 54 species and provides the base- line performance of SimpleBaseline [32] and HRNet [31]. Despite this progress, the volume of this dataset is still far smaller than the popular human pose dataset, such as MS COCO [20] with 200K images. With diverse species and limited data, current animal pose datasets usually have large variances in animal poses which include both intra-species and inter-species vari- 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. 23272 ances. More specifically, the same animal can have diverse poses, e.g., pandas can have poses like standing, crawling, sitting, and lying down. Besides, the difference in the poses of different animal species can also be significant, e.g., horses usually lie down to the ground, while monkeys can be in various poses. Furthermore, even with the same pose, different animals would have different appearances. As an example, the joints of monkeys are wrinkled and hairy, while those of hippos are smooth and hairless. As a result, it could be extremely challenging for current human pose esti- mation methods to perform well on the animal pose estima- tion task without sufficient training data. Although image- based pre-training methodologies can be helpful in mitigat- ing the problem of insufficient data, the huge gap between the pre-training datasets ( e.g. ImageNet [7] image classifi- cation dataset and MS COCO human pose dataset [20]) and the animal pose datasets could compromise the benefits of pre-training procedures. Rather than only using images to pre-train models, we notice that the keypoints of different poses and different an- imals share the same description in natural languages, thus the language-based pre-trained models can be beneficial to compensate for the shortage of animal image data. For example, if a pre-trained language model provides a text prompt of “a photo of the nose”, we can already use it to identify the presence of the nose keypoint in the image with- out involving too much training on the new dataset. For- tunately, a recently proposed Contrastive Language-Image Pre-training (CLIP) [28] model can provide a powerful mapping function to pair the image with texts effectively. Nevertheless, we found that fine-tuning the CLIP on the an- imal pose dataset could still suffer from large gaps between the language and the images depicting animals. In partic- ular, the vanilla CLIP model only learns to provide a text prompt with general language to describe the entire image, while the animal pose estimation requires pose-specific de- scriptions to identify several different keypoints with their locations estimated from the same image. To this end, it is important to adapt the pre-trained language model to the an- imal pose dataset and effectively exploit the rich language knowledge for animal pose estimation. To address the above issue, we propose a novel prompt- based contrastive learning scheme for effectively connect- ing language and animal pose (called CLAMP), enabling the first cross-modal animal pose estimation paradigm. In particular, we design pose-specific text prompts to describe different animal keypoints, which will be further embed- ded using the language model with rich prior knowledge. By adapting the pose-specific text prompts to visual animal keypoints, we can effectively utilize the knowledge from the pre-trained language model for the challenging animal pose estimation. However, there is a significant gap between the pre-trained CLIP model (which generally depicts the entireimage) and the animal pose task (which requires the specific keypoint feature discriminative to and aligned with given text descriptions). To this end, we decompose the compli- cated adaptation into a spatial and feature-aware process. Specifically, we devise a spatial-level contrastive loss to help establish spatial connections between text prompts and the image features. A feature-level contrastive loss is also devised to make the visual features and embedded prompts of different keypoints more discriminative to each other and align their semantics in a compatible multi-modal embed- ding space. With the help of the decomposed adaptation, ef- fective connections between the pre-trained language model and visual animal poses are established. Such connections with rich prior language knowledge can help deliver better animal pose prediction. In summary, the contribution of this paper is threefold: • We propose a novel cross-modal animal pose estima- tion paradigm named CLAMP to effectively exploit prior language knowledge from the pre-trained lan- guage model for better animal pose estimation. • We propose to decompose the cross-modal adaptation into a spatial-aware process and a feature-aware pro- cess with carefully designed losses, which could effec- tively align the language and visual features. • Experiments on two challenging datasets in three set- tings, i.e., 1) AP-10K [38] dataset (supervised learn- ing, few-shot learning, and zero-shot learning) and 2) Animal-Pose [4] dataset (supervised learning), vali- date the effectiveness of the CLAMP method.
Yao_Local_Implicit_Normalizing_Flow_for_Arbitrary-Scale_Image_Super-Resolution_CVPR_2023	Abstract Flow-based methods have demonstrated promising re- sults in addressing the ill-posed nature of super-resolution (SR) by learning the distribution of high-resolution (HR) images with the normalizing ﬂow. However, these methods can only perform a predeﬁned ﬁxed-scale SR, limiting their potential in real-world applications. Meanwhile, arbitrary- scale SR has gained more attention and achieved great progress. Nonetheless, previous arbitrary-scale SR meth- ods ignore the ill-posed problem and train the model with per-pixel L1 loss, leading to blurry SR outputs. In this work, we propose “Local Implicit Normalizing Flow” (LINF) as a uniﬁed solution to the above problems. LINF models the distribution of texture details under different scaling fac- tors with normalizing ﬂow. Thus, LINF can generate photo- realistic HR images with rich texture details in arbitrary scale factors. We evaluate LINF with extensive experiments and show that LINF achieves the state-of-the-art perceptual quality compared with prior arbitrary-scale SR methods.	1. Introduction Arbitrary-scale image super-resolution (SR) has gained increasing attention recently due to its tremendous appli- cation potential. However, this ﬁeld of study suffers from two major challenges. First, SR aims to reconstruct high- resolution (HR) image from a low-resolution (LR) counter- part by recovering the missing high-frequency information. This process is inherently ill-posed since the same LR im- age can yield many plausible HR solutions. Second, prior deep learning based SR approaches typically apply upsam- pling with a pre-deﬁned scale in their network architectures, such as squeeze layer [ 1], transposed convolution [ 2], and sub-pixel convolution [ 3]. Once the upsampling scale is de- termined, they are unable to further adjust the output res- olutions without modifying their model architecture. This causes inﬂexibility in real-world applications. As a result, * and †indicate equal contribution. This work was developed during the internship of Jie-En Yao and Li-Yuan Tsao at MediaTek Inc. Figure 1. A comparison of the previous arbitrary-scale SR ap- proaches and LINF. LINF models the distribution of texture details in HR images at arbitrary scales. Therefore, unlike the prior meth- ods that tend to produce blurry images, LINF is able to generate arbitrary-scale HR images with rich and photo-realistic textures. discovering a way to perform arbitrary-scale SR and pro- duce photo-realistic HR images from an LR image with a single model has become a crucial research direction. A natural approach to addressing the one-to-many in- verse problem in SR is to consider the solution as a dis- tribution. Consequently, a number of generative-based SR methods [ 1,4–8] have been proposed to tackle this ill- posed problem. Among them, ﬂow-based SR methods show promise, as normalizing ﬂow [ 9–12] offers several advantages over other generative models. For instance, ﬂow does not suffer from the training instability and mode collapse issues present in generative adversarial networks (GANs) [ 13]. Moreover, ﬂow-based methods are compu- tationally efﬁcient compared to diffusion [ 14] and autore- gressive (AR) [ 15,16] models. Representative ﬂow-based models, such as SRFlow [ 1] and HCFlow [ 7], are able to generate high-quality SR images and achieve state-of-the- art results on the benchmarks. However, these methods are restricted to ﬁxed-scale SR, limiting their applicability. Another line of research focuses on arbitrary-scale SR. LIIF [ 17] employs local implicit neural representation to represent images in a continuous domain. It achieves arbitrary-scale SR by replacing ﬁxed-scale upsample 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. 1776 ules with an MLP to query the pixel value at any coordi- nate. LTE [ 18] further estimates the Fourier information at a given coordinate to make MLP focus on learning high- frequency details. However, these works did not explicitly account for the ill-posed nature of SR. They adopt a per- pixel L1loss to train the model in a regression fashion. The reconstruction error favors the averaged output of all possi- ble HR images, leading the model to generate blurry results. Based on the observation above, combining ﬂow-based SR model with the local implicit module is a promising di- rection in which ﬂow can account for the ill-posed nature of SR, and the local implicit module can serve as a solu- tion to the arbitrary-scale challenge. Recently, LAR-SR [ 8] claimed that details in natural images are locally correlated without long-range dependency. Inspired by this insight, we formulated SR as a problem of learning the distribution of local texture patch. With the learned distribution, we per- form super-resolution by generating the local texture sepa- rately for each non-overlapping patch in the HR image. With the new problem formulation, we present Local Im- plicit Normalizing Flow (LINF) as the solution. Speciﬁ- cally, a coordinate conditional normalizing ﬂow model sthe local texture patch distribution, which is conditioned on the LR image, the central coordinate of local patch, and the scaling factor. To provide the conditional signal for the ﬂow model, we use the local implicit module to estimate Fourier information at each local patch. LINF excels the previous ﬂow-based SR methods with the capability to up- scale images with arbitrary scale factors. Different from prior arbitrary-scale SR methods, LINF explicitly addresses the ill-posed issue by learning the distribution of local tex- ture patch. As shown in Fig 1, hence, LINF can generate HR images with rich and reasonable details instead of the over-smoothed ones. Furthermore, LINF can address the is- sue of unpleasant generative artifacts, a common drawback of generative models, by controlling the sampling tempera- ture. Speciﬁcally, the sampling temperature in normalizing ﬂow controls the trade-off between PSNR (ﬁdelity-oriented metric) and LPIPS [ 19] (perceptual-oriented metric). The contributions of this work can be summarized as follows: •We proposed a novel LINF framework that leverages the advantages of a local implicit module and normal- izing ﬂow. To the best of our knowledge, LINF is the ﬁrst framework that employs normalizing ﬂow to gen- erate photo-realistic HR images at arbitrary scales. •We validate the effectiveness of LINF to serve as a uni- ﬁed solution for the ill-posed and arbitrary-scale chal- lenges in SR via quantitative and qualitative evidences. •We examine the trade-offs between the ﬁdelity- and perceptual-oriented metrics, and show that LINF does yield a better trade-off than the prior SR approaches.
Zhang_Lite-Mono_A_Lightweight_CNN_and_Transformer_Architecture_for_Self-Supervised_Monocular_CVPR_2023	Abstract Self-supervised monocular depth estimation that does not require ground truth for training has attracted attention in recent years. It is of high interest to design lightweight but effective models so that they can be deployed on edge devices. Many existing architectures beneﬁt from using heavier backbones at the expense of model sizes. This paper achieves comparable results with a lightweight architecture. Speciﬁcally, the efﬁcient combination of CNNs and Trans- formers is investigated, and a hybrid architecture called Lite-Mono is presented. A Consecutive Dilated Convolu- tions (CDC) module and a Local-Global Features Interac- tion (LGFI) module are proposed. The former is used to extract rich multi-scale local features, and the latter takes advantage of the self-attention mechanism to encode long- range global information into the features. Experiments demonstrate that Lite-Mono outperforms Monodepth2 by a large margin in accuracy, with about 80% fewer train- able parameters. Our codes and models are available at https://github.com/noahzn/Lite-Mono .	1. Introduction Many applications in the ﬁeld of robotics, autonomous driving, and augmented reality rely on depth maps, which represent the 3D geometry of a scene. Since depth sen- sors increase costs, research on inferring depth maps us- ing Convolutional Neural Networks (CNNs) from images emerged. With the annotated depth one can train a regres- sion CNN to predict the depth value of each pixel on a sin- gle image [ 10,11,22]. Lacking large-scale accurate dense ground-truth depth for supervised learning, self-supervised methods that seek supervisory signals from stereo-pairs of frames or monocular videos are favorable and have made great progress in recent years. These methods regard the depth estimation task as a novel view synthesis problem and minimize an image reconstruction loss [ 5,14,15,41,45]. *Corresponding Author Input Monodepth2 R-MSFM6 Lite-Mono (Ours) Figure 1. The proposed Lite-Mono has fewer parameters than Monodepth2 [ 15] and R-MSFM [ 46], but generates more accu- rate depth maps. The camera motion is known when using stereo-pairs of im- ages, so a single depth estimation network is adopted to pre- dict depth. But if only using monocular videos for training an additional pose network is needed to estimate the motion of the camera. Despite this, self-supervised methods that only require monocular videos are preferred, as collecting stereo data needs complicated conﬁgurations and data pro- cessing. Therefore, this paper also focuses on monocular video training. In addition to increasing the accuracy of monocular training by introducing improved loss functions [ 15] and semantic information [ 5,21] to mitigate the occlusion and moving objects problems, many works focused on design- ing more effective CNN architectures [ 17,33,39,41,46]. However, the convolution operation in CNNs has a local receptive ﬁeld, which cannot capture long-range global in- formation. To achieve better results a CNN-based model can use a deeper backbone or a more complicated archi- tecture [ 15,28,44], which also results in a larger model size. The recently introduced Vision Transformer (ViT) [ 8] is able to model global contexts, and some recent works ap- ply it to monocular depth estimation architectures [ 3,35] to obtain better results. However, the expensive calculation of the Multi-Head Self-Attention (MHSA) module in a Trans- former hinders the design of lightweight and fast inference models, compared with CNN models [ 35]. This paper pursues a lightweight and efﬁcient self- supervised monocular depth estimation model with a hy- brid CNN and Transformer architecture. In each stage of the proposed encoder a Consecutive Dilated Convolutions (CDC) module is adopted to capture enhanced multi-scale local features. Then, a Local-Global Features Interaction (LGFI) module is used to calculate the MHSA and encode 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. 18537 global contexts into the features. To reduce the computa- tional complexity the cross-covariance attention [ 1] is cal- culated in the channel dimension instead of the spatial di- mension. The contributions of this paper can be summa- rized in three aspects. •A new lightweight architecture, dubbed Lite-Mono, for self-supervised monocular depth estimation, is pro- posed. Its effectiveness with regard to the model size and FLOPs is demonstrated. •The proposed architecture shows superior accuracy on the KITTI [ 13] dataset compared with competitive larger models. It achieves state-of-the-art with the least trainable parameters. The model’s generalization abil- ity is further validated on the Make3D [ 32] dataset. Additional ablation experiments are conducted to ver- ify the effectiveness of different design choices. •The inference time of the proposed method is tested on an NVIDIA TITAN Xp and a Jetson Xavier platform, which demonstrates its good trade-off between model complexity and inference speed. The remainder of the paper is organized as follows. Sec- tion2reviews some related research work. Section 3illus- trates the proposed method in detail. Section 4elaborates on the experimental results and discussion. Section 5con- cludes the paper.
Yan_NeRF-DS_Neural_Radiance_Fields_for_Dynamic_Specular_Objects_CVPR_2023	Abstract Dynamic Neural Radiance Field (NeRF) is a powerful algorithm capable of rendering photo-realistic novel view images from a monocular RGB video of a dynamic scene. Although it warps moving points across frames from the observation spaces to a common canonical space for ren- dering, dynamic NeRF does not model the change of the reflected color during the warping. As a result, this ap- proach often fails drastically on challenging specular ob- jects in motion. We address this limitation by reformulat- ing the neural radiance field function to be conditioned on surface position and orientation in the observation space. This allows the specular surface at different poses to keep the different reflected colors when mapped to the common canonical space. Additionally, we add the mask of moving objects to guide the deformation field. As the specular sur- face changes color during motion, the mask mitigates the problem of failure to find temporal correspondences with only RGB supervision. We evaluate our model based on the novel view synthesis quality with a self-collected dataset of different moving specular objects in realistic environments. The experimental results demonstrate that our method sig- nificantly improves the reconstruction quality of moving specular objects from monocular RGB videos compared to the existing NeRF models. Our code and data are available at the project website1.	1. Introduction Neural Radiance Fields (NeRF) [25] trained with multi- view images can synthesize novel views for 3D scenes with photo-realistic quality. NeRF predicts the volume density and view dependent color of the sampled spatial points in the scene with a multi-layer perceptron (MLP). Recent works such as Nerfies [32] and NSFF [22] extend NeRF to reconstruct dynamic scenes from monocular videos. They resolve the lack of multi-view image supervision in dy- namic scenes using a deformation field, which warps dif- 1https://github.com/JokerYan/NeRF-DS HyperNeRF NeRF-DS (ours) Figure 1. Comparison of novel views rendered by HyperN- eRF [33] (left) and our NeRF-DS (right), on the “americano” scene in the HyperNeRF dataset [33]2(top) and the “basin” scene in our dynamic specular dataset (bottom). Our NeRF-DS model signifi- cantly improves the reconstruction quality by a surface-aware dy- namic NeRF and a mask guided deformation field. ferent observation spaces to a common canonical space. Despite showing promising results, we find that the ex- isting dynamic NeRFs do not consider specular reflections during warping and often fail drastically on challenging dy- namic specular objects as shown in Fig. 1. The quality of dynamic specular object reconstruction is important be- cause specular (e.g. metallic, plastic) surfaces are common in our daily environment and furthermore it indicates how accurate a dynamic NeRF represents the radiance field un- der motion or deformation. Previous works such as Ref- NeRF [50] and NeRV [45] have only focused on improving the specular reconstruction in static scenes. The problem of reconstructing dynamic specular objects with NeRF remain largely unexplored. We postulate that one of the reasons for dynamic models to fail on moving specular objects is because they do not 2The rendered frames come from the first 3 seconds of the “americano” scene when the cup is rotating. This part of the video is not included in the HyperNeRF [33] qualitative results. 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. 8285 consider the original surface information when rendering in a common canonical space. As suggested in rendering mod- els such as Phong shading [34], the specular color depends on the relative position and orientation of the surface with respect to the reflected environment. Nonetheless, existing dynamic NeRFs often ignore the original position and ori- entation of the surface when warping a specular object to a common canonical space for rendering. As the result, a point on a specular object reflecting different colors at dif- ferent positions and orientations can cause conflicts when warped to a common canonical space. Additionally, the key of existing dynamic models is to learn a deformation field for each frame such that correspondences can be established in a shared canonical space. However, the color of specular objects can vary significantly at different locations and ori- entations, which makes it hard to establish correspondences with the RGB supervision alone. These two limitations in- evitably lead to the failure of existing dynamic models when applied to specular objects. In this paper, we introduce NeRF-DS (Fig. 2) which models dynamic specular objects using a surface-aware dy- namic NeRF and a mask guided deformation field to miti- gate the two limitations mentioned above. 1) Our NeRF-DS still warps the points from the observation space to a com- mon canonical space and predicts their volume density. In contrast to other dynamic NeRFs, the color of each point is additionally conditioned on the spatial coordinate and sur- face normal in the observation space before warping. Cor- responding points from different frames can share the same geometry, but reflect different colors determined by their original surface position and orientation. 2) Our NeRF-DS reuses the moving object mask from the camera registra- tion stage as an additional input to the deformation field. This mask is a more consistent guidance for specular sur- faces in motion compared to the constantly changing color. The mask is also a strong cue to the deformation field on the moving and static regions. As shown in Fig. 1, our pro- posed NeRF-DS reconstructs and renders dynamic specular scenes with significantly higher quality. We implement our NeRF-DS on top of the state-of-the-art HyperNeRF [33] for dyanmic scenes. Since there are very limited dynamic specular objects in the existing datasets, we collect another dynamic specular dataset for evaluation. Our dataset consists of a variety of moving/deforming specular objects in realistic environ- ments. Experimental results on the dataset demonstrate that the NeRF-DS significantly improves the quality of novel view rendering on dynamic specular objects. The images rendered by our NeRF-DS avoid many serious artifacts compared to the existing NeRF models. In summary, we have made the following contributions: 1. A reparameterized dynamic NeRF that models dy- namic specular surface with additional observationspace coordinate and surface normal. 2. A mask guided deformation field that improves defor- mation learned for dynamic specular objects. 3. A dynamic specular scene dataset with training and testing monocular videos.
Yu_MVImgNet_A_Large-Scale_Dataset_of_Multi-View_Images_CVPR_2023	Abstract Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet [24] drives a remarkable trend of ‘learning from large-scale data’ in computer vision. Pretraining on ImageNet to ob- tain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serv- ing as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet , a large- scale dataset of multi-view images, which is highly conve- nient to gain by shooting videos of real-world objects in hu- man daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238classes, with rich annota- tions of object masks, camera parameters, and point clouds.The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision. We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, includ- ing radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations. Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet , cov- ering 87,200 samples from 150categories, with the class label on each point cloud. Experiments show that MVP- Net can benefit the real-world 3D object classification while posing new challenges to point cloud understanding. MVImgNet and MVPNet will be public, hoping to inspire the broader vision community. 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. 9150	1. Introduction Being data-driven, also known as data-hungry, is one of the most important attributes of deep learning algorithms. By training on large-scale datasets, deep neural networks are able to extract rich representations. In the past few years, the computer vision community has witnessed the bloom of such ‘learning from data’ regime [43, 44, 55], af- ter the birth of ImageNet [24] – the pioneer of large-scale real-world image datasets. Notably, pretraining on Ima- geNet is well-proven to boost the model performance when transferring the pretrained weights into not only high-level [34, 41, 42, 60, 61] but also low-level visual tasks [15, 57], and becomes a de-facto standard in 2D. Recently, various 3D datasets [5, 9, 11, 23, 33, 96, 104] are produced to facili- tate 3D visual applications. However, due to the non-trivial scanning and labori- ous labeling of real-world 3D data (commonly organized in point clouds or meshes), existing 3D datasets are ei- ther synthetic or their scales are not comparable with Ima- geNet [24]. Consequently, unlike in 2D vision where mod- els are usually pretrained on ImageNet to gain universal rep- resentation or commonsense knowledge, most of the current methods in 3D area are directly trained and evaluated on particular datasets for solving specific 3D visual tasks ( e.g., NeRF dataset [64] and ShapeNet [11] for novel view syn- thesis, ModelNet [96] and ScanObjectNN [88] for object classification, KITTI [33] and ScanNet [23] for scene un- derstanding). Here, two crucial and successive issues can be induced: (1)There is still no generic dataset in 3D vision, as a counterpart of ImageNet in 2D. (2)What benefit such a dataset can endow to 3D community is yet unknown . In this paper, we focus on investigating these two problems and set two corresponding targets: Build the primary dataset, then explore its effect through experiments. Milestone 1 – Dataset: For a clearer picture of the first goal, we start by carefully revisiting existing 3D datasets as well as ImageNet [24]. i)3D synthetic datasets [11,96] provide rich 3D CAD mod- els. However, they lack real-world cues ( e.g., context, oc- clusions, noises), which are indispensable for model robust- ness in practical applications. ScanObjectNN [88] extracts real-world 3D objects from indoor scene data, but is limited in scale. For 3D scene-level dataset [5, 10, 23, 33, 37, 81], their scales are still constrained by the laborious scanning and labeling ( e.g., millions of points per scene). Addition- ally, they contain specific inner-domain knowledge such as a particularly intricate indoor room or outdoor driving con- figurations, making it hard for general transfer learning. ii)Although ImageNet [24] contains the most comprehen- sive real-world objects, it only describes a 2D world that misses 3D-aware signals. Since humans live in a 3D world, 3D consciousness is vitally important for realizing human- like intelligence and solving real-life visual problems.Based on the above review, our dataset is created from a new insight – multi-view images , as a soft bridge be- tween 2D and 3D. It lies several benefits to remedying the aforementioned defects. Such data can be easily gained in considerable sizes via shooting an object around different views on common mobile devices with cameras ( e.g., smart- phones), which can be collected by crowd-sourcing in real world . Moreover, the multi-view constraint can bring nat- ural 3D visual signals (later experiments show that this not only benefits 3D tasks but also 2D image understanding). To this end, we build MVImgNet , containing 6.5million frames from 219,188 videos crossing real-life objects from 238classes, with rich annotations of object masks, camera parameters, and point clouds. You may take a glance at our MVImgNet from Fig. 1. Milestone 2 – Experimental Exploration: Now facing the second goal of this paper, we attempt to probe the power of our dataset by conducting some pilot experiments. Leveraging the multi-view nature of the data, we start by focusing on the view-based 3D reconstruction task and demonstrate that pretraining on MVImgNet can not only benefit the generalization ability of NeRF (Sec. 4.1), but also data-efficient multi-view stereo (Sec. 4.2). More- over, for image understanding, although humans can easily recognize one object from different viewpoints, deep learn- ing models can hardly do that robustly [26,38]. Considering MVImgNet provides numerous images of a particular ob- ject from different viewpoints, we verify that MVImgNet- pretrained models are endowed with decent view consis- tency in general image classification (supervised learning in Sec. 5.1, self-supervised contrastive learning in Sec. 5.2) andsalient object detection (Sec. 5.3). Bonus – A New 3D Point Cloud Dataset – MVPNet: Through dense reconstruction on MVImgNet, a new 3D object point cloud dataset is derived, called MVPNet , which contains 87,200 point clouds with 150 categories, with the class label on each point cloud (see Fig. 7). Exper- iments show that MVPNet not only benefits the real-world 3D object classification task but also poses new challenges and prospects to point cloud understanding (Sec. 6). MVImgNet and MVPNet will be public, hoping to in- spire the broader vision community.
Yue_Connecting_the_Dots_Floorplan_Reconstruction_Using_Two-Level_Queries_CVPR_2023	Abstract We address 2D floorplan reconstruction from 3D scans. Existing approaches typically employ heuristically de- signed multi-stage pipelines. Instead, we formulate floor- plan reconstruction as a single-stage structured predic- tion task: find a variable-size set of polygons, which in turn are variable-length sequences of ordered vertices. To solve it we develop a novel Transformer architec- ture that generates polygons of multiple rooms in paral- lel, in a holistic manner without hand-crafted intermedi- ate stages. The model features two-level queries for poly- gons and corners, and includes polygon matching to make the network end-to-end trainable. Our method achieves a new state-of-the-art for two challenging datasets, Struc- tured3D and SceneCAD, along with significantly faster in- ference than previous methods. Moreover, it can read- ily be extended to predict additional information, i.e., se- mantic room types and architectural elements like doors and windows. Our code and models are available at: https://github.com/ywyue/RoomFormer.	1. Introduction The goal of floorplan reconstruction is to turn observa- tions of an (indoor) scene into a 2D vector map in birds- eye view. More specifically, we aim to abstract a 3D point cloud into a set of closed polygons corresponding to rooms, optionally enriched with further structural and semantic el- ements like doors, windows and room type labels. Floorplans are an essential representation that enables a wide range of applications in robotics, AR/VR, inte- rior design, etc.Like prior work [2, 3, 8, 9, 29], we start from a 3D point cloud, which can easily be captured with RGB-D cameras, laser scanners or SfM systems. Several works [8, 9, 21, 29] have shown the effectiveness of project- ing the raw 3D point data along the gravity axis, to obtain a 2D density map that highlights the building’s structural el- ements ( e.g., walls). We also employ this early transition to 2D image space. The resulting density maps are compact and computationally efficient, but inherit the noise and data gaps of the underlying point clouds, hence floorplan recon-Input 3D Point Cloud Reconstructed Floorplan Figure 1. Semantic floorplan reconstruction. Given a point cloud of an indoor environment, RoomFormer jointly recovers multiple room polygons along with their associated room types, as well as architectural elements such as doors and windows. struction remains a challenging task. Existing methods can be split broadly into two categories that both operate in two stages: Top-down methods [8, 29] first extract room masks from the density map using neu- ral networks ( e.g., Mask R-CNN [15]), then employ opti- mization/search techniques ( e.g., integer programming [28], Monte-Carlo Tree-Search [4]) to extract a polygonal floor- plan. Such techniques are not end-to-end trainable, and their success depends on how well the hand-crafted opti- mization captures domain knowledge about room shape and layout. Alternatively, bottom-up methods [9, 21] first de- tect corners, then look for edges between corners ( i.e., wall segments) and finally assemble them into a planar floorplan graph. Both approaches are strictly sequential and therefore dependent on the quality of the initial corner, respectively room, detector. The second stage starts from the detected entities, therefore missing or spurious detections may sig- nificantly impact the reconstruction. We address those limitations and design a model that directly maps a density image to a set of room polygons. Our model, named RoomFormer , leverages the sequence prediction capabilities of Transformers and directly out- puts a variable-length, ordered sequence of vertices per room. RoomFormer requires neither hand-crafted, domain- specific intermediate products nor explicit corner, wall or room detections. Moreover, it predicts all rooms that make up the floorplan at once, exploiting the parallel nature of the Transformer architecture. 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. 845 In more detail, we employ a standard CNN backbone to extract features from the birds-eye view density map, fol- lowed by a Transformer encoder-decoder setup that con- sumes image features (supplemented with positional encod- ings) and outputs multiple ordered corner sequences, in par- allel. The floorplan is recovered by simply connecting those corners in the predicted order. Note that the described pro- cess relies on the ability to generate hierarchically struc- tured output of variable and a-priori unknown size, where each floorplan has a different number of rooms (with no nat- ural order), and each room polygon has a different number of (ordered) corners. We address this challenge by introduc- ing two-level queries with one level for the room polygons and one level for their corners. The varying numbers of both rooms and corners are accommodated by additionally classifying each query as valid or invalid. The decoder it- eratively refines the queries, through self-attention among queries and cross-attention between queries and image fea- tures. To enable end-to-end training, we propose a poly- gon matching strategy that establishes the correspondence between predictions and targets, at both room and corner levels. In this manner, we obtain an integrated model that holistically predicts a set of polygons to best explain the ev- idence in the density map, without hand-tuned intermediate rules of which corners, walls or rooms to commit to along the way. The model is also fast at inference, since it operates in single-stage feed-forward mode, without optimization or search and without any post-processing steps. Moreover, it is flexible and can, with few straight-forward modifications, predict additional semantic and structural information such as room types, doors and windows (Fig. 1). We evaluate our model on two challenging datasets, Structured3D [37] and SceneCAD [2]. For both of them, RoomFormer outperforms the state of the art, while at the same time being significantly faster than existing methods. In summary, our contributions are: • A new formulation of floorplan reconstruction, as the simultaneous generation of multiple ordered se- quences of room corners. • The RoomFormer model, an end-to-end trainable, Transformer-type architecture that implements the proposed formulation via two-level queries that pre- dict a set of polygons each consisting of a sequence of vertex coordinates. • Improved floorplan reconstruction scores on both Structured3D [37] and SceneCAD [2], with faster in- ference times. • Model variants able to additionally predict semantic room type labels, doors and windows.
Yu_Data-Free_Knowledge_Distillation_via_Feature_Exchange_and_Activation_Region_Constraint_CVPR_2023	Abstract Despite the tremendous progress on data-free knowledge distillation (DFKD) based on synthetic data generation, there are still limitations in diverse and efficient data syn- thesis. It is naive to expect that a simple combination of generative network-based data synthesis and data augmen- tation will solve these issues. Therefore, this paper proposes a novel data-free knowledge distillation method (Spaceship- Net) based on channel-wise feature exchange (CFE) and multi-scale spatial activation region consistency (mSARC) constraint. Specifically, CFE allows our generative net- work to better sample from the feature space and efficiently synthesize diverse images for learning the student network. However, using CFE alone can severely amplify the un- wanted noises in the synthesized images, which may result in failure to improve distillation learning and even have negative effects. Therefore, we propose mSARC to assure the student network can imitate not only the logit output but also the spatial activation region of the teacher network in order to alleviate the influence of unwanted noises in di- verse synthetic images on distillation learning. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, Im- agenette, and ImageNet100 show that our method can work well with different backbone networks, and outperform the state-of-the-art DFKD methods. Code will be available at: https://github.com/skgyu/SpaceshipNet. *Equal contribution This research was supported in part by the National Key R&D Pro- gram of China (grant 2018AAA0102500), and the National Natural Sci- ence Foundation of China (grant 62176249).	1. Introduction Knowledge distillation (KD) aims to train a lightweight student model that can imitate the capability of a pre-trained complicated teacher model. In the past decade, KD has been studied in a wide range of fields such as image recogni- tion, speech recognition, and natural language processing. Traditional KD methods usually assume that the whole or part of the training set used by the teacher network is ac- cessible by the student network [17, 24, 34]. But in practi- cal applications, there can be various kinds of constraints in accessing the original training data, e.g., due to pri- vacy issues in medical data [1, 2, 5, 20, 28, 35] and portrait data [3], and copyright and privateness of large data vol- ume such as JFT-300M [40] and text-image data [37]. The traditional KD methods no longer work under these sce- narios. Recently, data-free knowledge distillation (DFKD) [4,7,10,13,14,22,25,27,29,43,46] seeks to perform KD by generating synthetic data instead of accessing the original training data used by the teacher network to train the student network. Thus, the general framework of DFKD consists of two parts: synthetic data generation that replicates the orig- inal data distribution and constraint design between student and teacher network during distillation learning. Synthetic data generation methods in DFKD mainly consist of noise image optimization-based methods [4, 26, 31, 44] and gen- erative network-based methods [8,9, 12,13, 27,29, 45]. The former approaches optimize randomly initialized noise im- ages to make them have the similar distribution to the orig- inal training data. These methods can theoretically gener- ate an infinite number of independent and identically dis- tributed images for student network learning, but they are usually extremely time-consuming, and thus are difficult in generating sufficient synthetic data with high diversity. The later approaches learn a generator to synthesize im- ages that approximate the distribution of the original train- ing data. These methods can be much faster than the image 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. 24266 optimization-based approach, but the diversity of the syn- thesized data is usually limited because the generation of different images are not completely independent with each other. Despite the encouraging results achieved, DFKD re- mains a challenging task, because the synthetic data may have a different distribution from the original data, which could potentially result in bias in student network learn- ing. The possible reason is that the noises in the synthe- sized images can easily lead to the bias of the network’s region of interest. In addition, the widely used KL diver- gence constraint between student and teacher networks in existing DFKD methods may not work well with synthetic data [4]. This paper proposes a novel DFKD method that utilizes channel-wise feature exchange (CFE) and multi-scale spa- tial activation region consistency (mSARC) constraint to improve knowledge transfer from the teacher network to the student network. The proposed method enhances the diversity of synthetic training data and the robustness to un- wanted noises in synthetic images during distillation. Un- like previous generative network-based methods that em- ployed multiple generators to synthesize images [27] or reinitialization and retraining of generator [13] to enhance the synthetic training data diversity, our method improves the synthetic training data diversity by using the features of early synthetic images to perform CFE. When our gen- erative network and those of other methods have learned to generate the same number of synthetic images, the pro- posed method can produce more diverse training data for distillation. However, CFE also amplifies unwanted noise in synthetic images, which may hinder distillation learning (traditional data augmentation methods also suffer from this problem, e.g., CutMix [47] and Mixup [50]). To address this issue, we propose the mSARC constraint, which en- ables the student network to learn discriminative cues from similar regions to those used by the teacher network, effec- tively overcoming the limitations of the traditional KL di- vergence loss when applied to synthetic images during dis- tillation learning. Moreover, combining our mSARC with traditional data augmentation methods [47,50] can still sig- nificantly improve distillation learning with synthetic data. We evaluate our method on a number of datasets, includ- ing CIFAR-10 [19], CIFAR-100 [19], and Tiny-ImageNet [21]. Our approach demonstrates superior performance compared to the state-of-the-art DFKD methods. More- over, we observe that the student networks trained using our DFKD method achieve comparable performance to those trained using original training data. Additionally, we eval- uate our method on subsets of ImageNet [11] with 10 and 100 classes and a resolution of 224×224, validating the efficacy of our method on generating high-resolution syn- thetic images for distillation learning. In our ablation study,we verify the effectiveness of the key components (CFE and mSARC), we find that mSARC plays a particularly impor- tant role when strong data augmentation is applied to the synthetic images.
Xu_V2V4Real_A_Real-World_Large-Scale_Dataset_for_Vehicle-to-Vehicle_Cooperative_Perception_CVPR_2023	Abstract Modern perception systems of autonomous vehicles are known to be sensitive to occlusions and lack the capabil- ity of long perceiving range. It has been one of the key bottlenecks that prevents Level 5 autonomy. Recent re- search has demonstrated that the Vehicle-to-Vehicle (V2V) cooperative perception system has great potential to rev- olutionize the autonomous driving industry. However, the lack of a real-world dataset hinders the progress of this field. To facilitate the development of cooperative per- ception, we present V2V4Real, the first large-scale real- world multi-modal dataset for V2V perception. The data is collected by two vehicles equipped with multi-modal sensors driving together through diverse scenarios. Our V2V4Real dataset covers a driving area of 410 km, com- prising 20K LiDAR frames, 40K RGB frames, 240K anno- tated 3D bounding boxes for 5 classes, and HDMaps that cover all the driving routes. V2V4Real introduces three perception tasks, including cooperative 3D object detec- tion, cooperative 3D object tracking, and Sim2Real domain adaptation for cooperative perception. We provide compre- hensive benchmarks of recent cooperative perception algo- rithms on three tasks. The V2V4Real dataset can be found at research.seas.ucla.edu/mobility-lab/v2v4real/.	1. Introduction Perception is critical in autonomous driving (A V) for ac- curate navigation and safe planning. The recent develop- ment of deep learning brings significant breakthroughs in various perception tasks such as 3D object detection [22, 35, 42], object tracking [43, 56], and semantic segmenta- tion [47, 57]. However, single-vehicle vision systems still suffer from many real-world challenges, such as occlusions and short-range perceiving capability [15, 40,49], which *Corresponding Author, email address: [email protected] (a) Aggregated LiDAR data (b) HD map (c) Satallite Map Figure 1. A data frame sampled from V2V4Real: (a) aggregated LiDAR data, (b) HD map, and (c) satellite map to indicate the collective position. More qualitative examples of V2V4Real can be found in the supplementary materials. can cause catastrophic accidents. The shortcomings stem mainly from the limited field-of-view of the individual ve- hicle, leading to an incomplete understanding of the sur- rounding traffic. A growing interest and recent advancement in coop- erative perception systems have enabled a new paradigm that can potentially overcome the limitation of single- vehicle perception. By leveraging vehicle-to-vehicle (V2V) technologies, multiple connected and automated vehicles (CA Vs) can communicate and share captured sensor infor- mation simultaneously. As shown in a complex intersection in Fig. 1, for example, the ego vehicle (red liDAR) strug- gles to perceive the upcoming objects located across the way due to occlusions. Incorporating the LiDAR features from the nearby CA V (green scans) can largely broaden the sensing range of the vehicle and make it even see across the occluded corner. Despite the great promise, however, it remains chal- 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. 13712 Dataset YearReal/ SimV2XSize (km)RGB imagesLiDAR Maps3D boxesClasses Locations Kitti [14] 2012 Real No - 15k 15k No 200k 8 Karlsruhe nuScenes [2] 2019 Real No 33 1.4M 400k Yes 1.4M 23 Boston, SG Argo [3] 2019 Real No 290 107k 22k Yes 993k 15 2x USA Waymo Open [36] 2019 Real No - 1M 200k Yes 12M 4 3x USA OPV2V [50] 2022 Sim V2V - 44k 11k Yes 230k 1 CARLA V2X-Sim [20] 2022 Sim V2V&I - 60K 10k Yes 26.6k 1 CARLA V2XSet [49] 2022 Sim V2V&I - 44K 11k Yes 230k 1 CARLA DAIR-V2X [54] 2022 Real V2I 20 39K 39K No 464K 10 Beijing, CN V2V4Real (ours) 2022 Real V2V 410 40K 20K Yes 240K 5 Ohio, USA Table 1. Comparison of the proposed dataset and existing representative autonomous driving datasets. lenging to validate V2V perception in real-world scenarios due to the lack of public benchmarks. Most of the exist- ing V2V datasets, including OPV2V [50], V2X-Sim [20], and V2XSet [49], rely on open-source simulators like CARLA [11] to generate synthetic road scenes and traffic dynamics with simulated
Yang_Neural_Volumetric_Memory_for_Visual_Locomotion_Control_CVPR_2023	Abstract Legged robots have the potential to expand the reach of autonomy beyond paved roads. In this work, we consider the difﬁcult problem of locomotion on challenging terrains using a single forward-facing depth camera. Due to the partial observability of the problem, the robot has to rely on past observations to infer the terrain currently beneath it. To solve this problem, we follow the paradigm in com- puter vision that explicitly models the 3D geometry of the scene and propose Neural Volumetric Memory (NVM), a ge- ometric memory architecture that explicitly accounts for the SE(3) equivariance of the 3D world. NVM aggregates fea- ture volumes from multiple camera views by ﬁrst bringing them back to the ego-centric frame of the robot. We test the learned visual-locomotion policy on a physical robot and show that our approach, which explicitly introduces geo- metric priors during training, offers superior performance than more na ¨ıve methods. We also include ablation studies and show that the representations stored in the neural vol- umetric memory capture sufﬁcient geometric information to reconstruct the scene. Our project page with videos is https://rchalyang.github.io/NVM	1. Introduction Consider difﬁcult locomotion tasks such as walking up and down a ﬂight of stairs and stepping over gaps (Figure 1).The control of such behaviors requires tight coupling with perception because vision is needed to provide details of the terrain right beneath the robot and the 3D scene imme- diately around it. This problem is also partially-observable. Immediately relevant terrain information is often occluded from the robot’s current frame of observation, forcing it to rely on past observations for control decisions. For this rea- son, while blind controllers that are learned in simulation using reinforcement learning have achieved impressive re- sults in agility and robustness [ 33,36,38], there are clear limitations on how much they can do. How to incorporate perception into the pipeline to produce an integrated visuo- motor controller thus remains an open problem. A recent line of work combines perception with loco- motion using ego-centric cameras mounted on the robot. The predominant approach for addressing partial observ- ability is to do frame-stacking, where the robot maintains a visual buffer of recent images. This na ¨ıve heuristic suf- fers from two major problems: ﬁrst, frame-stacking on a moving robot ignores the equivariance structure of the 3D environment, making learning a lot more difﬁcult as pol- icy success now relies on being able to learn to account for spurious changes in camera poses. A second but subtler is- sue is that biological systems do not have the ability to save detailed visual observations pixel-by-pixel. These concerns motivate the creation of an intermediary, short-term mem- ory mechanism to functionally aggregate streams of obser- 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. 1430 Stages Stairs Obstacles Stones Sim2RealStages Stairs Obstacles Front-Facing Depth Camera Stones Figure 2. Overview of Simulated Environment & Real World Environment: Our simulated environments are shown on the left and real-world environments are shown on the right. For the real-world environment, the corresponding visual observations for each real-world environment are shown in the bottom row. All policies are trained in the simulation and transferred into the real world without ﬁne-tuning. vation into a single, coherent representation of the world. Motivated by these observations, we introduce a novel volumetric memory architecture for legged locomotion con- trol. Our architecture consists of a 2D to 3D feature volume encoder, and a pose estimator that can estimate the relative camera poses given two input images. When combined, the two networks constitute a Neural V olumetric Memory (NVM) that takes as input a sequence of depth images taken by a forward-looking, ego-centric camera and fuses them together into a coherent latent representation for locomo- tion control. We encourage the memory space to be SE(3) equivariant to changes in the camera pose of the robot by in- corporating translation and rotation operations based on es- timated relative poses from the pose network. This inverse transformation allows NVM to align feature volumes from the past to the present ego-centric frame, making both inte- grating over multiple timesteps into a coherent scene repre- sentation and learning a policy, less difﬁcult. Our training pipeline follows a two-step teacher-student process where the primary goal of the ﬁrst stage is to pro- duce behaviors in the form of a policy. After training com- pletes, this policy can traverse these difﬁcult terrains ro- bustly, but it relies on privileged sensory information such as an elevation map and ground-truth velocity. Elevation maps obtained in the real world are often biased, incom- plete, and full of errors [ 40], whereas ground-truth velocity information is typically only available in instrumented en- vironments. Hence in the visuomotor distillation stage of the pipeline, which still runs in the simulator, we feed the stream of ego-centric views from the forward depth cam- era into the neural volumetric memory. We feed the content of this memory into a small policy network and train ev- erything end-to-end including the two network components of the NVM using a behavior cloning loss where the state- only policy acts as the teacher. For completeness, we offer an additional self-supervised learning objective (Figure 4) that relies on novel-view consistency for learning. The end product of this visuomotor distillation pipeline is a memory- equipped visuomotor policy that can operate directly on theUniTree A1 robot hardware (see Figure 1). A single policy is used to handle all environments covered by this paper. We provide comprehensive experiment and ablation studies in both simulation and the real world, and show that our method outperforms all baselines by a large margin. It is thus essential to model the 3D structure of the environment.
Zhang_Generalization_Matters_Loss_Minima_Flattening_via_Parameter_Hybridization_for_Efficient_CVPR_2023	Abstract Most existing online knowledge distillation (OKD) tech- niques typically require sophisticated modules to produce diverse knowledge for improving students’ generalization ability. In this paper, we strive to fully utilize multi-model settings instead of well-designed modules to achieve a dis- tillation effect with excellent generalization performance. Generally, model generalization can be reflected in the flat- ness of the loss landscape. Since averaging parameters of multiple models can find flatter minima, we are inspired to extend the process to the sampled convex combinations of multi-student models in OKD. Specifically, by linearly weighting students’ parameters in each training batch, we construct a Hybrid-Weight Model (HWM) to represent the parameters surrounding involved students. The supervi- sion loss of HWM can estimate the landscape’s curva- ture of the whole region around students to measure the generalization explicitly. Hence we integrate HWM’s loss into students’ training and propose a novel OKD frame- work via parameter hybridization (OKDPH) to promote flatter minima and obtain robust solutions. Considering the redundancy of parameters could lead to the collapse of HWM, we further introduce a fusion operation to keep the high similarity of students. Compared to the state-of- the-art (SOTA) OKD methods and SOTA methods of seek- ing flat minima, our OKDPH achieves higher performance with fewer parameters, benefiting OKD with lightweight and robust characteristics. Our code is publicly available at https://github.com/tianlizhang/OKDPH.	1. Introduction Deep learning achieves breakthrough progress in a va- riety of tasks by constructing a large capacity network †Corresponding authorpre-trained on massive data [6]. In order to apply high- parameterized models in the real-world scene with lim- ited resources, the knowledge distillation (KD) technique [12] aims to obtain a compact and effective student model guided by a large-scale teacher model for model compres- sion. Based on the developments of KD, Zhang et al. [44] propose the concept of online knowledge distillation (OKD) to view all networks as students and achieve mutual learn- ing from scratch through peer teaching, liberating the dis- tillation process from the dependency on pre-trained teach- ers. Existing OKD methods mainly encourage students to acquire diverse and rich knowledge, including aggregating predictions [10, 33, 37], combining features [19, 22, 28], working with peers [39, 47], learning from group lead- ers [2], and receiving guidance from online teachers [39]. Nevertheless, these strategies focus on designing sophis- ticated architectures to exploit heterogeneous knowledge to enhance students’ generalization, but they lack explicit con- straints on generalization. The concept of generalization to deep models is the ability to fit correctly on previously un- seen data [25], which can be reflected by the flatness of the loss landscape [14, 18]. Flatness is the landscape’s local curvature, which is costly to direct calculate by the Hessian. Considering the setting of multiple students in OKD, we uti- lize the theory of multi-model fusion in parameter space [9] to estimate the local curvature by the linear combination of students’ parameters (we call it a hybrid-weight model , which is expressed as HWM). More specifically, HWM is a stochastic convex combination of parameters of multiple students on different data augmentations, which can sample multiple local points on the landscape. Intuitively, HWM’s loss reflects the upper and lower bounds of the local region’s loss and estimates the curvature of the landscape. Minimiz- ing HWM’s loss flattens the region and forms a landscape with smaller curvature. Based on the above observation, we propose a concise and effective OKD framework, termed online 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. 20176 distillation with parameter hybridization (OKDPH), to pro- mote flatter loss minima and achieve higher generaliza- tion. We devise a novel loss function for students’ train- ing that incorporates the standard Cross-Entropy (CE) loss and Kullback-Leibler (KL) divergence loss, but also a su- pervised learning loss from HWM. Specifically, HWM is constructed in each batch by linearly weighting multiple students’ parameters. The classification error of HWM ex- plicitly measures the flatness of the region around students on the loss landscape, reflecting their stability and general- ization. The proposed loss equals imposing stronger con- straints on the landscape, guiding students to converge in a more stable direction. For intuitive understanding, we vi- sualize the loss landscape of students obtained by different methods in Fig. 1. Our students converge to one broader and flatter basin (thus superior generalization performance), while the students obtained by DML [44] converge to differ- ent sharp basins, degrading the robustness and performance. Unfortunately, directly hybridizing students’ parameters can easily lead to the collapse of HWM due to the high nonlinearity of deep neural networks [26, 31]. Therefore, we restrict the differences between students through inter- mittent fusion operations to ensure the high similarity of multi-model parameters and achieve effective construction of HWM. Concretely, at regular intervals, we hybridize the parameter of HWM with each student and, conversely, assign the hybrid parameter to the corresponding student. This process shortens the distance between students, shown as very close loss trajectories of our students in Fig. 1. How- ever, it will not reduce diversity because students receive different types of data augmentation, and they can easily become diverse during training. Our method pulls students in the same direction, plays the role of strong regularization, and obtains one lightweight parameter that performs well in various scenarios. The solution derived from our method is expected to integrate the dark knowledge from multiple models while maintaining a compact architecture and can be competent for resource-constrained applications. To sum up, our contributions are organized as follows: • Inspired by the theory of multi-model fusion, we inno- vatively extend traditional weight averaging to an on- the-fly stochastic convex combination of students’ pa- rameters, called a hybrid-weight model (HWM). The supervision loss of HWM can estimate the curvature of the loss landscape around students and explicitly mea- sure the generalization. • We propose a brand-new extensible and pow- erful OKD framework via parameter hybridiza- tion (OKDPH) for loss minima flattening, which flex- ibly adapts to various network architectures without modifying peers’ structures and extra modules. It is the first OKD work that manipulates parameters. −30 −20 −10 0 10 20 30 P arameter1−20−100102030P arameter2Ours-S1 Ours-S2 DML-S1 DML-S2 0.00.40.81.21.62.02.42.83.23.6 LossFigure 1. The loss landscape visualization of four students (Ours- S1 and Ours-S2 are obtained by our method, and DML obtains DML-S1 and DML-S2), which are ResNet32 [11] trained by the same settings on CIFAR-10 [20]. Four students start from the ini- tial point (Red points in the center) and converge to three basins along different trajectories. The x-axis and y-axis represent the values of model parameters that PCA [23] obtains. • Extensive experiments on various backbones demon- strate that our OKDPH can considerably improve the students’ generalization and exceed the state-of-the- art (SOTA) OKD methods and SOTA approaches of seeking flat minima. Further loss landscape visualiza- tion and stability analysis verify that our solution lo- cates in the region having uniformly low loss and is more robust to perturbations and limited data.
Yang_Visual_Recognition-Driven_Image_Restoration_for_Multiple_Degradation_With_Intrinsic_Semantics_CVPR_2023	Abstract Deep image recognition models suffer a signiﬁcant per- formance drop when applied to low-quality images sincethey are trained on high-quality images. Although manystudies have investigated to solve the issue through imagerestoration or domain adaptation, the former focuses on vi- sual quality rather than recognition quality, while the lat- ter requires semantic annotations for task-speciﬁc training.In this paper , to address more practical scenarios, we pro-pose a Visual Recognition-Driven Image Restoration net-work for multiple degradation, dubbed VRD-IR, to recoverhigh-quality images from various unknown corruption types from the perspective of visual recognition within one model.Concretely, we harmonize the semantic representations of diverse degraded images into a uniﬁed space in a dynamicmanner , and then optimize them towards intrinsic semanticsrecovery. Moreover , a prior-ascribing optimization strat-egy is introduced to encourage VRD-IR to couple with var-ious downstream recognition tasks better . Our VRD-IR iscorruption- and recognition-agnostic, and can be insertedinto various recognition tasks directly as an image enhance-ment module. Extensive experiments on multiple image dis-tortions demonstrate that our VRD-IR surpasses existing image restoration methods and show superior performanceon diverse high-level tasks, including classiﬁcation, detec-tion, and person re-identiﬁcation.	1. Introduction We have witnessed the remarkable success made by deep learning in image recognition tasks in recent years, such asclassiﬁcation [ 25,33,66], detection [ 23,46,62,68], and seg- mentation [ 9,51]. However, most of these approaches lever- age the public datasets with high-quality images ( e.g., Ima- *Equal contribution †Corresponding Author515253545556575 12 13 14 15 16 17 18 19 20CUB Top-1 Accuracy on VGG16 (%) PSNR (dB)VRD-IR AODNet EPRNHazy Image DehazeNetFDGANAirNetMPRNetClean Image DDPURIE Figure 1. Illustration of visual quality and recognition quality us- ing different dehazing methods on hazy CUB [ 73]. The top-1 ac- curacy is evaluated by VGG16 [ 66] pre-trained on clean CUB. Our method, VRD-IR, is shown in bold. As we can see, higher visualquality doesn’t mean higher recognition quality. geNet [ 64], CoCo [ 47]) for training, and they suffer a signif- icant drop when applied to low-quality images ( e.g., hazy, rainy, and noisy), since the statistical properties of pixelsare ruined by image degradation [ 75]. An intuitive approach to tackle this issue is to restore the distorted images ﬁrst, and then feed them into the succeed-ing recognition models. With this line, various image en-hancement methods have been developed to improve the hu- man visual quality of corrupted images [ 15,90]. However, the visual quality and the recognition quality of an imagediffer fundamentally from one another. As shown in Fig. 1, the restored image with higher visual effect cannot guaran-tee satisfactory performance on downstream high-level vi-sion tasks [ 57,67,81]. Another feasible solution is to encourage the recognition models to learn corruption-invariant feature representations,which can be applied to low-quality images directly withoutimage recovery. For that purpose, numerous datasets have 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. 14059 been created [ 26,50,80]. One common method is to narrow the distribution distance between low- and high- quality im-ages in feature space [ 32,37,67,81]. While promising, most of these methods neglect the fact that the adverse impacts ofdifferent degradation are quite different on semantic level.On the other hand, they either assume that the task-speciﬁcannotations is available during training, or just could han-dle a single corruption/recognition task only, which hindersthe timely adaptation to changing external environment andadjustment to ﬂexible high-level tasks in real-world. In this paper, we propose a visual recognition-driven im- age restoration (VRD-IR) for multiple degradation, to re-cover the recognition-friendly high-quality image from itsgiven degraded version without knowing the speciﬁc degra-dation type and downstream high-level tasks. We ﬁrst har-monize the semantic features suffered from different degra-dation into a uniﬁed representation spaces, and then opti-mize them towards semantic recovery. Speciﬁcally, we de-sign a model paradigm: Intrinsic Semantics Enhancement(ISE), which can restore different degraded semantic repre-sentations in a dynamic manner. It consists of a DegradationNormalization and Compensation (DNC) module for map-ping different degraded features to a degradation-invariantspace, and a Fourier Guided Modulation (FGM) for guid-ing the feature enhancement from the statistical propertiesin amplitude spectrum. For better perception of different semantics, a prior-ascribing optimization strategy is pro- posed. A semantic aware decoder (SAD) is ﬁrst pre-trainedon both low- and high- quality images with the objective toreconstruct the high-quality image from the correspondingsemantic features. To make full use of semantic informationand provide good guidance for ISE, a similarity ranking loss is enforced during the pre-training of SAD. Then, we ﬁx thepre-trained SAD and force the ISE to improve the qualityof images reconstructed by SAD through enhancing the de- graded semantic representations. In this way, we encouragethe ISE to modulate the degraded input features from the perspective of machine vision. Moreover, the proposed VRD-IR can be plugged into pre-trained recognition models directly as a data enhance-ment module. Compared with feature distillation-based methods that require task-speciﬁc annotations for training,our VRD-IR enjoys better ﬂexibility and practicality. We summarize our main contributions as follows: • To the best of our knowledge, VRD-IR is the ﬁrst attempt towards a pure universal image restorationframework for high-level vision. As the VRD-IR canbe integrated with various recognition models directly,it is more practical in real world scenario. • Considering the adverse impacts of different degrada- tion in semantics, we design an Intrinsic Semantic En-hancement (ISE) module to modulate the degraded se-mantic representation in a dynamic manner. • A prior-ascribing optimization strategy is proposed to endow VRD-IR with capability to perceive degrada-tion effects on semantic level. Guided by this, our ISEcan modulate degraded features from the perspectiveof machine vision. • We verify the effectiveness of our framework on di- verse high-level vision tasks, including classiﬁcation,detection, and person re-identiﬁcation. Experimentsresults show the superiority of our method in recogni-tion tasks under multiple degradation.
Zhang_Learning_Neural_Proto-Face_Field_for_Disentangled_3D_Face_Modeling_in_CVPR_2023	Abstract Generative models show good potential for recovering 3D faces beyond limited shape assumptions. While plau-sible details and resolutions are achieved, these modelseasily fail under extreme conditions of pose, shadow orappearance, due to the entangled ﬁtting or lack of multi- view priors. To address this problem, this paper presents a novel Neural Proto-face Field (NPF) for unsupervisedrobust 3D face modeling. Instead of using constrainedimages as Neural Radiance Field (NeRF), NPF disentan- gles the common/speciﬁc facial cues, i.e., ID, expression and scene-speciﬁc details from in-the-wild photo collec- tions. Speciﬁcally, NPF learns a face prototype to aggregate3D-consistent identity via uncertainty modeling, extract- ing multi-image priors from a photo collection. NPF thenlearns to deform the prototype with the appropriate facialexpressions, constrained by a loss of expression consistency and personal idiosyncrasies. Finally, NPF is optimized to ﬁt a target image in the collection, recovering speciﬁc detailsof appearance and geometry. In this way, the generative model beneﬁts from multi-image priors and meaningful fa- cial structures. Extensive experiments on benchmarks showthat NPF recovers superior or competitive facial shapes and textures, compared to state-of-the-art methods.	1. Introduction 3D face reconstruction is a long-standing problem with applications including games, digital human and mobile photography. It is ill-posed in many cases requiring strong assumptions e.g., shape from shading [ 99]. With the 3D Morphable Model (3DMM) [ 10] proposed, such a problem can be solved by ﬁtting parameters to the target faces [ 67, 68,107]. Recently, deep-learning methods [ 22,25,43,64, Chengjie Wang and Ying Tai are corresponding authors5RWDWLRQ(' 3 7 , +HDG 1H5) 2XUV 'HIRUPDWLRQ E *HRPHWU\ 2XUV /$3 '')5 D Figure 1. (a) Comparison with graphics-renderer-based methods LAP [ 100] and D3DFR [ 20]. Our method models geometry de- tails and photo-realistic texture. (b) Results of neural rendering methods EG3D [ 13] + PTI [ 66], HeadNeRF [ 34] and our method. Our method produces high-quality geometry, robust texture mod-eling under rotation and deformation. 105] are proposed to regress 3DMM parameters from in- put images. These approaches are then improved by non-linear modeling [ 24,29,79,81,84,94] and multi-view con- sistency [ 7,15,76,90]. Besides 3DMM methods, recent ef- forts [ 91,100] attempt to model 3D face without shape as- sumptions. These non-parametric methods have potential ability to improve the modeling quality beyond 3DMM. Although the aforementioned methods achieve impres- sive performance, they also have obvious drawbacks. On the one hand, as the parametric models are usually builtfrom a small amount of subjects (e.g., BFM [ 58] with 200 subjects) and rigidly controlled conditions, they may befragile to large variations of identity [ 106], and have limi- 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. 382 tations on building teeth, skin details or anatomic grounded muscles [ 23]. On the other hand, all of these methods de- pend on graphics renderers [ 42,44,46] in the analysis-by- synthesis ﬁtting procedure, and thus yields hand-crafted ap-proximation or ill-posed decomposition on intrinsic clues. Hence, as illustrated in Fig. 1-(a), these methods struggle to produce photo-realistic texture or geometric details. Against these limitations, efforts are made to use a neu- ral renderer such as StyleGAN [ 38,39] to model faces by inverting the corresponding images [ 1,2] intoWspace. Ex- isting methods [ 11,18,59,62,92] mainly learn to embed 3DMM coefﬁcients to implicitly leverage 3D clues, but theyhave difﬁculty achieving precise 3D controls due to their en-tangled image formation. To disentangle neural rendering,recent works [ 13,14,34,54] employ explicit 3D pipelines, e.g., Neural Radiance Field (NeRF) [ 52] into the Style- GANs’ framework, so that face shapes and camera views can be extracted. In this way, precise 3D controls and de-tailed geometry can be obtained. However, these methods still show fragile performance under challenging conditions as shown in Fig. 1-(b). When confronting large poses, ex- treme appearance or lighting, the lack of facial priors dis- turbs the reconstruction and results in severe distortions. This is due to the essentially overﬁtting objective of invert-ing single target image, where the geometry ambiguity isunavoidable. On top of this, one solution is to leverage reliable pri- ors, e.g., multi-image consistency as a complement. While NeRF provides a natural paradigm to dig multi-view cues, it requires fully constrained images that are difﬁcult to obtain. Even conditioned by style codes [ 13,14,54], there is no di- rect way to build 3D faces from unconstrained portrait col- lections in such a neural rendering mechanism. In this work,we present a novel Neural Proto-face Field (NPF) for un- supervised robust 3D face modeling, where ID, expressionand scene-speciﬁc details can be disentangled from in-the- wild photo collections. To aggregate ID-aware cues, NPF leverages uncertainty modeling to extract multi-image pri-ors and recovers a face prototype with ID-consistent faceshape. To disentangle the expression, NPF then learns appropriate representations to deform the prototype, con- strained by a expression consistency loss. In this way, thelearned face shape is properly blended to avoid geometric ambiguity. Finally, to recover the scene-speciﬁc details,NPF is optimized to ﬁt a target image in the collection. The robustness of ﬁtting is guaranteed by a geometry and ap-pearance regularization. As shown in Fig. 1-(b), NPF makes the generative method beneﬁt from multi-image priors inunconstrained environments, and produces high-quality 3Dfaces under challenging conditions. In summary, our contributions are as follows: 1)A novel Neural Proto-face Field (NPF) is proposed to disentangle ID, expression and speciﬁc details from 3D faceMethods Rendering Pipeline Multi-view EMOCA [81], DECA [ 24], Unsup3D [ 91] Graphics Disentangled × LAP [100], FML [ 76], MVF [ 90] Graphics Disentangled /check DFG [18], StyleRig [ 77], StyleFlow [ 3] Neural Entangled × Pi-GAN [14], StyleSDF [ 54], EG3D [ 13] Neural Disentangled × Ours Neural Disentangled /check Table 1. Discussion with selected existing methods. modeling, which uses in-the-wild photo collections to ben- eﬁt the 3D generative model under challenging conditions. 2)With a novel face prototype aggregation method, NPF integrates multi-image face priors against the large varia-tions in unconstrained environments. 3)With a series of novel consistency losses, NPF is well ﬁt to speciﬁc scenes with personalized details, based on theguidance of face prototypes.
Yi_Weakly-Supervised_Single-View_Image_Relighting_CVPR_2023	Abstract We present a learning-based approach to relight a sin- gle image of Lambertian and low-frequency specular ob- jects. Our method enables inserting objects from pho- tographs into new scenes and relighting them under the new environment lighting, which is essential for AR appli- cations. To relight the object, we solve both inverse ren- dering and re-rendering. To resolve the ill-posed inverse rendering, we propose a weakly-supervised method by a low-rank constraint. To facilitate the weakly-supervised training, we contribute Relit, a large-scale (750K images) dataset of videos with aligned objects under changing il- luminations. For re-rendering, we propose a differen- tiable specular rendering layer to render low-frequency non-Lambertian materials under various illuminations of spherical harmonics. The whole pipeline is end-to-end and efficient, allowing for a mobile app implementation of AR object insertion. Extensive evaluations demonstrate that our method achieves state-of-the-art performance. Project page: https://renjiaoyi.github.io/relighting/.	1. Introduction Object insertion finds extensive applications in Mobile AR. Existing AR object insertions require a perfect mesh of the object being inserted. Mesh models are typically built by professionals and are not easily accessible to am- ateur users. Therefore, in most existing AR apps such as SnapChat and Ikea Place, users can use only built-in vir- tual objects for scene augmentation. This may greatly limit user experience. A more appealing setting is to allow the user to extract objects from a photograph and insert them into the target scene with proper lighting effects. This calls for a method of inverse rendering and relighting based on a single image, which has so far been a key challenge in the graphics and vision fields. Relighting real objects requires recovering lighting, ge- ometry and materials which are intertwined in the observed image; it involves solving two problems, inverse render- *Co-first authors. †Corresponding author: [email protected]. Input pho to Non- Lambertian object relighting Figure 1. Our method relights real objects into new scenes from single images, which also enables editing materials from diffuse to glossy with non-Lambertian rendering layers. ing [17] and re-rendering. Furthermore, to achieve real- istic results, the method needs to be applicable for non- Lambertian objects. In this paper, we propose a pipeline to solve both problems, weakly-supervised inverse render- ing and non-Lambertian differentiable rendering for Lam- bertian and low-frequency specular objects. Inverse rendering is a highly ill-posed problem, with sev- eral unknowns to be estimated from a single image. Deep learning methods excel at learning strong priors for reduc- ing ill-posedness. However, this comes at the cost of a large amount of labeled training data, which is especially cum- bersome to prepare for inverse rendering since ground truths of large-scale real data are impossible to obtain. Synthetic training data brings the problem of domain transfer. Some methods explore self-supervised pipelines and acquire ge- ometry supervisions of real data from 3D reconstruction by multi-view stereo (MVS) [34, 35]. Such approaches, how- ever, have difficulties in handling textureless objects. To tackle the challenge of training data shortage, we pro- pose a weakly-supervised inverse rendering pipeline based on a novel low-rank loss and a re-rendering loss. For low- rank loss, a base observation here is that the material re- flectance is invariant to illumination change, as an intrin- sic property of an object. We derive a low-rank loss for inverse rendering optimization which imposes that the re- flectance maps of the same object under changing illumi- nations are linearly correlated . In particular, we constrain 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. 8402 ... Encoder Decoder Skip connections Encoder Lighting coefficients+ MLP Skip connections ...Shape normal / Relit DatasetDifferentiable Render LayerTraining batch Rendered shading Reflectance Singular reflectanceNormal -Net Light -NetLowrank Constraint Module Self-supervision Normal - Net Light - Net+Render Layer Novel lightings/material Render Layer Inverse rendering decompositions + Training Relighting Spec - Net Spec -NetDiffuse branc h Specular parameters Figure 2. Overview of our method. At training time, Spec-Net separates input images into specular and diffuse branches. Spec-Net, Normal-Net and Light-Net are trained in a self-supervised manner by the Relit dataset. At inference time, inverse rendering properties are predicted to relight the object under novel lighting and material. The non-Lambertian render layers produce realistic relit images. the reflectance matrix with each row storing one of the re- flectance maps to be rank one. This is achieved by minimiz- ing a low-rank loss defined as the Frobenius norm between the reflectance matrix and its rank-one approximation. We prove the convergence of this low-rank loss. In contrast, traditional Euclidean losses lack a convergence guarantee. To facilitate the learning, we contribute Relit, a large- scale dataset of videos of real-world objects with changing illuminations. We design an easy-to-deploy capturing sys- tem: a camera faces toward an object, both placed on top of a turntable. Rotating the turntable will produce a video with the foreground object staying still and the illumination changing. To extract the foreground object from the video, manual segmentation of the first frame suffices since the ob- ject is aligned across all frames. As shown in Figure 2, a fixed number of images under different lighting are randomly selected as a batch. We first devise a Spec-Net to factorize the specular highlight, trained by the low-rank loss on the chromaticity maps of diffuse im- ages (image subtracts highlight) which should be consistent within the batch. With the factorized highlight, we further predict the shininess and specular reflectance, which is self- supervised with the re-rendering loss of specular highlight. For the diffuse branch, we design two networks, Normal- Net and Light-Net, to decompose the diffuse component by predicting normal maps and spherical harmonic lighting coefficients, respectively. The diffuse shading is rendered by normal and lighting, and diffuse reflectance (albedo) is computed by diffuse image and shading. Both networks aretrained by low-rank loss on diffuse reflectance. Regarding the re-rendering phase, the main difficulty is the missing 3D information of the object given a single- view image. The Normal-Net produces a normal map which is a partial 3D representation, making the neural rendering techniques and commercial renderers inapplicable. The ex- isting diffuse rendering layer for normal maps of [20] can- not produce specular highlights. Pytorch3D and [9,11] ren- der specular highlights for point lights only. To this end, we design a differentiable specular renderer from normal maps, based on the Blinn-Phong specular re- flection [5] and spherical harmonic lighting [6]. Combining with the differentiable diffuse renderer, we can render low- frequency non-Lambertian objects with prescribed parame- ters under various illuminations, and do material editing as a byproduct. We have developed an Android app based on our method which allows amateur users to insert and relight arbitrary objects extracted from photographs in a target scene. Exten- sive evaluations on inverse rendering and image relighting demonstrate the state-of-the-art performance of our method. Our contributions include: • A weakly-supervised inverse rendering pipeline trained with a low-rank loss. The correctness and con- vergence of the loss are mathematically proven. • A large-scale dataset of foreground-aligned videos col- lecting 750Kimages of 100+ real objects under differ- ent lighting conditions. 8403 • An Android app implementation for amateur users to make a home run.
Zhang_Painting_3D_Nature_in_2D_View_Synthesis_of_Natural_Scenes_CVPR_2023	Abstract We introduce a novel approach that takes a single seman- tic mask as input to synthesize multi-view consistent color images of natural scenes, trained with a collection of single images from the Internet. Prior works on 3D-aware image synthesis either require multi-view supervision or learning category-level prior for specific classes of objects, which are inapplicable to natural scenes. Our key idea to solve this challenge is to use a semantic field as the intermedi- ate representation, which is easier to reconstruct from an input semantic mask and then translated to a radiance field with the assistance of off-the-shelf semantic image synthe- sis models. Experiments show that our method outperforms baseline methods and produces photorealistic and multi- view consistent videos of a variety of natural scenes. The ∗Affiliated with the State Key Lab of CAD&CG, Zhejiang University. †Corresponding author: Xiaowei Zhou.project website is https://zju3dv.github.io/paintingnature/.	1. Introduction Natural scenes are indispensable content in many appli- cations such as film production and video games. This work focuses on a specific setting of synthesizing novel views of natural scenes given a single semantic mask, which en- ables us to generate 3D contents by editing 2D semantic masks. With the development of deep generative models, 2D semantic image synthesis methods [24, 46, 61, 66] have achieved impressive advances. However, they do not con- sider the underlying 3D structure and cannot generate multi- view consistent free-viewpoint videos. To address this problem, a straightforward approach is first utilizing semantics-driven image generator like SPADE [46] to synthesize an image from the input semantic mask and then predicting novel views based on the gener- ated image. Although the existing single-view view syn- 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. 8518 thesis methods [31, 34, 45, 52, 67, 70] achieve impressive rendering results, they typically require training networks on posed multi-view images. Compared to urban or indoor scenes, learning to synthesize natural scenes is a challeng- ing task as it is difficult to collect 3D data or posed videos of natural scenes for training, as demonstrated in [32], making the aforementioned methods not applicable. AdaMPI [17] designs a training strategy to learn the view synthesis net- work on single-view image collections. It warps images to random novel views and warps them back to the original view. An inpainting network is trained to fill the holes in disocclusion regions to match the original images. After training, the inpainting network is used to generate pseudo multi-view images for training a view synthesis network. Our experimental results in Section 4.5 show that the in- painting network struggles to output high-quality image contents in missing regions under large viewpoint changes, thus limiting the rendering quality. In this paper, we propose a novel framework for semantics-guided view synthesis of natural scenes by learn- ing prior from single-view image collections. Based on the observation that semantic masks have much lower complex- ity than images, we divide this task into two simpler sub- problems: we first generate semantic masks at novel views and then translate them to RGB images through SPADE. For view synthesis of semantic masks, the input semantic mask is first translated to a color image by SPADE, and a depth map is predicted from the color image by a depth es- timator [50]. Then, the input semantic mask is warped to novel views using the predicted depth map and refined by an inpainting network trained by a self-supervised learning strategy on single-view image collections. Our experiments show that, in contrast to images, the novel view synthesis of semantic masks is much easier to learn by the network. It is observed that semantic masks generated by the in- painting network tend to be view-inconsistent. As a result, SPADE could generate quite different contents in these re- gions even when the inconsistency is minor between the se- mantic masks. Fig. 4 presents two examples. To solve this issue, we learn a neural semantic field to fuse and denoise these semantic masks for better multi-view consistency. Fi- nally, we translate the multi-view semantic masks to color images by SPADE and reconstruct a neural scene represen- tation for view-consistent rendering. Extensive experiments are conducted on the LHQ dataset [58], a widely-used benchmark dataset for semantic image synthesis. The results demonstrate that our approach significantly outperforms baseline methods both qualita- tively and quantitatively. We also show that by editing the input semantic mask, our approach is capable of generating various high-quality rendering results of natural scenes, as shown in Fig. 1.
Zhang_Exploiting_Completeness_and_Uncertainty_of_Pseudo_Labels_for_Weakly_Supervised_CVPR_2023	Abstract Weakly supervised video anomaly detection aims to iden- tify abnormal events in videos using only video-level labels. Recently, two-stage self-training methods have achieved significant improvements by self-generating pseudo labels and self-refining anomaly scores with these labels. As the pseudo labels play a crucial role, we propose an enhance- ment framework by exploiting completeness and uncertainty properties for effective self-training. Specifically, we first design a multi-head classification module (each head serves as a classifier) with a diversity loss to maximize the distri- bution differences of predicted pseudo labels across heads. This encourages the generated pseudo labels to cover as many abnormal events as possible. We then devise an it- erative uncertainty pseudo label refinement strategy, which improves not only the initial pseudo labels but also the up- dated ones obtained by the desired classifier in the sec- ond stage. Extensive experimental results demonstrate the proposed method performs favorably against state-of-the- art approaches on the UCF-Crime, TAD, and XD-Violence benchmark datasets.	1. Introduction Automatically detecting abnormal events in videos has attracted increasing attention for its broad applications in intelligent surveillance systems. Since abnormal events are sparse in videos, recent studies are mainly developed within the weakly supervised learning framework [5,12,19,25,27, *Corresponding author. (b)(c) (a) Figure 1. Illustration of the completeness: (a) represents a video that contains multiple abnormal clips (ground truth anomalies are in the orange area). Existing methods tend to focus on the most anomalous clip as shown in (b). We propose to use the multi-head classification module together with a diversity loss to encourage pseudo labels to cover the complete abnormal events as depicted in (c). 29, 32, 34, 37–41], where only video-level annotations are available. However, the goal of anomaly detection is to pre- dict frame-level anomaly scores during test. This results in great challenges for weakly supervised video anomaly de- tection. Existing methods broadly fall into two categories: one- stage methods based on Multiple Instance Learning (MIL) and two-stage self-training methods. One-stage MIL-based 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. 16271 methods [19, 27, 29, 39, 41] treat each normal and abnor- mal video as a negative and positive bag respectively, and clips of a video are the instances of a bag. Formulating anomaly detection as a regression problem, these methods adopt ranking loss to encourage the highest anomaly score in a positive bag to be higher than that in a negative bag. Due to the lack of clip-level annotations, the anomaly scores generated by MIL-based methods are usually less accurate. To alleviate this problem, two-stage self-training methods are proposed [5, 12]. In the first stage, pseudo labels for clips are generated by MIL-based methods. In the second stage, MIST [5] utilizes these pseudo labels to refine dis- criminative representations. In contrast, MSL [12] refines the pseudo labels via a transformer-based network. Despite progress, existing methods still suffer two limitations. First, the ranking loss used in the pseudo label generator ignores the completeness of abnormal events. The reason is that a positive bag may contain multiple abnormal clips as shown in Figure 1, but MIL is designed to detect only the most likely one. The second limitation is that the uncertainty of generated pseudo labels is not taken into account in the sec- ond stage. As the pseudo labels are usually noisy, directly using them to train the final classifier may hamper its per- formance. To address these problems, we propose to enhance pseudo labels via exploiting completeness and uncertainty properties. Specifically, to encourage the complete detec- tion of abnormal events, we propose a multi-head module to generate pseudo labels (each head serves as a classifier) and introduce a diversity loss to ensure the distribution dif- ference of pseudo labels generated by the multiple classi- fication heads. In this way, each head tends to discover a different abnormal event, and thus the pseudo label genera- tor covers as many abnormal events as possible. Then, in- stead of directly training a final classifier with all pseudo la- bels, we design an iterative uncertainty-based training strat- egy. We measure the uncertainty using Monte Carlo (MC) Dropout [6] and only clips with lower uncertainty are used to train the final classifier. At the first iteration, we use such uncertainty to refine pseudo labels obtained in the first stage, and in the remaining iterations, we use it to refine the output of the desired final classifier. The main contributions of this paper are as follows: • We design a multi-head classifier scheme together with a diversity loss to encourage the pseudo labels to cover as many abnormal clips as possible. • We design an iterative uncertainty aware self-training strategy to gradually improve the quality of pseudo la- bels. • Experiments on UCF-Crime, TAD, and XD-Violence datasets demonstrate the favorable performance com- pared to several state-of-the-art methods.
Ye_NEF_Neural_Edge_Fields_for_3D_Parametric_Curve_Reconstruction_From_CVPR_2023	Abstract We study the problem of reconstructing 3D feature curves of an object from a set of calibrated multi-view im- ages. To do so, we learn a neural implicit ﬁeld repre- senting the density distribution of 3D edges which we re- fer to as Neural Edge Field (NEF). Inspired by NeRF [20], NEF is optimized with a view-based rendering loss where a 2D edge map is rendered at a given view and is com- pared to the ground-truth edge map extracted from the im- age of that view. The rendering-based differentiable opti- mization of NEF fully exploits 2D edge detection, without needing a supervision of 3D edges, a 3D geometric oper- ator or cross-view edge correspondence. Several technical designs are devised to ensure learning a range-limited and view-independent NEF for robust edge extraction. The ﬁ- nal parametric 3D curves are extracted from NEF with an iterative optimization method. On our benchmark with syn- thetic data, we demonstrate that NEF outperforms exist- ing state-of-the-art methods on all metrics. Project page: https://yunfan1202.github.io/NEF/.	1. Introduction Feature curves “deﬁne” 3D shapes to an extent, not only geometrically (surface reconstruction from curve net- works [15, 16]) but also perceptually (feature curve based shape perception [4, 35]). Therefore, feature curve extrac- *Corresponding author.tion has been a long-standing problem in both graphics and vision. Traditional approaches to 3D curve extraction often work directly on 3D shapes represented by, e.g., polygo- nal meshes or point clouds. Such approaches come with a major difﬁculty: Sharp edges may be partly broken or com- pletely missed due to imperfect 3D acquisition and/or re- construction. Consequently, geometrically-based methods, even the state-of-the-art ones, are sensitive to parameter set- tings and error-prune near rounded edges, noise, and sparse data. Recently, learning-based methods are proposed to ad- dress these issues but with limited generality [18,19,33,39]. In many cases, edges are visually prominent and easy to detect in the 2D images of a 3D shape. To resolve occlusion, one may think of 3D curve reconstruction from multi-view edges. This solution, however, relies strongly on cross-view edge correspondence which itself is a highly difﬁcult prob- lem [28]. This explains why there is rarely a work on multi- view curve reconstruction even in the deep learning era. We ask this question: Can we learn 3D feature curve extraction directly from the input of multi-view images? In this work, we try to answer the question through learn- ing a neural implicit ﬁeld representing the density distribu- tion of 3D edges from a set of calibrated multi-view im- ages, inspired by the recent success of neural radiance ﬁeld (NeRF) [20]. We refer to this edge density ﬁeld as Neu- ral Edge Field (NEF). Similar to NeRF, NEF is optimized with a view-based rendering loss where a 2D edge map is rendered at a given view and is compared to the ground- truth edge map extracted from the image of that view. 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. 8486 volumetric rendering is based on edge density and color (gray-scale) predicted by MLPs along viewing rays. Dif- ferent from NeRF, however, our goal is merely to optimize the NEF which is later used for extracting parametric 3D curves; no novel view synthesis is involved. The rendering- based differentiable optimization of 3D edge density fully exploits 2D edge detection, without needing a 3D geomet- ric operator or cross-view edge correspondence. The latter is implicitly learned with multi-view consistency. Directly optimizing NEF as NeRF-like density is prob- lematic since the range of density can be arbitrarily large and different from scene to scene, and it is hard to select a proper threshold to extract useful geometric shapes (e.g., 3D surfaces for NeRF and 3D edges for NEF). Moreover, NeRF density usually does not approximate the underlying 3D shape well due to noise. Therefore, we seek for conﬁn- ing the edge density in the range of [0;1]through learning a mapping function with a learnable scaling factor to map the edge density to the actual NEF density. By doing so, we can easily choose a threshold to extract edges robustly from the optimized edge density. Another issue with NEF optimization is the incompatible visibility of the edge density ﬁeld and the edges detected in images. While the former is basically a wireframe represen- tation of the underlying 3D shape and all edges are visible from any view (i.e., no self-occlusion), edges in 2D images can be occluded by the object itself. This leads to inconsis- tent supervisions of different views with different visibility and may cause false negative: An edge that should have been present in NEF according to one view visible to the edge may be suppressed by other views invisible. To ad- dress this issue, we opt to 1) impose consistency between density and color in NEF and 2) give less punishment on non-edge pixels in the rendering loss, to allow the NEF to keep all edges seen from all views. This essentially makes NEF view-independent which is reasonable. Having obtained the edge density, we ﬁt parametric curves by treating the 3D density volume as a point cloud of edges. We optimize the control points of curves in a coarse- to-ﬁne manner. Since initialization is highly important to such a non-convex optimization, we ﬁrst apply line ﬁtting in a greedy fashion to cover most points. Based on the ini- tialization, we then upgrade lines to cubic B ´ezier curves by adding extra control points and optimize all curves simulta- neously with an extra endpoint regularization. We build a benchmark with a synthetic dataset consist- ing of 115 CAD models with complicated shape structures from ABC dataset [14] and utilize BlenderProc [7] to ren- der posed images. Extensive experiments on the proposed dataset show that NEF, which is self-trained with only 2D supervisions, outperforms existing state-of-the-art methods on all metrics. Our contributions include: • A self-supervised 3D edge detection from multi-view2D edges based neural implicit ﬁeld optimization. • Several technical designs to ensure learning a range- limited and view-independent NEF and an iterative op- timization strategy to reconstruct parametric curves. • A benchmark for evaluating and comparing various edge/curve extraction methods.
Yu_CelebV-Text_A_Large-Scale_Facial_Text-Video_Dataset_CVPR_2023	Abstract Text-driven generation models are flourishing in video generation and editing. However, face-centric text-to-video generation remains a challenge due to the lack of a suitable dataset containing high-quality videos and highly relevant texts. This paper presents CelebV-Text , a large-scale, di- verse, and high-quality dataset of facial text-video pairs, to facilitate research on facial text-to-video generation tasks. CelebV-Text comprises 70,000 in-the-wild face video clips with diverse visual content, each paired with 20 texts gen- erated using the proposed semi-automatic text generation strategy. The provided texts are of high quality, describ- ing both static and dynamic attributes precisely. The supe- riority of CelebV-Text over other datasets is demonstrated via comprehensive statistical analysis of the videos, texts, and text-video relevance. The effectiveness and potential of CelebV-Text are further shown through extensive self- evaluation. A benchmark is constructed with representa- tive methods to standardize the evaluation of the facial text- to-video generation task. All data and models are publicly available1. *Equal contribution. 1Project page: https://celebv-text.github.io	1. Introduction Text-driven video generation has recently garnered sig- nificant attention in the fields of computer vision and com- puter graphics. By using text as input, video content can be generated and controlled, inspiring numerous applications in both academia and industry [5,34,43,47]. However, text- to-video generation still faces many challenges, particularly in the face-centric scenario where generated video frames often lack quality [18, 34, 37] or have weak relevance to in- put texts [2,4,39,67]. We believe that one of the main issues is the absence of a well-suited facial text-video dataset con- taining high-quality video samples and text descriptions of various attributes highly relevant to videos. Constructing a high-quality facial text-video dataset poses several challenges, mainly in three aspects. 1) Data collection. The quality and quantity of video samples largely determine the quality of generated videos [11, 45, 48, 60]. However, obtaining such a large-scale dataset with high-quality samples while maintaining a natural distribu- tion and smooth video motion is challenging. 2) Data an- notation. The relevance of text-video pairs needs to be en- sured. This requires a comprehensive coverage of text for describing the content and motion appearing in the video, such as light conditions and head movements. 3) Text gen- 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. 14805 eration. Producing diverse and natural texts are non-trivial. Manual text generation is expensive and not scalable. While auto-text generation is easily extensible, it is limited in nat- uralness. To overcome the challenges mentioned above, we care- fully design a comprehensive data construction pipeline that includes data collection and processing, data annotation, and semi-auto text generation. First, to obtain raw videos, we follow the data collection steps of CelebV-HQ, which has proven to be effective in [66]. We introduce a minor modification to the video processing step to improve the video’s smoothness further. Next, to ensure highly relevant text-video pairs, we analyze videos from both temporal dy- namics and static content and establish a set of attributes that may or may not change over time. Finally, we propose a semi-auto template-based method to generate texts that are diverse and natural. Our approach leverages the advan- tages of both auto- and manual-text methods. Specifically, we design a rich variety of grammar templates as [10,52] to parse annotation and manual texts, which are flexibly com- bined and modified to achieve high diversity, complexity, and naturalness. With the proposed pipeline, we create CelebV-Text , a Large-Scale Facial Text-Video Dataset, which includes 70,000in-the-wild video clips with a resolution of at least 512×512and1,400,000text descriptions with 20 for each clip. As depicted in Figure 1, CelebV-Text consists of high- quality video samples and text descriptions for realistic face video generation. Each video is annotated with three types of static attributes (40 general appearances, 5 detailed ap- pearances, and 6 light conditions) and three types of dy- namic attributes (37 actions, 8 emotions, and 6 light direc- tions). All dynamic attributes are densely annotated with start and end timestamps, while manual-texts are provided for labels that cannot be discretized. Furthermore, we have designed three templates for each attribute type, resulting in a total of 18 templates that can be flexibly combined. All attributes and manual-texts are naturally described in our generated texts. CelebV-Text surpasses existing face video datasets [11] in terms of resolution (over 2 times higher), number of sam- ples, and more diverse distribution. In addition, the texts in CelebV-Text exhibit higher diversity, complexity, and natu- ralness than those in text-video datasets [19, 66]. CelebV- Text also shows high relevance of text-video pairs, validated by our text-video retrieval experiments [17]. To further ex- amine the effectiveness and potential of CelebV-Text, we evaluate it on a representative baseline [19] for facial text- to-video generation. Our results show better relevance be- tween generated face videos and texts when compared to a state-of-the-art large-scale pretrained model [26]. Fur- thermore, we show that a simple modification of [19] with text interpolation can significantly improve temporal coher- ence. Finally, we present a new benchmark for text-to-video generation to standardize the facial text-to-video generation task, which includes representative models [5, 19] on three text-video datasets. The main contributions of this work are summarized asfollows: 1) We propose CelebV-Text, the first large-scale facial text-video dataset with high-quality videos, as well as rich and highly-relevant texts, to facilitate research in fa- cial text-to-video generation. 2) Comprehensive statistical analyses are conducted to examine video/text quality and diversity, as well as text-video relevance, demonstrating the superiority of CelebV-Text. 3) A series of self-evaluations are performed to demonstrate the effectiveness and poten- tial of CelebV-Text. 4) A new benchmark for text-to-video generation is constructed to promote the standardization of the facial text-to-video generation task.
Zhang_Delivering_Arbitrary-Modal_Semantic_Segmentation_CVPR_2023	Abstract Multimodal fusion can make semantic segmentation more robust. However, fusing an arbitrary number of modalities remains underexplored. To delve into this prob- lem, we create the DELIVER arbitrary-modal segmenta- tion benchmark, covering De pth, Li DAR, multiple V iews, Events, and R GB. Aside from this, we provide this dataset in four severe weather conditions as well as ﬁve sensor failure cases to exploit modal complementarity and resolve par- tial outages. To make this possible, we present the arbi- trary cross-modal segmentation model CMN EXT. It en- compasses a Self-Query Hub (SQ-Hub) designed to extract effective information from any modality for subsequent fu- sion with the RGB representation and adds only negligible amounts of parameters ( ∼0.01M) per additional modal- ity. On top, to efﬁciently and ﬂexibly harvest discrimina- tive cues from the auxiliary modalities, we introduce the simple Parallel Pooling Mixer (PPX) . With extensive experi- ments on a total of six benchmarks, our CMN EXTachieves state-of-the-art performance on the DELIVER , KITTI-360, MFNet, NYU Depth V2, UrbanLF , and MCubeS datasets, allowing to scale from 1to81modalities. On the freshly collected DELIVER , the quad-modal CMN EXTreaches up to66.30% in mIoU with a +9.10% gain as compared to the mono-modal baseline.1	1. Introduction With the explosion of modular sensors, multimodal fu- sion for semantic segmentation has progressed rapidly re- cently [ 5,11,48] and in turn has stirred growing inter- est to assemble more and more sensors to reach higher and higher segmentation accuracy aside from more robust scene understanding. However, most works [ 34,75,103] and multimodal benchmarks [ 29,61,91] focus on speciﬁc sensor pairs, which lack behind the current trend of fusing *Equal contribution. †Corresponding author (e-mail: [email protected] ). 1The D ELIVER dataset and our code will be made publicly available at:https://jamycheung.github.io/DELIVER.html . D-E- -D- D-E- 5256606468mIoU (%) (a) RGB-D-E-L fusion. D-E- -D- D-E- 46474849505152mIoU (%) (b) RGB-A-D-N fusion. D-E- -D- D-E- 7778798081mIoU (%) (c) RGB-Light Field. Figure 1. Arbitrary-modal segmentation results of CMNeXt using: (a).{RGB ,Depth,Event,LiDAR}on our D ELIVER dataset; (b).{RGB ,Angle of Linear Polarization ( AoLP) ,Degree of Lin- ear Polarization ( DoLP) ,Near-Infrared ( NIR)}on MCubeS [ 44]; (c).{RGB ,8/33/80 sub-aperture Light Fields (LF8/LF33/LF80) on UrbanLF-Syn [ 59], respectively. LiDAR Event Depth RGB 57.2 57.0 49.1 38.3 57.2 57.4 65.37 65.92mIoU (%) Figure 2. Comparing CMX [ 48], HRFuser [ 4], and our CMNeXt in sensor failure ( i.e., LiDAR Jitter) on the D ELIVER dataset. more and more modalities [ 4,70],i.e., progressing towards Arbitrary-Modal Semantic Segmentation (AMSS). When looking into AMSS, two observations become ap- parent. Firstly, an increasing amount of modalities should provide more diverse complementary information, mono- tonically increasing segmentation accuracy. This is di- rectly supported by our results when incrementally adding and fusing modalities as illustrated in Fig. 1a(RGB- Depth-Event-LiDAR), Fig. 1b(RGB-AoLP-DoLP-NIR), and Fig. 1cwhen adding up to 80sub-aperture light-ﬁeld modalities (RGB-LF 8/-LF33/-LF80). Unfortunately, this great potential cannot be uncovered by previous cross- 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. 1136 Hub distribute ..hub2fusemergeR D LER D LER D LE (a) Separate (b) Joint (c) Asymmetric Figure 3. Comparison of multimodal fusion paradigms, such as (a) merging with separate branches [ 4], (b) distributing with a joint branch [ 70], and (c) our hub2fuse with asymmetric branches. modal fusion methods [ 9,77,99], which follow designs for pre-deﬁned modality combinations. The second observa- tion is that the cooperation of multiple sensors is expected to effectively combat individual sensor failures. Most of the existing works [ 67,72,76] are built on the assump- tion that each modality is always accurate. Under par- tial sensor faults, which are common in real-life robotic systems, e.g. LiDAR Jitter, fusing misaligned sensing data might even degrade the segmentation performance, as de- picted with CMX [ 48] and HRFuser [ 4] in Fig. 2. These two critical observations remain to a large extent neglected. To address these challenges, we create a benchmark based on the CARLA simulator [ 19], with Depth,LiDAR, Views, Events, and RGB images: The DELIVER Multi- modal dataset. It features severe weather conditions and ﬁve sensor failure modes to exploit complementary modalities and resolve partial sensor outages. To proﬁt from all this, we present the arbitrary cross-modal CMNeXt segmenta- tion model. Without increasing the computation overhead substantially when adding more modalities CMNeXt incor- porates a novel Hub2Fuse paradigm (Fig. 3c). Unlike re- lying on separate branches (Fig. 3a) which tend to be com- putationally costly or using a single joint branch (Fig. 3b) which often discards valuable information, CMNeXt is an asymmetric architecture with two branches, one for RGB and another for diverse supplementary modalities. The key challenge lies in designing the two branches to pick up multimodal cues. Speciﬁcally, at the hubstep ofHub2Fuse , to gather useful complementary informa- tion from auxiliary modalities, we design a Self-Query Hub (SQ-Hub), which dynamically selects informative fea- tures from all modality-sources before fusion with the RGB branch. Another great beneﬁt of SQ-Hub is the ease of ex- tending it to an arbitrary number of modalities, at negligible parameters increase ( ∼0.01Mper modality). At the fusion step, fusing sparse modalities such as LiDAR or Event data can be difﬁcult to handle for joint branch architectures with- out explicit fusion such as TokenFusion [ 70]. To circum- vent this issue and make best use of both dense and sparse modalities, we leverage cross-fusion modules [ 48] and cou- ple them with our proposed Parallel Pooling Mixer (PPX)which efﬁciently and ﬂexibly harvests the most discrimina- tive cues from any auxiliary modality. These design choices come together in our CMNeXt architecture, which paves the way for AMSS (Fig. 1). By carefully putting together alter- native modalities, CMNeXt can overcome individual sensor failures and enhances segmentation robustness (Fig. 2). With comprehensive experiments on D ELIVER and ﬁve additional public datasets, we gather insight into the strength of the CMNeXt model. On D ELIVER, CMNeXt obtains66.30% in mIoU with a +9.10% gain compared to the RGB-only baseline [ 78]. On UrbanLF-Real [ 59] and MCubeS [ 44] datasets, CMNeXt surpasses the previ- ous best methods by +3.90% and+8.68%, respectively. Compared to previous state-of-the-art methods, our model achieves comparable perfomance on bi-modal NYU Depth V2 [61] as well as MFNet [ 29] and outperforms all previous modality-speciﬁc methods on KITTI-360 [ 45]. On a glance, we deliver the following contributions: • We create the new benchmark D ELIVER for Arbitrary-Modal Semantic Segmentation (AMSS) with four modalities, four adverse weather conditions, and ﬁve sensor failure modes. • We revisit and compare different multimodal fusion paradigms and present the Hub2Fuse paradigm with an asymmetric architecture to attain AMSS. • The universal arbitrary cross-modal fusion model CM- NeXt is proposed, with a Self-Query Hub (SQ-Hub) for selecting informative features and a Parallel Pool- ing Mixer (PPX) for harvesting discriminative cues. • We investigate AMSS by fusing up to a total of 80 modalities and notice that CMNeXt achieves state-of- the-art performances on six datasets.
Yao_Explicit_Boundary_Guided_Semi-Push-Pull_Contrastive_Learning_for_Supervised_Anomaly_Detection_CVPR_2023	Abstract Most anomaly detection (AD) models are learned using only normal samples in an unsupervised way, which mayresult in ambiguous decision boundary and insufﬁcient dis- criminability. In fact, a few anomaly samples are often available in real-world applications, the valuable knowl-edge of known anomalies should also be effectively ex- ploited. However , utilizing a few known anomalies dur-ing training may cause another issue that the model maybe biased by those known anomalies and fail to generalize to unseen anomalies. In this paper , we tackle supervised anomaly detection, i.e., we learn AD models using a few available anomalies with the objective to detect both theseen and unseen anomalies. We propose a novel explicitboundary guided semi-push-pull contrastive learning mech-anism, which can enhance model’s discriminability whilemitigating the bias issue. Our approach is based on twocore designs: First, we ﬁnd an explicit and compact sepa- rating boundary as the guidance for further feature learn- ing. As the boundary only relies on the normal feature dis-tribution, the bias problem caused by a few known anoma-lies can be alleviated. Second, a boundary guided semi-push-pull loss is developed to only pull the normal fea-tures together while pushing the abnormal features apart from the separating boundary beyond a certain margin re-gion. In this way, our model can form a more explicit and discriminative decision boundary to distinguish known andalso unseen anomalies from normal samples more effec- tively. Code will be available at http s:// github. com/xcyao00/BGAD .	1. Introduction Anomaly detection (AD) has received widespread atten- tion in diverse domains, such as industrial defect inspec- Corresponding Author.tion [ 4,8,10,50] and medical lesion detection [ 12,38]. Most previous anomaly detection methods [ 1,3,5,6,8,10,15,30, 33,47,50–52] are unsupervised and pay much attention to normal samples while inadvertently overlooking anomalies,because it is difﬁcult to collect sufﬁcient and all kinds of anomalies. However, learning only from normal samples may limit the discriminability of the AD models [ 12,24]. As illustrated in Figure 1(a), without anomalies, the de- cision boundaries are generally implicit and not discrimi- native enough. The insufﬁcient discriminability issue is a common issue in unsupervised anomaly detection due to thelack of knowledge about anomalies. In fact, a few anoma-lies are usually available in real-world applications, whichcan be exploited effectively to address or alleviate this issue. Recently, methods that can be called semi-supervised AD [ 27,34,36] or AD with outlier exposure [ 16,17] begin to focus on those available anomalies. These methods attemptto learn knowledge from anomalies by one-class classiﬁca-tion with anomalies as negative samples [ 34,36]o rb ys u - pervised binary classiﬁcation [ 16,17] or by utilizing the de- viation loss to optimize one anomaly scoring network [ 27]. They show a fact that the detection performance can be im- proved signiﬁcantly even with a few anomalies. However, the known anomalies can’t represent all kinds of anomalies.These methods may be biased by the known anomalies and fail to generalize to unseen anomalies (see Figure 5). Therefore, to address the two above issues, we tackle supervised anomaly detection [ 12], in which a few known anomalies can be effectively exploited to train discrimina- tive AD models with the objective to improve detection per-formance on the known anomalies and generalize well to unseen anomalies. Compared with unsupervised AD, su- pervised AD is more meaningful for real-world AD appli- cations, because the detected anomalies can be used to fur-ther improve the discriminability and generalizability of themodel. To this end, we propose a novel Boundary Guided Anomaly Detection ( BGAD ) model, which has two core 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. 24490 /RJOLNHOLKRRG,PSOLFLW %RXQGDULHV 'HQVLW\'HQVLW\'HQVLW\ $EQRUPDO %RXQGDU\1RUPDO %RXQGDU\)LQGLQJ([SOLFLW6HSDUDWLQJ%RXQGDU\ ([SOLFLW %RXQGDULHV $PELJXRXV /RJOLNHOLKRRG /RJOLNHOLKRRG 8QDPELJXRXV D E F([SOLFLW%RXQGDU\HQHUDWLQJ3KDVH %RXQGDU\*XLGHG2SWLPL]LQJ3KDVH Figure 1. Conceptual illustration of our method. (a) In most unsupervised AD models, the anomaly score distribution usually has ambigu- ous regions, which makes it difﬁcult to get one ideal decision boundary. E.g. , the left boundary will cause many false negatives, while the right boundary may induce many false positives. (b) With the normalized normal feature distribution, a pair of explicit and compact (close to the normal distribution) boundaries can be obtained easily. (c) With the proposed BG-SPP loss, boundary guided optimizing canbe implemented to obtain an unambiguous anomaly score distribution: a signiﬁcant gap between the normal and abnormal distributions. designs as illustrated in Figure 1: explicit boundary gener- ating and boundary guided optimizing. •Explicit Boundary Generating. We ﬁrst employ nor- malizing ﬂow [ 14] to learn a normalized normal feature dis- tribution, and obtain an explicit separating boundary, whichis close to the normal feature distribution edge and con- trolled by a hyperparameter β(i.e., the normal boundary in Figure 1(b)). The obtained explicit separating boundary only relies on the normal distribution and has no relation with the abnormal samples, thus the bias problem causedby the insufﬁcient known anomalies can be mitigated. •Boundary Guided Optimizing. After obtaining the explicit separating boundary, we then propose a boundaryguided semi-push-pull (BG-SPP) loss to exploit anomalies for learning more discriminative features. With the BG-SPP loss, only the normal features whose log-likelihoods are smaller than the boundary are pulled together to forma more compact normal feature distribution (semi-pull); while the abnormal features whose log-likelihoods arelarger than the boundary are pushed apart from the bound- ary beyond a certain margin region (semi-push). In this way, our model can form a more explicit and dis- criminative separating boundary and also a reliable margin region for distinguishing anomalies more effectively (see Figure 1(c), 6). Furthermore, rarity is a critical problem of anomalies and may cause feature learning inefﬁcient. We thus propose RandAugment-based Pseudo Anomaly Gener- ation, which can simulate anomalies by creating local irreg-ularities in normal samples, to tackle the rarity challenge. In summary, we make the following main contributions:1. We propose a novel Explicit Boundary Guided super- vised AD modeling method, in which both normal and ab-normal samples are exploited effectively by well-designedexplicit boundary generating and boundary guided optimiz-ing. With the proposed AD method, higher discriminabilityand lower bias risk can be achieved simultaneously. 2. To exploit a few known anomalies effectively, we pro- pose a BG-SPP loss to pull together normal features while pushing abnormal features apart from the separating bound-ary, thus more discriminative features can be learned. 3. We achieve SOTA results on the widely-used MVTecAD benchmark, with the performance of 99.3% image-level AUROC and 99.2% pixel-level AUROC.
Yang_Progressive_Open_Space_Expansion_for_Open-Set_Model_Attribution_CVPR_2023	Abstract Despite the remarkable progress in generative technol- ogy, the Janus-faced issues of intellectual property protec- tion and malicious content supervision have arisen. Efforts have been paid to manage synthetic images by attributing them to a set of potential source models. However, the closed-set classification setting limits the application in real-world scenarios for handling contents generated by arbitrary models. In this study, we focus on a challenging task, namely Open-Set Model Attribution (OSMA), to simul- taneously attribute images to known models and identify those from unknown ones. Compared to existing open- set recognition (OSR) tasks focusing on semantic novelty, OSMA is more challenging as the distinction between images from known and unknown models may only lie in visually imperceptible traces. To this end, we propose a Progressive O pen S pace E xpansion (POSE) solution, which simulates open-set samples that maintain the same se- mantics as closed-set samples but embedded with different imperceptible traces. Guided by a diversity constraint, the open space is simulated progressively by a set of lightweight augmentation models. We consider three real-world sce- narios and construct an OSMA benchmark dataset, includ- ing unknown models trained with different random seeds, architectures, and datasets from known ones. Extensive experiments on the dataset demonstrate POSE is superior to both existing model attribution methods and off-the-shelf OSR methods. Github: https://github.com/ICTMCG/POSE	1. Introduction Advanced generative modeling technology can create extremely realistic visual content, leading to dramatic changes in the field of AI-enhanced design, arts, and meta- universe [23, 41, 47]. Whereas, the broadcasting of mali- cious content generated by open source generation models *Corresponding author Figure 1. Open-set model attribution problem: The unknown classes include unknown models different from known models in training seeds, architectures, or training datasets. The goal is to simultaneously attribute images to known models and identify those from unknown ones. has brought severe social impacts [9, 18, 43]. Furthermore, new challenges have arisen for the ownership protection of copyrighted digital generative models. To solve these problems, model attribution, i.e., identifying the source model of generated contents, has drawn increasing attention recently [5, 33, 48, 50, 53]. Marra et al . [33] are among the first to point out that GAN models leave specific fingerprints in the generated images, just like camera devices. Further researches [5, 14, 48, 50, 53] validate the existence of GAN fingerprints and show the feasibility of attributing fake images to a fixed and finite set of known models. However, most of these works focus on finding discriminative fingerprints among the contents generated by different GAN models following a simple closed-set setup. The ever-growing number of unseen source models in the real-world scenario appeal for a more generic approach. In this paper, we focus on 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. 15856 the problem of Open-Set Model Attribution (OSMA), i.e., simultaneously attributing images to known source models and identifying those from unknown ones. An intuitive way to solve OSMA is to apply off-the-shelf open-set recognition (OSR) approaches to closed-set model attribution classifiers. Traditional OSR methods leverage the output logits to either reject or categorize the input images [2, 44]. However, following the discriminative line, the performance highly depends on the closed-set classi- fier [46]. The learned feature space is not rectified for open- set samples. Another mainstream of OSR methods is based on simulating open-set samples or features [7,27,35,36,58]. By the simulation of open space, the learned feature space is more compact for closed-set categories [58], leading the detection of unknown samples more valid. Nevertheless, existing works only leverage a single generator [7, 27, 35] or mechanism [58] for simulating open-set samples or features, which are not diverse enough to reduce the open space risk for OSMA. A generator could produce open-set samples of different semantics, but its fingerprint is fixed and thus not suitable for the expansion of open space. In this study, we propose Progressive Open Space Ex- pansion (POSE) tailored for open-set model attribution, which simulates the potential open space of challenging unknown models through involving a set of augmentation models progressively. For augmentation model construc- tion, it can be daunting to consider all types of unknown models with a variety of architectures. Instead, lightweight networks composed of a few convolution layers are em- ployed. They serve as “virtual” follow-up blocks of known models, augmenting closed-set samples to open-set samples surrounding them by modifying their fingerprints with re- construction residuals. Despite the simple structure, these augmentation models show the potential to model traces of a variety of unknown models. To enrich the simulated open space, multiple augmentation models are involved. Instead of training them independently, we design a progressive training mechanism to ensure the diversity of simulated open space across models in a computation-effective way. To validate the effectiveness of POSE in the open world, we construct a benchmark dataset considering three chal- lenging unknown scenarios as shown in Figure 1, which includes unknown models trained with either a different random seed, architecture or dataset from known mod- els. Extensive experiments on the benchmark demonstrate POSE is superior to both existing GAN attribution methods and OSR methods. In summary, our contributions are: •We tackle an important challenge for applying model attribution to open scenarios, the open-set model attribution problem, which attributes images to known models and identifies images from unknown ones. •We propose a novel solution named POSE, which sim- ulates the potential open space of unknown models pro-gressively by a set of lightweight augmentation models, and consequently reduces open space risk. •We construct an OSMA benchmark simulating the real- world scenarios, on which extensive experiments prove the superiority of POSE compared with existing GAN attribution methods and off-the-shelf OSR methods.
Yu_On_the_Difficulty_of_Unpaired_Infrared-to-Visible_Video_Translation_Fine-Grained_Content-Rich_CVPR_2023	Abstract Explicit visible videos can provide sufficient visual in- formation and facilitate vision applications. Unfortunately, the image sensors of visible cameras are sensitive to light conditions like darkness or overexposure. To make up for this, recently, infrared sensors capable of stable imaging have received increasing attention in autonomous driving and monitoring. However, most prosperous vision mod- els are still trained on massive clear visible data, facing huge visual gaps when deploying to infrared imaging sce- narios. In such cases, transferring the infrared video to a distinct visible one with fine-grained semantic patterns is a worthwhile endeavor. Previous works improve the out- puts by equally optimizing each patch on the translated vis- ible results, which is unfair for enhancing the details on content-rich patches due to the long-tail effect of pixel dis- tribution. Here we propose a novel CPTrans framework to tackle the challenge via balancing gradients of different patches, achieving the fine-grained Content-rich Patches Transferring. Specifically, the content-aware optimization module encourages model optimization along gradients of target patches, ensuring the improvement of visual details. Additionally, the content-aware temporal normalization module enforces the generator to be robust to the motions of target patches. Moreover, we extend the existing dataset In- fraredCity to more challenging adverse weather conditions (rain and snow), dubbed as InfraredCity-Adverse1. Exten- sive experiments show that the proposed CPTrans achieves state-of-the-art performance under diverse scenes while re- quiring less training time than competitive methods.	1. Introduction Visible light cameras have broad applicability in com- puter vision algorithms for the sufficient visual informa- Corresponding author 1The code and dataset are available at https://github.com/BIT-DA/I2V- Processing ROMA OursOutput GradCAM++ (a) (b) Figure 1. (a) Visualization of pixel category distribution on dataset IRVI [25] and semantic examples in random selected frames. We conduct semantic segmentation via a pre-trained SegFormer [47] on all visible video frames of IRVI and predict all pixels accord- ing to the predefined categories in ADE20K [52]. (b) Outputs and GradCAM++ results of different methods. ROMA pays equal at- tention on the whole output, and the long-tail effect of training data leads to the generation optimization along prejudiced gradi- ents caused by the large proportion of pixels (e.g., sky and road). We can generate more vivid details for content-rich patches (e.g., cars and road signs) than other methods. tion (e.g., structure, texture, and color) of their captured results. Most state-of-the-art vision algorithms have been observed to show admirable performance under clear visi- bility conditions [10, 16, 50]. Unfortunately, in most cases, the real-world weather is unpredictable and diverse, leading to complex and variable light conditions like overexposure on snowing days. While image sensors of visible cameras are sensitive to light conditions, their imaging results are ambiguous in adverse weather. Under such circumstances, people take infrared sensors to make up for the deficien- cies of visible cameras. These infrared sensors can capture stable structural information in diverse environments due to the thermal imaging principle. In emergency avoidance or hazard detection, they could be applied in autonomous driv- ing and monitoring scenarios [28,30]. However, most com- puter vision models are trained under visible data. Although infrared videos outline surrounding objects all the time, the existing huge gaps and semantic lacking problems hinder 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. 1631 the applications in infrared imaging scenarios. Therefore, it is worth translating stable and accessible infrared videos into clear visible ones. The translated visible results may provide visual information for supporting visual applica- tions like object detection and semantic segmentation. To tackle the unpaired infrared-to-visible translation challenge, previous methods [12, 13, 33, 37] mainly fo- cus on learning color mapping functions with complex manual coloring. The high costs and inevitable human bias limit the application of such approaches. Inspired by GANs [11], unpaired image translation methods have emerged. For instance, cycle-based methods [17,21,22,48] preserve content during the translation via the cycle con- sistency [19]. Furthermore, one-sided methods [20, 31, 51] maintain the content through hand-designed feature-level restraints. However, substantial visual gaps between in- frared and visible data lead to difficulties in generating fine-grained visible results. Additionally, continuous in- frared video signals are more challenging to transfer be- cause of the need to ensure temporal consistency. Thus, taking long-term information into account, [3, 7, 25] pro- pose their temporal consistency losses to refine frameworks based on unpaired image translation methods. Besides, I2V-GAN [25] and ROMA [49] are tailored approaches for unpaired infrared-to-visible video translation. Espe- cially, ROMA has achieved state-of-art performance, il- lustrating the importance of retaining structural informa- tion and proposing cross-similarity consistency for struc- ture. Despite its success, experiments indicate that cross- similarity still faces challenges in accurately transferring fine-grained (i.e., realistic and delicate) details, especially for the content-rich patches. In fact, most GAN-based methods utilize the PatchGAN discriminator [19] for style optimization. Similar to the classification task, the discriminator outputs w×hpredic- tions (True or False) for corresponding patches. To ana- lyze the optimizing behavior of discriminators in the train- ing process, we visualize the gradients via GradCAM++ [6] and pixel category distribution as shown in Fig. 1. Grad- CAM++ utilizes the gradients of the classification score to identify the parts of interest. The left part (a) shows that a few majority categories occupy most of the pixels while most minority categories contain a limited number of pix- els. Additionally, content-rich patches (including rich vi- sual details like patches of cars) are mostly the minority categories, while those content-lacking patches (including lacking visual details like patches of the sky) are mostly the majority. Upon exposure to new data, gradient-based optimization methods, without any constraint, change the learned encoding to minimize the objective function with global respect [36]. Thus, equal optimization for each patch (GradCAM++ of ROMA on Fig. 1 (b)) faces prejudiced gra- dients to content-lacking patches (i.e., major pixels) whenapplied to the generation. Moreover, it will lead to the in- ability of discriminators to improve the qualities of content- rich patches. An approach is needed to break the prejudice on optimization caused by the usually exhibiting long-tail distribution in real-world training data [24, 45, 54]. In this paper, we start with the analysis of difficulty for fine-grained Content-rich Patches Trans fer on unpaired infrared-to-visible video translation and propose the CP- Trans framework To improve the results of content-rich patches, we introduce two novel modules: Content-aware Optimization (CO), balancing the gradients of patches for improving generated content-rich patches, and Content- aware Temporal Normalization (CTN), which enforces the generator to be robust to the motion of content. Be- sides, we extend the InfraredCity dataset to adverse weather conditions (i.e., raining and snowing scenes), noted as InfraredCity-Adverse , for promoting infrared-related re- search. Our extensive evaluations of diverse datasets show that our approach improves upon the previous ROMA method, setting new state-of-the-art performances on un- paired infrared-to-visible video translation. Remarkably, further applications validate our task’s value and confirm our approach’s admirable performance. Contributions are: • We focus on the difficulty of fine-grained unpaired infrared-to-visible video translation and point out the ex- isting problem that models are optimized along preju- diced gradients due to the long-tail effect. • We propose a novel CPTrans framework consisting of content-aware optimization and temporal normalization, which benefits the generation of content-rich patches. • We extend the InfraredCity to more challenging ad- verse weather conditions (rain and snow), noted as InfraredCity-Adverse for infrared-related study and val- idate the remarkable success of CPTrans through suffi- cient experiments (including further applications, i.e., ob- ject detection, video fusion, and semantic segmentation).
Yin_AGAIN_Adversarial_Training_With_Attribution_Span_Enlargement_and_Hybrid_Feature_CVPR_2023	Abstract The deep neural networks (DNNs) trained by adversarial training (AT) usually suffered from significant robust gener- alization gap, i.e., DNNs achieve high training robustness but low test robustness. In this paper, we propose a generic method to boost the robust generalization of AT methods from the novel perspective of attribution span. To this end, compared with standard DNNs, we discover that the gen- eralization gap of adversarially trained DNNs is caused by the smaller attribution span on the input image. In other words, adversarially trained DNNs tend to focus on specific visual concepts on training images, causing its limitation on test robustness. In this way, to enhance the robustness, we propose an effective method to enlarge the learned at- tribution span. Besides, we use hybrid feature statistics for feature fusion to enrich the diversity of features. Extensive experiments show that our method can effectively improves robustness of adversarially trained DNNs, outperforming previous SOTA methods. Furthermore, we provide a the- oretical analysis of our method to prove its effectiveness.	1. Introduction Deep neural networks (DNNs) have shown remarkable success in solving complex prediction tasks. However, re- cent studies have shown that they are particularly vulnera- ble to adversarial attacks [22], which take the form of small perturbations to the input that cause DNNs to predict in- * Zhen Xiao and Kelu Yao are the corresponding authors. (a) (b) (c) (d) (e)ASC: 0.585 ASC: 0.491 ASC: 0.562 ASC: 0.479 ASC: 0.583 ASC: 0.497Figure 1. A visual illustration of attribution span under ResNet- 18. (a) is the original image; (b) and (c) are attribution spans of the standard model and robust model in the inference phase, re- spectively. ASC is Attribution SpanCoverage; (d) is the differ- ence between the standard model and the robust model in terms of attribution span; (e) is the result after partial feature erasure of the original image using (d). correct outputs. The defense of adversarial examples has been intensively studied in recent years and several defenses against adversarial attacks have been proposed in a great deal of work [17, 18]. Among the various existing defense strategies, adversar- ial training (AT) [10, 15] has been shown to be one of the most effective defenses [16] and has received a lot of atten- tion from the research community. However, adversarially 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. 20544 trained DNNs typically show a significant robust general- ization gap [27]. Intuitively, there is a large gap between the training robustness and test robustness of the adversarially trained model on the adversarial examples. Some existing methods [23, 25, 27] narrow the robust generalization gap from the perspective of weight loss landscapes. Other exist- ing methods [13,24,32] enhance robust generalization from the perspective of training strategies. However, this work ignores a critical factor affecting generalization robustness, which is the learned knowledgeable representation. Training DNNs with robust generalization is particu- larly difficult, typically possessing significantly higher sam- ple complexity [8, 29, 31] and requiring more knowledge- able [4, 19]. Compared with standard DNNs, we discover that the generalization gap of adversarially trained DNNs is caused by the smaller attribution span on the input im- age. In other words, adversarially trained DNNs tend to focus on specific visual concepts on training images [8], causing its limitation on test robustness. Specifically, we explore the difference between the standard model (training w/o AT) and the robust model (training w/ AT) in the in- ference phase through empirical experiments. As shown in Figure 1 (b) and Figure 1 (c), the standard model and the robust model have different attribution span for the same image in the inference phase, and the attribution span of the standard model is larger than that of the robust model in general. Through our further exploration, we find that these different spans (see Figure 1 (d)) affect the model’s deci- sion on clean data and hardly affect the model’s decision on adversarial examples. This indicates that AT enables the model to learn robust features, but ignores the features of generalization. This motivates us to design a method to en- large the attribution span to ensure that the model focuses on robust features while enhancing the focus on other features to improve the generalization ability of the robust model. To this end , we propose a generic method to boost the robust generalization of AT from the novel perspective of attribution span. Specifically, we use the class activation mapping to obtain the attribution span of the model under real and fake labels, and mix these two spans proportion- ally to complete the enlargement of the attribution span and make the model focus on the features within this span dur- ing the training process. In addition, in order to increase the diversity of features and ensure the stable training of the model under the enlarged attribution span, we adopt the fea- ture fusion implemented by hybrid feature statistics to fur- ther improve the generalization ability of the model. Com- pared to other methods, our method can further improve the accuracy of the model on clean data and adversarial exam- ples. Meanwhile, our work provides new insights into the lack of good generalization of robust models. Our main contributions are summarized as follows. • We find that adversarially trained DNNs focus on asmaller span of features in the inference phase and ignores some other spans of features. These spans are generally associated with generalization ability and have little impact on robustness. • We propose a method to boost AT, called AGAIN, which is short for Attribution Span Enlar Gement and Hybrid Fe Ature Fus IoN. During model training, we expand the region where the model focuses its features while ensuring that it learns robust features, and com- bine feature fusion to enhance the generalization of the model over clean data and adversarial examples. • Extensive experiments have shown that our proposed method can better improve the accuracy of the model on clean data and adversarial examples compared to state-of-the-art AT methods. Particularly, it can be eas- ily combined with other methods to further enhance the effectiveness of the method.
Yoshida_Light_Source_Separation_and_Intrinsic_Image_Decomposition_Under_AC_Illumination_CVPR_2023	Abstract Artificial light sources are often powered by an elec- tric grid, and then their intensities rapidly oscillate in re- sponse to the grid’s alternating current (AC). Interestingly, the flickers of scene radiance values due to AC illumina- tion are useful for extracting rich information on a scene of interest. In this paper, we show that the flickers due to AC illumination is useful for intrinsic image decomposition (IID). Our proposed method conducts the light source sepa- ration (LSS) followed by the IID under AC illumination. In particular, we reveal the ambiguity in the blind LSS via ma- trix factorization and the ambiguity in the IID assuming the diffuse reflection model, and then show why and how those ambiguities can be resolved via a physics-based approach. We experimentally confirmed that our method can recover the colors of the light sources, the diffuse reflectance values, and the diffuse and specular intensities (shadings) under each of the light sources, and that the IID under AC illumi- nation is effective for application to auto white balancing.	1. Introduction Artificial light sources in our surroundings are often powered by an electric grid, and therefore their intensities rapidly oscillate in response to the grid’s alternating current (AC). Such intensity oscillations cause flickers in the radi- ance values of a scene illuminated by artificial light sources. The flickers are usually too fast to notice with our naked eyes, but can be captured by using cameras with short ex- posure time settings [32]. It is known that the flickers could make auto white balance unnatural [15]. Interestingly, the flickers are useful for extracting rich in- formation on a scene of interest. Sheinin et al. [28] propose a method for light source separation (LSS) under AC illu- mination. Their method decomposes an image sequence of a scene illuminated by multiple AC light sources into thebasis images of the scene, each of which is illuminated by only one of the light sources, and the temporal intensity pro- files of the light sources. They make use of their self-built coded-exposure camera synchronized to AC and the dataset of temporal intensity profiles of various light sources, and then achieve LSS even for dark scenes such as a city-scale scene at night. In this paper, we show that the flickers due to AC illu- mination is useful also for intrinsic image decomposition (IID). Originally, IID recovers the shading and reflectance images of a scene of interest from a single input image on the basis of the Retinex theory [2, 19]. Those intrinsic prop- erties of a scene is useful for computer vision applications such as image segmentation [6], object recognition [25], and shape from shading [14]. Our proposed method assumes a scene illuminated by multiple AC light sources, and recovers the intrinsic prop- erties of the scene and the light sources from an image se- quence captured by using a consumer high speed camera. In contrast to the conventional methods for IID, our method assumes the dichromatic reflection model [27], and then re- covers the intrinsic properties more than the reflectance and shading images: the colors of the light sources, the diffuse reflectance values, and the diffuse and specular intensities (shadings) under each of the light sources. Specifically, our proposed method conducts the blind LSS via matrix factorization followed by the IID assuming the dichromatic reflection model. In particular, we reveal the ambiguity in the blind LSS under AC illumination via matrix factorization [26], and then resolve the ambiguity by integrating the LSS and the IID assuming the diffuse reflec- tion model. Furthermore, we reveal the ambiguity in the IID assuming the diffuse reflection model under AC illumi- nation, and then resolve the ambiguity on the basis of the dichromatic reflection model by taking specular highlights into consideration1. 1It is analogous to uncalibrated photometric stereo; the GBR ambigu- ity [3] can be resolved from specularity [8] 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. 5735 To show the effectiveness of our proposed method, we conducted a number of experiments using both synthetic and real images. We confirmed that our method works well, i.e.can resolve the ambiguities in the LSS and the IID on real images as well as synthetic images. In addition, we show that the IID under AC illumination is effective for ap- plication to auto white balancing. The main contributions of this study are threefold. First, we tackle a novel problem of the IID under AC illumination. We conduct the blind LSS via matrix factorization followed by the IID assuming the dichromatic reflection model, and show that the flickers due to AC illumination are useful not only for LSS but also for IID. Second, we reveal the am- biguity in the blind LSS via matrix factorization and the ambiguity in the IID assuming the diffuse reflection model. Then, we show why and how those ambiguities can be re- solved via a physics-based approach. Third, we experimen- tally confirmed that our method can recover the colors of the light sources, the diffuse reflectance values, and the diffuse and specular intensities (shadings) under each of the light sources, and that the IID under AC illumination is effective for application to auto white balancing.
Xu_Open-Vocabulary_Panoptic_Segmentation_With_Text-to-Image_Diffusion_Models_CVPR_2023	Abstract We present ODISE: Open-vocabulary DIffusion-based panoptic SEgmentation, which unifies pre-trained text- image diffusion and discriminative models to perform open- vocabulary panoptic segmentation. Text-to-image diffu- sion models have the remarkable ability to generate high- quality images with diverse open-vocabulary language de- scriptions. This demonstrates that their internal represen- tation space is highly correlated with open concepts in the real world. Text-image discriminative models like CLIP , on the other hand, are good at classifying images into open- vocabulary labels. We leverage the frozen internal repre- sentations of both these models to perform panoptic seg- mentation of any category in the wild. Our approach out- performs the previous state of the art by significant margins on both open-vocabulary panoptic and semantic segmen- tation tasks. In particular, with COCO training only, our method achieves 23.4 PQ and 30.0 mIoU on the ADE20K dataset, with 8.3 PQ and 7.9 mIoU absolute improvement over the previous state of the art. We open-source our code and models at https://github.com/NVlabs/ ODISE . *Jiarui Xu was an intern at NVIDIA during the project. †equal contri- bution.	1. Introduction Humans look at the world and can recognize limitless categories. Given the scene presented in Fig. 1, besides identifying every vehicle as a “truck”, we immediately un- derstand that one of them is a pickup truck requiring a trailer to move another truck. To reproduce an intelligence with such a fine-grained and unbounded understanding, the prob- lem of open-vocabulary recognition [36, 57, 76, 89] has re- cently attracted a lot of attention in computer vision. How- ever, very few works are able to provide a unified frame- work that parses all object instances and scene semantics at the same time, i.e., panoptic segmentation. Most current approaches for open-vocabulary recogni- tion rely on the excellent generalization ability of text- image discriminative models [30, 57] trained with Internet- scale data. While such pre-trained models are good at clas- sifying individual object proposals or pixels, they are not necessarily optimal for performing scene-level structural understanding. Indeed, it has been shown that CLIP [57] often confuses the spatial relations between objects [69]. We hypothesize that the lack of spatial and relational under- standing in text-image discriminative models is a bottleneck for open-vocabulary panoptic segmentation. On the other hand, text-to-image generation using dif- fusion models trained on Internet-scale data [1, 59, 61, 62, 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. 2955 90] has recently revolutionized the field of image synthe- sis. It offers unprecedented image quality, generalizabil- ity, composition-ability and, semantic control via the input text. An interesting observation is that to condition the im- age generation process on the provided text, diffusion mod- els compute cross-attention between the text’s embedding and their internal visual representation. This design im- plies the plausibility of the internal representation of dif- fusion models being well-differentiated and correlated to high/mid-level semantic concepts that can be described by language. As a proof-of-concept, in Fig.1 (center), we vi- sualize the results of clustering a diffusion model’s internal features for the image on the left. While not perfect, the discovered groups are indeed semantically distinct and lo- calized. Motivated by this finding, we ask the question of whether Internet-scale text-to-image diffusion models can be exploited to create universal open-vocabulary panoptic segmentation learner for any concept in the wild? To this end, we propose ODISE : Open-vocabulary DIffusion-based panoptic SEgmentation (pronounced o-di- see), a model that leverages both large-scale text-image dif- fusion and discriminative models to perform state-of-the- art panoptic segmentation of any category in the wild. An overview of our approach is illustrated in Fig. 2. At a high-level it contains a pre-trained frozen text-to-image dif- fusion model into which we input an image and its cap- tion and extract the diffusion model’s internal features for them. With these features as input, our mask generator pro- duces panoptic masks of all possible concepts in the image. We train the mask generator with annotated masks avail- able from a training set. A mask classification module then categorizes each mask into one of many open-vocabulary categories by associating each predicted mask’s diffusion features with text embeddings of several object category names. We train this classification module with either mask category labels or image-level captions from the training dataset. Once trained, we perform open-vocabulary panop- tic inference with both the text-image diffusion and discrim- inative models to classify a predicted mask. On many differ- ent benchmark datasets and across several open-vocabulary recognition tasks, ODISE achieves state-of-the-art accuracy outperforming the existing baselines by large margins. Our contributions are the following: • To the best of our knowledge, ODISE is the first work to explore large-scale text-to-image diffusion models for open-vocabulary segmentation tasks. • We propose a novel pipeline to effectively leverage both text-image diffusion and discriminative models to perform open-vocabulary panoptic segmentation. • We significantly advance the field forward by out- performing all existing baselines on many open- vocabulary recognition tasks, and thus establish a new state of the art in this space.
Zhang_A_Loopback_Network_for_Explainable_Microvascular_Invasion_Classification_CVPR_2023	Abstract Microvascular invasion (MVI) is a critical factor for prog- nosis evaluation and cancer treatment. The current diagnosis of MVI relies on pathologists to manually find out cancer- ous cells from hundreds of blood vessels, which is time- consuming, tedious, and subjective. Recently, deep learning has achieved promising results in medical image analysis tasks. However, the unexplainability of black box models and the requirement of massive annotated samples limit the clini- cal application of deep learning based diagnostic methods. In this paper, aiming to develop an accurate, objective, and explainable diagnosis tool for MVI, we propose a Loop- back Network (LoopNet) for classifying MVI efficiently. With the image-level category annotations of the collected Patho- logic Vessel Image Dataset (PVID), LoopNet is devised to be composed binary classification branch and cell locating branch. The latter is devised to locate the area of cancer- ous cells, regular non-cancerous cells, and background. For healthy samples, the pseudo masks of cells supervise the cell locating branch to distinguish the area of regular non- cancerous cells and background. For each MVI sample, the cell locating branch predicts the mask of cancerous cells. Then the masked cancerous and non-cancerous areas of the same sample are input back to the binary classification branch separately. The loopback between two branches en- ables the category label to supervise the cell locating branch to learn the locating ability for cancerous areas. Experiment results show that the proposed LoopNet achieves 97.5%ac- curacy on MVI classification. Surprisingly, the proposed loopback mechanism not only enables LoopNet to predict the cancerous area but also facilitates the classification back- bone to achieve better classification performance. *Corresponding author vessels healthy/MVI vessels 117140×273140 px (a) (b) (c) healthy/cancerous cellsFigure 1. Examples of MVI and healthy vessels extracted from a pathological image of liver cancer. (a) The super large sample contains numerous blood vessels of varied sizes. (b) The healthy vessels are composed of a variety of cells with similar appearances. (c) The cancerous cells have varied types and similar appearances to parts of healthy cells.	1. Introduction Microvascular invasion (MVI), referring to the appear- ance of cancerous cells within microscopic venules or veins, is a histological feature of cancer-related to aggressive biological behavior [27, 56]. In clinical, MVI is usually used as a reference standard for assessing cancer spread- ing, which is a critical factor for prognosis evaluation and treatment [8, 15, 43]. Accurate prognosis evaluation along with appropriate treatment can effectively improve patient’s life quality and prolong their life-span. Currently, the diagnosis of MVI relies on pathologists to manually find out cancerous cells from hundreds of blood vessels, each of which usually contains dozens of cells. As shown in Fig.1, each pathological sample is an image of 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. 7443 about 100,000×250,000px. These super-large pathological images have three characteristics. Firstly, each sample con- tains numerous blood vessels (Fig.1a). Secondly, each blood vessel usually has a variety of cells with similar appearances (Fig.1b). Thirdly, types of cancerous cells are also varied (Fig.1c). Therefore, diagnosis of MVI requires the profes- sional pathologist to discriminate cancerous/non-cancerous cells carefully, which is time-consuming and tedious. The discrimination relies on the individual pathologist’s prior knowledge, which is subjective and leads to misdiagnosis occasionally. In recent years, deep learning has achieved promising results in many areas [28 –30, 68 –72], including medical image analysis. Many researchers focus on applying deep learning techniques to image-based tumor analysis tasks, such as tumor grading [3, 73], lesion area detection [14, 35], vessel segmentation [18,31], cell detection/segmentation [42, 54, 67, 75], etc. The successful application of deep learning relies on massive annotated samples. However, annotating cancerous cells of all MVI images is very time-consuming. What’s more, the black-box characteristic of deep learning leads to unexplainable classification results, which limits the clinical application of deep learning based diagnostic methods. In order to apply the deep learning technique to the MVI analysis task, we collect the first Pathologic Vessel Image Dataset (PVID) containing healthy blood vessel samples and MVI samples from the pathological image of liver cancer patients. In this paper, we aim to develop an accurate, objective, and explainable method for MVI diagnosis with as few anno- tations as possible. As annotating the cell in each MVI vessel is time-consuming, we only adopt easily-obtained image- level category labels for developing the new approach. For the explainable MVI classification, the developed ap- proach should provide credible evidence, such as cancerous areas and classification results. Therefore, the proposed ap- proach is devised to be composed of two branches: the binary classification branch and the cell locating branch. The binary classification branch is used to classify the healthy blood ves- sels and MVI vessels with corresponding vessel image-level category labels as supervision. The initial goal of the cell locating branch is to distinguish the cancerous cells. How- ever, the supervision information for the cell locating branch is insufficient, which requires exploring more supervision information from the characteristic of MVI itself. Firstly, based on the characteristic of blood vessel sam- ples that most cells can be distributed into some similar templates according to structure and color, the correlation filter [9, 22], which is widely adopted in the object tracking area, can be used for locating most of the cells; hence the results of this filter can be interpreted as pseudo masks of cells for supervising the cell locating branch to distinguishcell area from the background. Secondly, the healthy vessel sample only containing non-cancerous cells and background is used for supervising the cell locating branch distinguishing healthy area (non-cancerous cells and background) from the cancerous cells. Lastly, we devise loopback strategy between the binary classification branch and cell locating branch to discover the cancerous area from each MVI sample. For the loopback strategy, the cell locating branch first predicts the cancerous area of the MVI sample, then the can- cerous and non-cancerous areas of the same sample masked with the predicted results are input back into the classifi- cation branch separately. The devised a loopback strategy effectively achieves two goals: 1) utilizing the image-level category label to supervise the cell locating branch distin- guishing the cancerous area from other areas. 2) building the direct relation between the predicted cancerous areas and the final classification result. Experiment results show that the loopback strategy not only enables the proposed framework to predict precious cancerous areas but also facilitate the classification branch achieve better classification performance. The two-branch framework with the loopback strategy, termed as Loopback Network (LoopNet), achieves 97.5%accuracy on MVI clas- sification. In conclusion, the main contributions of our work are summarized as follows: •We propose the first deep learning based network, termed as LoopNet, for explainable MVI classifica- tion. LoopNet fully exploits the characteristics of MVI samples to achieve blood vessel classification and cell locating results simultaneously and can be extended to MVI analysis tasks on various organs. •The loopback strategy is devised for utilizing the cat- egory label to supervise LoopNet distinguishing the cancerous area from other regions, which effectively builds the direct relation between the located cancerous area and the final classification result. •We collect the first Pathologic Vessel Image Dataset (PVID) containing 4130 healthy blood vessel samples and857MVI samples from the pathological image of 103liver cancer patients. •Experiment show that LoopNet achieves 97.5%accu- racy on PVID, which verifies the potential of deep learn- ing on MVI classification task.
Yang_VectorFloorSeg_Two-Stream_Graph_Attention_Network_for_Vectorized_Roughcast_Floorplan_Segmentation_CVPR_2023	Abstract Vector graphics (VG) are ubiquitous in industrial de- signs. In this paper, we address semantic segmentation of a typical VG, i.e., roughcast floorplans with bare wall structures, whose output can be directly used for further applications like interior furnishing and room space mod- eling. Previous semantic segmentation works mostly pro- cess well-decorated floorplans in raster images and usu- ally yield aliased boundaries and outlier fragments in seg- mented rooms, due to pixel-level segmentation that ignores the regular elements (e.g. line segments) in vector floor- plans. To overcome these issues, we propose to fully uti- lize the regular elements in vector floorplans for more in- tegral segmentation. Our pipeline predicts room segmen- tation from vector floorplans by dually classifying line seg- ments as room boundaries, and regions partitioned by line segments as room segments. To fully exploit the structural relationships between lines and regions, we use two-stream graph neural networks to process the line segments and par- titioned regions respectively, and devise a novel modulated graph attention layer to fuse the heterogeneous information from one stream to the other. Extensive experiments show that by directly operating on vector floorplans, we outper- form image-based methods in both mIoU and mAcc. In ad- dition, we propose a new metric that captures room integrity and boundary regularity, which confirms that our method produces much more regular segmentations. Source code is available at https://github.com/DrZiji/VecFloorSeg.	1. Introduction Vector graphics are widely used in industrial designs, in- cluding graphic designs [26], 2D interfaces [5] and floor- plans [15]. In particular, 2D floorplans consisting of ge- ometric primitives (e.g., lines and curves) are the de-facto §Haiyong is the Project Lead. ∗Corresponding Author. (b)Furnished Counterpart (a)Input (d)Label (e)Ours (f)Image Based Segmentation(b)Furnished Counterpart (c)3D Rendering Figure 1. Comparing the results of our vector graphics based method (e) and raster image-based method [39] (f). Our result has straight boundaries and consistent region labels, compared with image-based result where red squares highlight semantic confu- sion and the green square underscores missing room prediction. data representation for interior designs, indoor construction and property development. In contrast to raster images with fixed resolutions, vector graphics can be arbitrarily scaled without artifacts such as blurring and aliasing details. On the other hand, due to the irregularly structured data, it is difficult to apply image-based backbone networks directly to vector graphics for various applications. Semantic segmentation of roughcast floorplans into rooms with labeled types (e.g. bedroom, kitchen, etc.) is a fundamental task for downstream applications. Interior de- signers usually first draw the roughcast floorplan, including basic elements like wall blocks and pipe barrels for property development (Fig. 1(a)) [29, 34]. Afterwards, interior fur- nishing, furniture layout, and 3D room spaces can be con- structed and customized (Fig. 1(b)&(c)) [33]. During this procedure, it is important to obtain semantic segmentation of room spaces to cater to above needs. While recogniz- ing room layouts from wall structures is straightforward for humans, automatic recognition with accurate semantics and clean boundaries is challenging. 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. 1358 Recent works [19, 22, 23, 39] use powerful image-based segmentation networks on rasterized floorplans to predict room segmentation in a pixel-wise manner. Due to the pixel-wise processing that ignores the integrity of structural elements, their results tend to have jigsaw boundaries and fragmented semantic regions as shown in Fig. 1(f). Besides, these methods usually rely on texts and furniture to deter- mine the semantic labels, which are not available in rough- cast floorplans. Another line of prior works processes vec- tor graphics for recognition, e.g., object detection [14, 28] and symbol spotting [9,10,41]. However, to our best knowl- edge, semantic segmentation of vector graphics, roughcast floorplans in particular, has not been investigated before. In this work, we make a first attempt at semantic seg- mentation of 2D roughcast floorplans directly as vector graphics. On one hand, by working with vector floor- plans directly, the segmentation output is naturally regular and compact vector graphics rather than dense pixels (cf. Fig. 1(e)&(f)), which greatly facilitates downstream appli- cations. On the other hand, the vector roughcast floorplans pose challenges in the following aspects. First, rooms in vector floorplans seldomly contain complete contour lines formed by the input line segments (see Fig. 1(a)&(d)). Sec- ond, the type of a room is determined not only by its shape but also by the relative relationships with its neighboring rooms and within the overall floorplan. To address the above challenges, we make two obser- vations. First, room spaces can be subdivided into a set of polygonal regions by input lines together with their ex- tensions (Fig. 2), and their semantic classification as room types defines room segmentation. Second, lines (including extended lines) are potential boundaries of different rooms, and their being classified as boundaries or not should assist room segmentation in a dual direction. Based on the two observations, we design a two-stream graph attention network (GAT) for the task. As illustrated in Fig. 2, the primal stream takes as input the primal graph that encodes line endpoints as vertices and line segments as edges, and predicts the boundary classification of edges; the dual stream takes as input the dual graph that encodes parti- tioned regions as vertices and their adjacency as edges, and predicts the vertex classification of regions, which effec- tively defines the semantic segmentation of a vector floor- plan. Furthermore, the two streams should enhance each other rather than being separated. To facilitate data ex- change between two streams, we present a novel modulated GAT layer to fuse information from one stream into the graph network computation of the other stream. We evalu- ate our approach on two large-scale floorplan datasets; both classical metrics and a new metric that we develop to focus on integral segmentation show that our results improve pre- vious image-based results significantly. To summarize, we make the following contributions:•We approach semantic segmentation of vector rough- cast floorplans through the dual aspects of boundary line classification and region classification. •We design two-stream graph neural networks to pro- cess dual regions and primal lines respectively, and devise a novel modulated GAT layer to exchange data across streams. •We propose a new metric to capture both accuracy and integrity of the segmentation results. •We obtain vector segmentation results on two floorplan datasets, which show much more compact boundaries and better integrity than raster image-based results.
Zhang_NICO_Towards_Better_Benchmarking_for_Domain_Generalization_CVPR_2023	Abstract Despite the remarkable performance that modern deep neural networks have achieved on independent and iden- tically distributed (I.I.D.) data, they can crash under dis- tribution shifts. Most current evaluation methods for do- main generalization (DG) adopt the leave-one-out strat- egy as a compromise on the limited number of domains. We propose a large-scale benchmark with extensive labeled domains named NICO++along with more rational eval- uation methods for comprehensively evaluating DG algo- rithms. To evaluate DG datasets, we propose two metrics to quantify covariate shift and concept shift, respectively. Two novel generalization bounds from the perspective of data construction are proposed to prove that limited con- cept shift and significant covariate shift favor the evalua- tion capability for generalization. Through extensive ex- periments, NICO++shows its superior evaluation capabil- ity compared with current DG datasets and its contribu- tion in alleviating unfairness caused by the leak of oracle knowledge in model selection. The data and code for the benchmark based on NICO++are available at https: //github.com/xxgege/NICO-plus .	1. Introduction Machine learning has illustrated its excellent capability in a wide range of areas [37, 65, 82]. Most current algo- rithms minimize the empirical risk in training data relying on the assumption that training and test data are indepen- dent and identically distributed (I.I.D.). However, this ideal hypothesis is hardly satisfied in real applications, especially those high-stake applications such as healthcare [10, 49], autonomous driving [1, 13, 39] and security systems [6], owing to the limitation of data collection and intricacy of the scenarios. Distribution shifts between training and test data may lead to the unreliable performance of current approaches in practice. Hence, instead of generalization †Equal contribution *Corresponding Author 0.1 0.3 0.5 0.7 Concept shift0.150.250.35Covariate shiftPACS VLCS DomainNet Oﬃce-Home iWildCam (WILDS) FMoW (WILDS) Meta-shift NICO NICO++Figure 1. Covariate shift ( Mcovin Equation (1)) and concept shift (Mmax cpt in Equation (2)) of NICO++and current DG datasets. NICO++has the lowest concept shift and highest covariate shift, showing the superiority in evaluation capability. within the training distribution, the ability to generalize un- der distribution shift, domain generalization (DG) [75, 94], is of more critical significance in realistic scenarios. In the field of computer vision, benchmarks that pro- vide the common ground for competing approaches often play a role of catalyzer promoting the advance of research [14]. An advanced DG benchmark should provide sufficient diversity in distributions for both training and evaluating DG algorithms [74, 78] while ensuring essential common knowledge of categories for inductive inference across do- mains [33, 34, 93]. The first property drives generalization challenging, and the second ensures the solvability [81]. This requires adequate distinct domains and instructive fea- tures for each category shared among all domains. Current DG benchmarks, however, either lack sufficient domains (e.g., 4 domains in PACS [40], VLCS [18] and Office-Home [73] and 6 in DomainNet [53]) or too simple or limited to simulating significant distribution shifts in real scenarios [2, 21, 30]. To enrich the diversity and perplexing distribution shifts in training data as much as possible, most of the current evaluation methods for DG adopt the leave- one-out strategy, where one domain is considered as the test domain and the others for training. This is not an ideal eval- uation for generalization but a compromise due to the lim- ited number of domains in current datasets, which impairs 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. 16036 the evaluation capability. To address this issue, we suggest testing DG methods on multiple test domains instead of one specific domain in each evaluation after training . To benchmark DG methods comprehensively and sim- ulate real scenarios where a trained model may encounter any possible test data while providing sufficient diversity in the training data, we construct a large-scale DG dataset named NICO++with extensive domains and two proto- cols supported by aligned and flexible domains across cate- gories, respectively, for better evaluation. Our dataset con- sists of 80 categories, 10 aligned common domains for all categories, 10 unique domains specifically for each cat- egory, and more than 230,000 images. Abundant diver- sity in both domain and category supports flexible assign- ments for training and test, controllable degree of distribu- tion shifts, and extensive evaluation on multiple target do- mains. Images collected from real-world photos and consis- tency within category concepts provide sufficient common knowledge for recognition across domains on NICO++. To evaluate DG datasets in-depth, we investigate dis- tribution shifts on images (covariate shift) and common knowledge for category discrimination across domains (concept agreement) within them. Formally, we present quantification for covariate shift and the opposite of concept agreement, namely concept shift, via two novel metrics. We propose two novel generalization bounds and analyze them from the perspective of data construction instead of models. Through these bounds, we prove that limited concept shift and significant covariate shift favor the evaluation capabil- ity for generalization. Moreover, a critical yet common problem in DG is the model selection and the potential unfairness in the compar- ison caused by leveraging the knowledge of target data to choose hyperparameters that favors test performance [3,27]. This issue is exacerbated by the notable variance of test per- formance with various algorithm irrelevant hyperparame- ters on current DG datasets. Intuitively, strong and unsta- ble concept shift such as confusing mapping relations from images to labels across domains embarrasses training con- vergence and enlarges the variance. We conduct extensive experiments on three levels. First, we evaluate NICO++and current DG datasets with the pro- posed metrics and show the superiority of NICO++in eval- uation capability, as shown in Figure 1. Second, we con- duct copious experiments on NICO++to benchmark cur- rent representative methods with the proposed protocols. Results show that the room for improvement of generaliza- tion methods on NICO++is spacious. Third, we show that NICO++helps alleviate the issue by squeezing the possible improvement space of oracle leaking and contributes as a fairer benchmark to the evaluation of DG methods, which meets the proposed metrics.
Yadav_Habitat-Matterport_3D_Semantics_Dataset_CVPR_2023	Abstract We present the Habitat-Matterport 3D Semantics (HM3DS EM) dataset. HM3DS EMis the largest dataset of 3D real-world spaces with densely annotated seman- tics that is currently available to the academic commu- nity. It consists of 142,646 object instance annotations across 216 3D spaces and 3,100 rooms within those spaces. The scale, quality, and diversity of object annotations far exceed those of prior datasets. A key difference setting apart HM3DS EMfrom other datasets is the use of tex- ture information to annotate pixel-accurate object bound- aries. We demonstrate the effectiveness of HM3DS EM dataset for the Object Goal Navigation task using differ- ent methods. Policies trained using HM3DS EMperform outperform those trained on prior datasets. Introduction ofHM3DS EMin the Habitat ObjectNav Challenge lead to an increase in participation from 400 submissions in 2021 to 1022 submissions in 2022. Project page: https: //aihabitat.org/datasets/hm3d-semantics/	1. Introduction Over the recent past, work on acquiring and semantically annotating datasets of real-world spaces has significantly accelerated research into embodied AI agents that can per- ceive, navigate and interact with realistic indoor scenes [ 1–5]. However, the acquisition of such datasets at scale is a labori- ous process. HM3D [ 5] which is one of the largest available datasets with 1000 high-quality and complete indoor space reconstructions, reportedly required 800+ hours of human effort to carry out mainly data curation and verification of 3D reconstructions. Moreover, dense semantic annotation of such acquired spaces remains incredibly challenging. We present the Habitat-Matterport 3D Dataset Seman- tics ( HM3DS EM). This dataset provides a dense semantic annotation ‘layer’ augmenting the spaces from the original *Equal Contribution, Correspondence: [email protected] †Equal ContributionHM3D dataset. This semantic ‘layer’ is implemented as a set of textures that encode object instance semantics and cluster objects into distinct rooms. The semantics include architectural elements (walls, floors, ceilings), large objects (furniture, appliances etc.), as well as ‘stuff’ categories (ag- gregations of smaller items such as books on bookcases). This semantic instance information is specified in the seman- tic texture layer, providing pixel-accurate correspondences to the original acquired RGB surface texture and underlying geometry of the objects. TheHM3DS EMdataset currently contains annotations for142,646object instances distributed across 216spaces and3,100rooms within those spaces. Figure 1 shows some examples of the semantic annotations from the HM3DS EM dataset. The achieved scale is larger than prior work (2.8x rel- ative to Matterport3D [ 6] (MP3D) and 2.1x relative to ARK- itScenes [ 7] in terms of total number of object instances). We demonstrate the usefulness of HM3DS EMon the Ob- jectGoal navigation task. Training on HM3DS EMresults in higher cross-dataset generalization performance. Surpris- ingly, the policies trained on HM3DS EMperform better on average across scene datasets compared to training on the datasets themselves. We also show that increasing the size of training datasets improve the navigation performance. These results highlight the importance of improving the quality and scale of 3D datasets with dense semantic annotations for improving downstream embodied AI task performance.
Xu_Visual-Tactile_Sensing_for_In-Hand_Object_Reconstruction_CVPR_2023	Abstract Tactile sensing is one of the modalities humans rely on heavily to perceive the world. Working with vision, this modality refines local geometry structure, measures defor- mation at the contact area, and indicates the hand-object contact state. With the availability of open-source tactile sensors such as DIGIT, research on visual-tactile learning is becoming more accessible and reproducible. Leverag- ing this tactile sensor, we propose a novel visual-tactile in-hand object reconstruction framework VTacO , and ex- tend it to VTacOH for hand-object reconstruction. Since our method can support both rigid and deformable ob- ject reconstruction, no existing benchmarks are proper for the goal. We propose a simulation environment, VT-Sim, which supports generating hand-object interaction for both rigid and deformable objects. With VT-Sim, we gener- ate a large-scale training dataset and evaluate our method on it. Extensive experiments demonstrate that our pro- posed method can outperform the previous baseline meth- ods qualitatively and quantitatively. Finally, we directly ap- ply our model trained in simulation to various real-worldtest cases, which display qualitative results. Codes, mod- els, simulation environment, and datasets are available at https://sites.google.com/view/vtaco/ .	1. Introduction Human beings have a sense of object geometry by seeing and touching, especially when the object is in manipulation and undergoes a large portion of occlusion, where visual information is not enough for the details of object geom- etry. In such cases, vision-based tactile sensing is a good supplement as a way of proximal perception. In the past, few vision-based tactile sensors were commercially avail- able or open-source, so the visual-tactile sensing techniques could not be widely studied. Previous works [27, 34] on in-hand object reconstruction either studied rigid objects or were limited to simple objects with simple deformation. * indicates equal contributions. § Cewu Lu is the corresponding author, the member of Qing Yuan Research Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China and Shanghai Qi Zhi institute. 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. 8803 Vision-based tactile sensors [6, 15, 30, 33] can produce colorful tactile images indicating local geometry and defor- mation in the contact areas. In this work, we mainly work with DIGIT [15] as it is open-source for manufacture and is easier to reproduce sensing modality. With tactile images, we propose a novel Visual- Tactile in-hand Object recon- struction framework named VTacO . VTacO reconstructs the object geometry with the input of a partial point cloud observation and several tactile images. The tactile and ob- ject features are extracted by neural networks and fused in the Winding Number Field (WNF) [13], and the object shape is extracted by Marching Cubes algorithm [17]. WNF can represent the object shape with open and thin structures. The poses of tactile sensors can be determined either by markers attached or by hand kinematics. By default, VTacO assumes the tactile sensor poses can be obtained indepen- dently, but we also discuss how to obtain the tactile sensor poses alongside the object with hand pose estimation. The corresponding method is named VTacOH . With tactile information, we can enhance pure visual in- formation from three aspects: (1) Local geometry refine- ment . We use tactile sensing as proximal perception to complement details of local geometry. (2) Deformation at contact area . Objects, even those we consider rigid, can undergo considerable deformation given external forces ex- erted by hand. (3) Hand-object contact state . Tactile sen- sors indicate whether the hand is in contact with the object’s surface. To demonstrate such merits, we conduct the object reconstruction tasks in both rigid and non-rigid settings. Since obtaining the ground truth of object deformation in the real world is hard, we first synthesize the training data from a simulator. DIGIT has an official simulation im- plementation, TACTO [29]. However, it is based on py- bullet [5], which has limited ability to simulate deformable objects. Thus, we implement a tactile simulation environ- ment VT-Sim in Unity. In VT-Sim, we generate hand poses with GraspIt! [18], and simulate the deformation around the contact area with an XPBD-based method. In the sim- ulation, we can easily obtain depth image, tactile image, DIGIT pose, and object WNF as training samples for both rigid and non-rigid objects. To evaluate the method, we compare the proposed visual-tactile models with its visual-only setting, and the previous baseline 3D Shape Reconstruction from Vision and Touch (3DVT) [23]. Extensive experiments show that our method can achieve both quantitative and qualitative improvements on baseline methods. Besides, since we make the tactile features fused with winding number predic- tion, we can procedurally gain finer geometry reconstruc- tion results by incrementally contacting different areas of objects. It can be useful for robotics applications [8, 22]. Then, we directly apply the model trained with synthesis data to the real world. It shows great generalization ability.We summarize our contributions as follows: • A visual-tactile learning framework to reconstruct an object when it is being manipulated. We provide the object-only version VTacO, and the hand-object ver- sion VTacOH. • A simulation environment, VT-Sim, which can gener- ate training samples. We also validate the generaliza- tion ability of the models trained on the simulated data to the real-world data.
Yuan_Robust_Test-Time_Adaptation_in_Dynamic_Scenarios_CVPR_2023	Abstract Test-time adaptation (TTA) intends to adapt the pre- trained model to test distributions with only unlabeled test data streams. Most of the previous TTA methods have achieved great success on simple test data streams such as independently sampled data from single or multiple distri- butions. However, these attempts may fail in dynamic sce- narios of real-world applications like autonomous driving, where the environments gradually change and the test data is sampled correlatively over time. In this work, we ex- plore such practical test data streams to deploy the model on the fly, namely practical test-time adaptation (PTTA). To do so, we elaborate a Robust Test-Time Adaptation (RoTTA) method against the complex data stream in PTTA. More specifically, we present a robust batch normalization scheme to estimate the normalization statistics. Meanwhile, a memory bank is utilized to sample category-balanced data with consideration of timeliness and uncertainty. Further, to stabilize the training procedure, we develop a time-aware reweighting strategy with a teacher-student model. Exten- sive experiments prove that RoTTA enables continual test- time adaptation on the correlatively sampled data streams. Our method is easy to implement, making it a good choice for rapid deployment. The code is publicly available at https://github.com/BIT-DA/RoTTA	1. Introduction In recent years, many machine learning problems have made considerable headway with the success of deep neu- ral networks [13, 22, 33, 38]. Unfortunately, the perfor- mance of deep models drops significantly when training data and testing data come from different distributions [59], which limits their utility in real-world applications. To re- duce the distribution shift, a handful of works focus on transfer learning field [56], in particular, domain adapta- tion (DA) [17, 42, 45, 48, 69, 72] or domain generalization (DG) [40,41,52,71,83], in which one or more different but Corresponding author Test data streamContinual TTANon-i.i.d.TTAPractical TTACategoryDistributionFully TTA Correlation samplingDistributionchangingFigure 1. We consider the practical test-time adaptation (TTA) setup and compare it with related ones. First, Fully TTA [70] adapts models on a fixed test distribution with an independently sampled test stream. Then, on this basis, Continual TTA [73] takes the continually changing distributions into consideration. Next, Non-i.i.d. TTA [19] tries to tackle the correlatively sampled test streams on a single test distribution, where the label distribution among a batch of data deviates from that of the test distribution. To be more practical, Practical TTA strives to connect both worlds: distribution changing and correlation sampling. related labeled datasets (a.k.a. source domain) are collected to help the model generalize well to unlabeled or unseen samples in new datasets (a.k.a. target domain). While both DA and DG have extensively studied the problem of distribution shifts, they typically assume acces- sibility to the raw source data. However, in many practical scenarios like personal consumption records, the raw data should not be publicly available due to data protection reg- ulations. Further, existing methods have to perform heavy backward computation, resulting in unbearable training costs. Test-time adaptation (TTA) [3,11,16,24,26,54,65,81] attempts to address the distribution shift online at test time with only unlabeled test data streams. Unequivocally, TTA has drawn widespread attention in a variety of applications, e.g., 2D/3D visual recognition [2, 29, 49, 65, 82], multi- modality [63, 64] and document understanding [15]. Prior TTA studies [7, 20, 70, 73] mostly concentrate on a simple adaptation scenario, where test samples are inde- pendently sampled from a fixed target domain. To name a few, Sun et al. [65] adapt to online test samples drawn from a constant or smoothly changing distribution with an auxil- iary self-supervised task. Wang et al. [70] adapt to a fixed 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. 15922 Table 1. Comparison between our proposed practical test-time adaptation (PTTA) and related adaptation settings. SettingAdaptation Stage Available Data Test Data Stream Train Test Source Target Distribution Sampling Protocol Domain Adaptation ! % ! ! - - Domain Generalization ! % ! % - - Test-Time Training [65] ! ! ! ! stationary independently Fully Test-Time Adaptation [70] % ! % ! stationary independently Continual Test-Time Adaptation [73] % ! % ! continually changing independently Non-i.i.d. Test-Time Adaptation [5, 19] % ! % ! stationary correlatively Practical Test-Time Adaptation (Ours) % ! % ! continually changing correlatively target distribution by performing entropy minimization on- line. However, such an assumption is violated when the test environments change frequently [73]. Later on, Boudiaf et al. [5] and Gong et al. [19] consider the temporal correlation ship within test samples. For example, in autonomous driv- ing, test samples are highly correlated over time as the car will follow more vehicles on the highway or will encounter more pedestrians in the streets. More realistically, the data distribution changes as the surrounding environment alerts in weather, location, or other factors. In a word, distribution change and data correlation occur simultaneously in reality. Confronting continually changing distributions, tradi- tional algorithms like pseudo labeling or entropy minimiza- tion become more unreliable as the error gradients cumu- late. Moreover, the high correlation among test samples re- sults in the erroneous estimation of statistics for batch nor- malization and collapse of the model. Driven by this analy- sis, adapting to such data streams will encounter two major obstacles: 1) incorrect estimation in the batch normaliza- tion statistics leads to erroneous predictions of test samples, consequently resulting in invalid adaptation; 2) the model will easily or quickly overfit to the distribution caused by the correlative sampling. Thus, such dynamic scenarios are pressing for a new TTA paradigm to realize robust adapta- tion. In this work, we launch a more realistic TTA setting, where distribution changing and correlative sampling oc- cur simultaneously at the test phase. We call this Practical Test-Time Adaptation , or briefly, PTTA . To understand more clearly the similarities and differences between PTTA and the previous setups, we visualize them in Figure 1 and sum- marize them in Table 1. To conquer this challenging prob- lem, we propose a Robust Test-TimeAdaptation ( RoTTA ) method, which consists of three parts: 1) robust statistics es- timation, 2) category-balanced sampling considering time- liness and uncertainty and 3) time-aware robust training. More concretely, we first replace the erroneous statistics of the current batch with global ones maintained by the expo- nential moving average. It is a more stable manner to esti- mate the statistics in BatchNorm layers. Then, we simulate a batch of independent-like data in memory with category- balanced sampling while considering the timeliness and un- certainty of the buffered samples. That is, samples that arenewer and less uncertain are kept in memory with higher priority. With this batch of category-balanced, timely and confident samples, we can obtain a snapshot of the current distribution. Finally, we introduce a time-aware reweight- ing strategy that considers the timeliness of the samples in the memory bank, with a teacher-student model to perform robust adaptation. With extensive experiments, we demon- strate that RoTTA can robustly adapt in the practical setup, i.e., PTTA. In a nutshell, our contributions can be summarized as: • We propose a new test-time adaptation setup that is more suitable for real-world applications, namely practical test-time adaptation (PTTA). PTTA considers both distribution changing and correlation sampling. • We benchmark the performance of prior methods in PTTA and uncover that they only consider one aspect of the problem, resulting in ineffective adaptation. • We propose a robust test-time adaptation method (RoTTA), which has a more comprehensive considera- tion of PTTA challenges. Ease of implementation and effectiveness make it a practical deployment option. • We extensively demonstrate the practicality of PTTA and the effectiveness of RoTTA on common TTA benchmarks [23], i.e., CIFAR-10-C and CIFAR-100- C and a large-scale DomainNet [58] dataset. RoTTA obtains state-of-the-art results, outperforming the best baseline by a large margin (reducing the averaged classification error by over 5.9%, 5.5% and 2.2% on CIFAR-10-C, CIFAR-100-C and DomainNet, respec- tively).
Zhang_Frame_Flexible_Network_CVPR_2023	Abstract Existing video recognition algorithms always conduct different training pipelines for inputs with different frame numbers, which requires repetitive training operations and multiplying storage costs. If we evaluate the model us- ing other frames which are not used in training, we ob- serve the performance will drop significantly (see Fig. 1), which is summarized as Temporal Frequency Deviation phenomenon. To fix this issue, we propose a general frame- work, named Frame Flexible Network (FFN), which not only enables the model to be evaluated at different frames to adjust its computation, but also reduces the memory costs of storing multiple models significantly. Concretely, FFN in- tegrates several sets of training sequences, involves Multi- Frequency Alignment (MFAL) to learn temporal frequency invariant representations, and leverages Multi-Frequency Adaptation (MFAD) to further strengthen the representa- tion abilities. Comprehensive empirical validations us- ing various architectures and popular benchmarks solidly demonstrate the effectiveness and generalization of FFN (e.g., 7.08/5.15/2.17 %performance gain at Frame 4/8/16 on Something-Something V1 dataset over Uniformer). Code is available at https://github.com/BeSpontaneous/FFN.	1. Introduction The growing number of online videos boosts the research on video recognition, laying a solid foundation for deep learning which requires massive data. Compared with im- age classification, video recognition methods need a series of frames to represent the video which scales the compu- tation. Thus, the efficiency of video recognition methods has always been an essential factor in evaluating these ap- proaches. One existing direction to explore efficiency is designing lightweight networks [9, 40] which are hardware friendly. Even if they increase the efficiency with an accept- able performance trade-off, these methods cannot make fur- ther customized adjustments to meet the dynamic-changing resource constraint in real scenarios. In community, there 5 25 45 65 85 GFLOPs152025303540455055Acc (%) TSM TSM_ST SlowFast SlowFast_ST Uniformer Uniformer_ST(a) Temporal Frequency Deviation phenomenon exists in various video recognition architectures. 5 35 65 95 125 GFLOPs15202530354045Acc (%) TSM(R18) TSM(R18)_ST TSM(R50) TSM(R50)_ST TSM(R101) TSM(R101)_ST(b) Temporal Frequency Devia- tion phenomenon exists in different depths of deep networks. Figure 1. Temporal Frequency Deviation phenomenon widely exists in video recognition. All methods are trained with high frame number and evaluated at other frames to compare with Sep- arated Training (ST) which individually trains the model at differ- ent frames on Something-Something V1 dataset. are two lines of research being proposed to resolve this is- sue. The first one is to design networks that can execute at various depths [10] or widths [37] to adjust the computa- tion from the model perspective. The other line of research considers modifying the resolutions of input data [15,34] to accommodate the cost from the data aspect. However, these methods are carefully designed for 2D CNNs, which may hinder their applications on video recognition where 3D CNNs and Transformer methods are crucial components. Different from image-related tasks, we need to sample multiple frames to represent the video, and the computa- tional costs will grow proportionally to the number of sam- pled frames. Concretely, standard protocol trains the same network with different frames separately to obtain multi- ple models with different performances and computations. This brings challenges to applying these networks on edge devices as the parameters will be multiplied if we store all models, and downloading and offloading models to switch them will cost non-negligible time. Moreover, the same video may be sampled at various temporal rates on differ- ent platforms, employing a single network that is trained at a certain frame number for inference cannot resist the vari- ance of frame numbers in real scenarios. 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. 10504 Training the model with a high frame number (i.e., high temporal frequency) and directly evaluating it at fewer frames (i.e., low temporal frequency) to adjust the cost is a naive and straightforward solution. To test its effective- ness, we compare it with Separated Training (ST) which trains the model at different temporal frequency individu- ally and tests it with the corresponding frame. We conduct experiments on 2D-network TSM [18], 3D-network Slow- Fast [6] and Transformer-network Uniformer [16], and find obvious performance gaps between the inference results and ST from Fig. 1, which means these methods will exhibit sig- nificantly inferior performance if they are not evaluated at the frame number used in training. Further, we conduct the same experiments on different depths of deep networks and a similar phenomenon appears. We denote this generally existing phenomenon as Temporal Frequency Deviation. The potential reason for Temporal Frequency Devia- tion has been explored in Sec. 3 and briefly summarized as the shift in normalization statistics. To address this is- sue, we propose a general framework, named Frame Flexi- ble Network (FFN), which only requires one-time training, but can be evaluated at multiple frame numbers with great flexibility. We import several input sequences with differ- ent sampled frames to FFN during training and propose Multi-Frequency Alignment (MFAL) to learn the temporal frequency invariant representations for robustness towards frame change. Moreover, we present Multi-Frequency Adaptation (MFAD) to further strengthen the representation abilities of the sub-networks which helps FFN to exhibit strong performance at different frames during inference. Although normalization shifting problem [36, 37] and resolution-adaptive networks [15, 34] have been studied, we stress that designing frame flexible video recognition frameworks to accommodate the costs and save parameters is non-trivial and has practical significance for the follow- ing reasons. First, prior works [15, 34] carefully analyzed the detailed structure of 2D convolutions in order to pri- vatize the weights for different scale images. While our method does not touch the specific design of the spatial- temporal modeling components and shares their weights for inputs with different frames. This procedure not only en- ables our method to be easily applied to various architec- tures (2D/3D/Transformer models), but also enforces FFN to learn temporal frequency invariant representations. Sec- ond, it is, indeed, a common practice to conduct Separated Training (ST) in video recognition, which needs multiply- ing memory costs to store individual models, and the mod- els are hard to resist the variance in temporal frequency which limits their applications in actual practice. While FFN provides a feasible solution to these challenges which significantly reduces the memory costs of storing multiple models and can be evaluated at different frames to adjust the cost with even higher accuracy compared to ST.With the proposed framework, we can resolve Tempo- ral Frequency Deviation and enable these methods to adjust their computation based on the current resource budget by sampling different frames, trimming the storage costs of ST remarkably. Moreover, we provide a naive solution that en- ables FFN to be evaluated at any frame and increases its flexibility during inference. Validation results prove that FFN outperforms ST even at frames that are not used in training. The contributions are summarized as follows: • We reveal the phenomenon of Temporal Frequency Deviation that widely exists in video recognition. It is detailedly analyzed and practically inspires our study. • We propose a general framework Frame Flexible Net- work (FFN) to resolve Temporal Frequency Devia- tion. We design Multi-Frequency Alignment (MFAL) to learn temporal frequency invariant representations and present Multi-Frequency Adaptation (MFAD) to further strengthen the representation abilities. • Comprehensive empirical validations show that FFN, which only requires one-shot training, can adjust its computation by sampling different frames and outper- form Separated Training (ST) at different frames on various architectures and datasets, reducing the mem- ory costs of storing multiple models significantly.
Yang_Revisiting_Weak-to-Strong_Consistency_in_Semi-Supervised_Semantic_Segmentation_CVPR_2023	Abstract In this work, we revisit the weak-to-strong consistency framework, popularized by FixMatch from semi-supervised classification, where the prediction of a weakly perturbed im- age serves as supervision for its strongly perturbed version. Intriguingly, we observe that such a simple pipeline already achieves competitive results against recent advanced works, when transferred to our segmentation scenario. Its success heavily relies on the manual design of strong data augmen- tations, however, which may be limited and inadequate to explore a broader perturbation space. Motivated by this, we propose an auxiliary feature perturbation stream as a supplement, leading to an expanded perturbation space. On the other, to sufficiently probe original image-level augmen- tations, we present a dual-stream perturbation technique, enabling two strong views to be simultaneously guided by a common weak view. Consequently, our overall Unified Dual-Stream Perturbations approach (UniMatch) surpasses all existing methods significantly across all evaluation proto- cols on the Pascal, Cityscapes, and COCO benchmarks. Its superiority is also demonstrated in remote sensing interpre- tation and medical image analysis. We hope our reproduced FixMatch and our results can inspire more future works.	1. Introduction Semantic segmentation aims to provide pixel-level pre- dictions to images, which can be deemed as a dense classi- fication task and is fundamental to real-world applications, e.g., autonomous driving. Nevertheless, conventional fully- supervised scenario [43, 73, 77] is extremely hungry for deli- cately labeled images by human annotators, greatly hinder- ing its broad application to some fields where it is costly and even infeasible to annotate abundant images. Therefore, semi-supervised semantic segmentation [56] has been pro- posed and is attracting increasing attention. Generally, it wishes to alleviate the labor-intensive process via leverag- *Corresponding author. 183 366 732 1464 # labeled images (10582 in total)6568717477mIOU (%) PseudoSeg [ICLR'2021] CPS [CVPR'2021] PC2Seg [ICCV'2021] ReCo [ICLR'2022] ST++ [CVPR'2022] U2PL [CVPR'2022] Reproduced FixMatchFigure 1. Comparison between state-of-the-art methods and our reproduced FixMatch [55] on the Pascal dataset. ing a large quantity of unlabeled images, accompanied by a handful of manually labeled images. Following closely the research line of semi-supervised learning (SSL), advanced methods in semi-supervised se- mantic segmentation have evolved from GANs-based adver- sarial training paradigm [21, 47, 56] into the widely adopted consistency regularization framework [13, 19, 28, 29, 49, 61, 81] and reborn self-training pipeline [23, 27, 68, 70]. In this work, we focus on the weak-to-strong consistency regulariza- tion framework, which is popularized by FixMatch [55] from the field of semi-supervised classification, and then impacts many other relevant tasks [42, 45, 57, 62, 66, 67]. The weak- to-strong approach supervises a strongly perturbed unlabeled image xswith the prediction yielded from its correspond- ing weakly perturbed version xw, as illustrated in Figure 2a. Intuitively, its success lies in that the model is more likely to produce high-quality prediction on xw, while xsis more effective for our model to learn, since the strong perturba- tions introduce additional information as well as mitigate confirmation bias [2]. We surprisingly notice that, so long as coupled with appropriate strong perturbations, FixMatch can indeed still exhibit powerful generalization capability in our scenario, obtaining superior results over state-of-the-art (SOTA) methods, as compared in Figure 1. Thus, we select this simple yet effective framework as our baseline. Through investigation of image-level strong perturbations, 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. 7236 Method# labeled images (10582 in total) 92 183 366 732 1464 w/o any SP 39.5 52.7 65.5 69.2 74.6 w/ CutMix 56.7 67.9 71.9 75.1 78.3 w/ whole SP 63.9 73.0 75.5 77.8 79.2 Table 1. The importance of image-level strong perturbations (SP) to FixMatch on the Pascal dataset. w/o any SP : directly utilize hard label of xwto supervise its logits. w/ CutMix : only use CutMix [71] as a perturbation. w/ whole SP : strong perturbations contain color transformations from ST++ [68], together with CutMix. we observe that they play an indispensable role in making the FixMatch a rather strong competitor in semi-supervised semantic segmentation. As demonstrated in Table 1, the performance gap between whether to adopt perturbations is extremely huge. Greatly inspired by these clues, we hope to inherit the spirit of strong perturbations from FixMatch, but also further strengthen them from two different perspectives and directions, namely expanding a broader perturbation space , and sufficiently harvesting original perturbations . Each of these two perspectives is detailed in the following two paragraphs respectively. Image-level perturbations, e.g., color jitter and CutMix [71], include heuristic biases, which actually introduce ad- ditional prior information into the bootstrapping paradigm of FixMatch, so as to capture the merits of consistency reg- ularization. In case not equipped with these perturbations, FixMatch will be degenerated to a naïve online self-training pipeline, producing much worse results. Despite its effective- ness, these perturbations are totally constrained at the image level, hindering the model to explore a broader perturbation space and to maintain consistency at diverse levels. To this end, in order to expand original perturbation space, we de- sign a unified perturbation framework for both raw images and extracted features. Concretely, on raw images, similar to FixMatch, pre-defined image-level strong perturbations are applied, while for extracted features of weakly perturbed im- ages, an embarrassingly simple channel dropout is inserted. In this way, our model pursues the equivalence of predictions on unlabeled images at both the image and embedding level. These two perturbation levels can be complementary to each other. Distinguished from [33, 41], we separate different levels of perturbations into independent streams to avoid a single stream being excessively hard to learn. On the other hand, current FixMatch framework merely utilizes a single strong view of each unlabeled image in a mini-batch, which is insufficient to fully exploit the manually pre-defined perturbation space. Considering this, we present a simple yet highly effective improvement to the input, where dual independent strong views are randomly sampled from the perturbation pool. They are then fed into the student model in parallel, and simultaneously supervised by theirshared weak view. Such a minor modification even easily turns the FixMatch baseline into a stronger SOTA framework by itself. Intuitively, we conjecture that enforcing two strong views to be close to a common weak view can be regarded as minimizing the distance between these strong views. Hence, it shares the spirits and merits of contrastive learning [11,25], which can learn more discriminative representations and is proved to be particularly beneficial to our current task [40, 61]. We conduct comprehensive studies on the effectiveness of each proposed component. Our contributions can be summarized in four folds: •We notice that, coupled with appropriate image-level strong perturbations, FixMatch is still a powerful frame- work when transferred to the semantic segmentation scenario. A plainly reproduced FixMatch outperforms almost all existing methods in our current task. •Built upon FixMatch, we propose a unified perturba- tion framework that unifies image-level and feature- level perturbations in independent streams, to exploit a broader perturbation space. •We design a dual-stream perturbation strategy to fully probe pre-defined image-level perturbation space, as well as to harvest the merits of contrastive learning for discriminative representations. •Our framework that integrates above two components, surpasses existing methods remarkably across all evalu- ation protocols on the Pascal, Cityscapes, and COCO. Notably, it also exhibits strong superiority in medical image analysis and remote sensing interpretation.
Zhang_Improving_Graph_Representation_for_Point_Cloud_Segmentation_via_Attentive_Filtering_CVPR_2023	Abstract Recently, self-attention networks achieve impressive per- formance in point cloud segmentation due to their superior- ity in modeling long-range dependencies. However, com- pared to self-attention mechanism, we find graph convolu- tions show a stronger ability in capturing local geometry information with less computational cost. In this paper, we employ a hybrid architecture design to construct our Graph Convolution Network with Attentive Filtering ( AF-GCN ), which takes advantage of both graph convolution and self- attention mechanism. We adopt graph convolutions to ag- gregate local features in the shallow encoder stages, while in the deeper stages, we propose a self-attention-like mod- ule named Graph Attentive Filter (GAF) to better model long-range contexts from distant neighbors. Besides, to fur- ther improve graph representation for point cloud segmen- tation, we employ a Spatial Feature Projection (SFP) mod- ule for graph convolutions which helps to handle spatial variations of unstructured point clouds. Finally, a graph- shared down-sampling and up-sampling strategy is intro- duced to make full use of the graph structures in point cloud processing. We conduct extensive experiments on multi- ple datasets including S3DIS, ScanNetV2, Toronto-3D, and ShapeNetPart. Experimental results show our AF-GCN ob- tains competitive performance.	1. Introduction With the rapid development of 3D sensing technolo- gies (such as LiDARs and RGB-D cameras), 3D point clouds have demonstrated great potential in many applica- tions such as robotics, autonomous driving, virtual reality and augmented reality [10]. Consequently, point cloud seg- mentation has attracted more and more attention. Unlike regular pixel grids in 2D images, 3D points in point clouds are irregular and unstructured, thereby posing significant Corresponding author: Wei Gao. (a) Input (b) Ground Truth (c) Low -level Features (d) High- level FeaturesFigure 1. (a) and (b) are the input point cloud and corresponding semantic labels, respectively. (c) and (d) are the visualization of the low-level and high-level point features after the first and last down-sampling, respectively. Differences in color indicate differ- ences in features. As shown in (c), neighbors in the same object may have low feature correlations due to the differences in RGB attributes or geometry structures. As shown in (d), points after sev- eral down-sampling are sparse and the distant neighbors in high- level feature aggregation should be filtered because of containing possible irrelevant information. challenges for point cloud segmentation. Several researches [18, 21, 46] adopt graph convolution networks to utilize the topological structure of point cloud for segmentation. Graph convolution networks learn fea- tures from points and their neighbors for better captur- ing local geometric features while maintaining permuta- tion invariance, which have intrinsic advantages for han- dling non-Euclidean point cloud data. Furthermore, many works [17, 44, 51, 59] improve the graph convolution net- works by proposing well-designed convolution kernels and get promising performance in point cloud segmentation. Recently inspired by the great success of vision trans- formers [8,11,23,34,56], several works [9,15,36,52,55,58] 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. 1244 introduce self-attention mechanism into point cloud anal- ysis for its superiority in modeling long-range dependen- cies and high-level relations, which obtain significant per- formance improvement, especially in point cloud segmen- tation. However, self-attention mechanism exhibits certain limitations in capturing local geometry information. Com- pared with graph convolutions, self-attention mechanisms require additional computation for feature correlations, and assign large weights to neighbors which have high feature correlations. As illustrated in Figure 1, points in low-level feature learning phases are dense and low-level features are mainly extracted from the colors and geometry structures (like edges, corners and surfaces). Therefore, self-attention mechanisms are inefficient in low-level feature aggregation and may neglect information about neighbors which have considerable differences in colors or geometry structures. To exploit the advantages of graph convolution in cap- turing local geometry information and self-attention mech- anism in modeling long-range contexts simultaneously, we design a hybrid network, namely Graph Convolution Net- work with Attentive Filtering (AF-GCN). In the shallow stages of the encoder, we adopt graph convolutions to ag- gregate local geometry information. While in the deeper stages, we propose a self-attention-like module in the graph convolution form called Graph Attentive Filter (GAF) to improve graph representation for point cloud segmentation. Different from previous studies [9, 15, 36, 50, 58], our pro- posed Graph Attentive Filter estimates the correlation be- tween the points from both features and spatial structures in- formation, then suppresses irrelevant information from the distant neighbors to better capture high-level relations. To further improve our graph convolution networks for point cloud segmentation, we adopt a Spatial Feature Pro- jection (SFP) module for graph convolutions. The spa- tial feature projection module projects the spatial informa- tion of points into the feature space, which helps graph convolutions with isotropic kernels to model spatial vari- ations effectively. Moreover, we design a graph-shared down-sampling and up-sampling strategy to better utilize the graph structures in the decoder. In general, our key con- tributions are summarized as follows: • We construct a hierarchical graph convolution network AF-GCN with a hybrid architecture design for point cloud segmentation, which takes advantage of graph convolution and self-attention mechanism. • We propose a novel Graph Attentive Filters module to suppress irrelevant information from distant neighbors by estimating the correlation between the points from both features and spatial structure information. • We employ a Spatial Feature Projection module for graph convolutions to handle the spatial variation of ir- regular point clouds. To better exploit the graph struc-tures, we design a graph-shared down-sampling and up-sampling strategy. • Experimental results demonstrate our model achieves state-of-the-art performance on multiple point cloud segmentation datasets. Ablation studies also verify the effectiveness of each proposed component.
Yu_Accidental_Light_Probes_CVPR_2023	Abstract Recovering lighting in a scene from a single image is a fundamental problem in computer vision. While a mir- ror ball light probe can capture omnidirectional lighting, light probes are generally unavailable in everyday images. In this work, we study recovering lighting from accidental light probes (ALPs)—common, shiny objects like Coke cans, which often accidentally appear in daily scenes. We propose a physically-based approach to model ALPs and estimate lighting from their appearances in single images. The main idea is to model the appearance of ALPs by photogram- metrically principled shading and to invert this process via differentiable rendering to recover incidental illumination. We demonstrate that we can put an ALP into a scene to allow high-fidelity lighting estimation. Our model can also recover lighting for existing images that happen to contain an ALP *. I’d rather be Shiny. — Tamatoa from Moana, 2016	1. Introduction Traditionally, scene lighting has been captured through the use of light probes, typically a chromium mirror ball; their shape (perfect sphere) and material (perfect mirror) allow for a perfect measurement of all light that intersects the probe. Unfortunately, perfect light probes rarely appear in everyday photos, and it is unusual for people to carry them around to place in scenes. Fortunately, many everyday objects share the desired properties of light probes: Coke cans, rings, and thermos bottles are shiny (high reflectance) and curved (have a variety of surface normals). These ob- jects can reveal a significant amount of information about the scene lighting, and can be seen as imperfect “accidental” light probes (e.g., the Diet Pepsi in Figure 1). Unlike perfect light probes, they can easily be found in casual photos or acquired and placed in a scene. In this paper, we explore us- ing such everyday, shiny, curved objects as Accidental Light Probes (ALPs) to estimate lighting from a single image. *Project website: https://kovenyu.com/ALP Figure 1. (Left) From an image that has an accidental light probe (a Diet Pepsi can), we insert a virtual object (a Diet Coke can) with estimated lighting using the accidental light probe (Middle), and using estimated lighting from a recent state-of-the-art lighting estimation method [49] (Right). Note how our method better re- lights the inserted can to produce an appearance consistent with the environment (e.g., the highlight reflection and overall intensity). In general, recovering scene illumination from a single view is fundamental for many computer vision applications such as virtual object insertion [9], relighting [46], and pho- torealistic data augmentation [51]. Yet, it remains an open problem primarily due to its highly ill-posed nature. Images are formed through a complex interaction between geometry, material, and lighting [21], and without precise prior knowl- edge of a scene’s geometry or materials, lighting estimation is extremely under-constrained. For example, scenes that consist primarily of matte materials reveal little information about lighting, since diffuse surfaces behave like low-pass filters on lighting during the shading process [38], eliminat- ing high-frequency lighting information. To compensate for the missing information, the computer vision community has explored using deep learning to extract data-driven priors for lighting estimation [14, 44]. However, these methods gener- ally do not leverage physical measurements to address these ambiguities, yet physical measurements can offer substantial benefits in such an ill-posed setting. For images with ALPs, we propose a physically-based 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. 12521 modeling approach for lighting estimation. The main idea is to model the ALP appearance using physically-based shad- ing and to invert this process to estimate lighting. This inversion process involves taking an input image, estimating the ALP’s 6D pose and scale, and then using the object’s surface geometry and material to infer lighting. Compared to purely data-driven learning approaches that rely on diverse, high-quality lighting datasets, which are hard to acquire, our physically-based approach generalizes to different indoor and outdoor scenes. To evaluate this technique, we collect a test set of real images, where we put ALPs in daily scenes and show that our approach can estimate high-fidelity lighting. We also demonstrate lighting estimation and object insertion based on existing images (Figure 1). In summary, we make the following three contributions: •We propose the concept of accidental light probes (ALPs), which can provide strong lighting cues in ev- eryday scenes and casual photos. •We develop a physically-based approach for lighting estimation for images with an ALP and show improved visual performance compared to existing light estima- tion techniques. •We collect a dataset of ALPs and a dataset of images with ALPs and light probes in both indoor and out- door scenes. We demonstrate that our physically-based model outperforms existing methods on these datasets.
Zhang_Object_Detection_With_Self-Supervised_Scene_Adaptation_CVPR_2023	Abstract This paper proposes a novel method to improve the per- formance of a trained object detector on scenes with fixed camera perspectives based on self-supervised adaptation. Given a specific scene, the trained detector is adapted using pseudo-ground truth labels generated by the detector itself and an object tracker in a cross-teaching manner. When the camera perspective is fixed, our method can utilize the background equivariance by proposing artifact-free object mixup as a means of data augmentation, and utilize accu- rate background extraction as an additional input modal- ity. We also introduce a large-scale and diverse dataset for the development and evaluation of scene-adaptive ob- ject detection. Experiments on this dataset show that our method can improve the average precision of the original detector, outperforming the previous state-of-the-art self- supervised domain adaptive object detection methods by a large margin. Our dataset and code are published at https://github.com/cvlab-stonybrook/scenes100.	1. Introduction The need to detect objects in video streams from station- ary cameras arises in many computer vision applications, including video surveillance and autonomous retail. In gen- eral, different applications require the detection of different object categories, and each computer-vision-based product will have its own detector. However, for a specific product, there is typically a single detector that will be used for many cameras/scenes. For example, a typical video surveillance product would use the same detector to detect pedestrians and vehicles for network cameras installed at different loca- tions. Unfortunately, a single detector might not work well for all scenes, leading to trivial and unforgiving mistakes. This fundamental problem of many computer vision products stems from the limited generalization power of a single model, due to limited training data, limited model ca- pacity, or both. One can attempt to address this problem byusing more training data, but it will incur additional cost for data collection and annotation. Furthermore, in many cases, due to the low latency requirement or the limited comput- ing resources for inference, a product is forced to use a very lightweight network, and this network will have limited rep- resentation capacity to generalize across many scenes. In this paper, instead of having a single scene-generic detector, we propose using scene-specific detectors. This yields higher detection performance as each detector is customized for a specific scene, and allows us to use a lightweight model without sacrificing accuracy as each de- tector is only responsible for one scene. Obtaining scene-specific detectors, however, is very challenging. A trivial approach is to train a detector for each scene separately, but this requires an enormous amount of annotated training data. Instead, we propose a self- supervised method to adapt a pre-trained detector to each scene. Our method records the unlabeled video frames in the past, uses the trained detector to detect objects in those frames, and generates augmented training data based on those detections. Although the detections made by the pre-trained model can be noisy, they can still be useful for generating pseudo annotated data. We further extend those pseudo bounding boxes by applying object tracking [2, 57] along the video timeline, aiming to propagate the detections to adjacent frames to recover some of the false negatives not returned by the detector. We also use multiple detectors to obtain the pseudo labels and train the detector in a cross- teaching manner, taking the advantage of the ensemble of models [13, 24]. Exploiting the stationary nature of the camera, we pro- pose two additional techniques to boost the detection perfor- mance: location-aware mixup and background-augmented input. The former is to generate more samples during train- ing through object mixup [76] that contains less artifacts, based on the aforementioned pseudo boxes generated from detection and tracking. The latter involves estimating the background image and fusing it with the detector’s input. In short, the main contribution of our paper is a novel framework that utilizes self-supervision, location-aware ob- 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. 21589 ject mixup, and background modeling to improve the detec- tion performance of a pre-trained object detector on scenes with stationary cameras. We also contribute a large scale and diverse dataset for the development of scene adap- tive object detection, which contains sufficient quantity and quality annotations for evaluation.
Zhang_Frequency-Modulated_Point_Cloud_Rendering_With_Easy_Editing_CVPR_2023	Abstract We develop an effective point cloud rendering pipeline for novel view synthesis, which enables high fidelity local detail reconstruction, real-time rendering and user-friendly editing. In the heart of our pipeline is an adaptive fre- quency modulation module called Adaptive Frequency Net (AFNet) , which utilizes a hypernetwork to learn the local texture frequency encoding that is consecutively injected into adaptive frequency activation layers to modulate the implicit radiance signal. This mechanism improves the fre- quency expressive ability of the network with richer fre- quency basis support, only at a small computational bud- get. To further boost performance, a preprocessing mod- ule is also proposed for point cloud geometry optimization via point opacity estimation. In contrast to implicit render- ing, our pipeline supports high-fidelity interactive editing based on point cloud manipulation. Extensive experimental results on NeRF-Synthetic, ScanNet, DTU and Tanks and Temples datasets demonstrate the superior performances achieved by our method in terms of PSNR, SSIM and LPIPS, in comparison to the state-of-the-art. Code is released at https://github.com/yizhangphd/FreqPCR.	1. Introduction Photo-realistic rendering and editing of 3D representa- tions is a key problem in 3D computer vision and graph- ics with numerous applications, such as computer games, *Equal contribution. †Corresponding author: Bingbing Ni.VR/AR, and video creation. In particular, recently intro- duced neural radiance field (NeRF) [21] has inspired some follow-up works aiming to editable rendering [15, 17, 20, 47, 49, 55]. Due to the deeply coupled black-box net- work, NeRF-based object-level editing usually requires a pre-trained segmentation model to separate the objects to be edited [17, 55]. Although some recent voxel-based variants of NeRF [47,58] achieve multi-scene composition, they still lack the ability to extract target objects from voxels. In contrast to implicit rendering, point cloud render- ing [1,6,13,18,33,36,57] is a promising editable rendering model. On the one hand, explicit 3D representations are bet- ter for interactive editing. On the other hand, the geometric priors provided by point clouds can help us avoid massive sampling in volume rendering methods, which can meet the requirements of some real-time applications. As a class of representative point cloud rendering methods, NPBG and NPBG++ [1,33] achieve real-time rendering by using point- wise features for encoding appearance information and an U-Net [35] for decoding, respectively. However, the param- eter quantity increases with the size of point cloud, which may limit their application due to the excessive computa- tional and memory complexity. Huang et al. [13] combine point clouds with implicit rendering, where explicit point clouds are used to estimate the geometry, and implicit radi- ance mapping is used to predict view-dependent appearance of surfaces. However, quantitative evaluation of its render- ing results is significantly lower than that of implicit render- ing methods such as NeRF [21], mainly due to the following reasons: 1) the color of each viewing ray only depends on 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. 119 a single surface point, thus without multiple sample color aggregation for error attenuation, surface based rendering techniques require radiance mapping to have a more precise and expressive frequency encoding ability; and 2) defects of the point cloud geometry reconstructed by MVSNet [56] cause wrong surface estimation. To this end, we introduce Adaptive Frequency Net (AFNet) to improve frequency ex- pression ability of the radiance mapping and a preprocess- ing module for point cloud geometry optimization. Radiance mapping, also known as radiance field, is a type of Implicit Neural Representation (INR). There have been some studies [2, 7, 32, 41, 45, 46, 60] on the expressive power and inductive bias of INRs. The standard Multi-layer Perceptrons (MLPs) with ReLU activation function are well known for the strong spectral bias towards reconstructing low frequency signals [32]. Some recent works [7, 41, 46] introduce strategies to enhance the high-frequency repre- sentation of MLPs from a global perspective. However, from a local perspective, the frequencies of a 3D scene are region-dependent and most real objects are composed by both weak textured regions and strong textured ones. Moti- vated by this, we design a novel adaptive frequency modula- tion mechanism based on HyperNetwork architecture [11], which learns the local texture frequency and injects it into adaptive frequency activation layers to modulate the im- plicit radiance signal. The proposed mechanism can pre- dict suitable frequency without frequency supervision and modulate the radiance signal with adaptive frequency ba- sis support to express more complex textures at negligible computational overhead. Previous surface point-based works [1, 13, 33] could not optimize the point cloud geometry because they keep only the closest point as a surface estimation for each ray. But if we sample multiple points per ray during rendering, it will greatly reduce the rendering speed. Therefore, we use the volume rendering method as a preprocessing module to op- timize the point cloud geometry. Specifically, we keep more points in the pixel buffer and learn the opacity of each point based on volume rendering. We find in our experiments that for some poorly reconstructed scenes, point cloud geome- try optimization can improve the rendering PSNR by 2-4dB and avoid rendering artifacts. For rigid object editing, we follow the deformation field construction [28–31, 48] to render the edited point cloud. Point cloud can be seen as a bridge between user editing and deformation field to achieve interactive editing and ren- dering. Users only need to edit the point cloud, and the deformation field between the original 3D space and the de- formed space is easy to obtain by querying the correspond- ing transformations performed on point cloud. Moreover, to avoid cross-scene training in multi-scene composition ap- plication, we develop a masking strategy based on depth buffer to combine multiple scenes in pixel level.We evaluate our method on NeRF-Synthetic [21], Scan- Net [5], DTU [14] and Tanks and Temples [16] datasets and comprehensively compare the proposed method with other works in terms of performance (including PSNR, SSIM and LPIPS), model complexity, rendering speed, and editing ability. Our performance outperforms NeRF [21] and all surface point-based rendering methods [1, 13, 33], and is comparable to Compressible-composable NeRF (CC- NeRF), i.e., the latest NeRF-based editable variant. We achieve a real-time rendering speed of 39.27 FPS on NeRF- Synthetic, which is 1700 ×faster than NeRF and 37 × faster than CCNeRF. We also reproduce the scene editing of Object-NeRF [55] and CCNeRF [47] on ToyDesk [55] and NeRF-Synthetic [21] dataset, respectively. As shown in Fig. 1, we achieve real-time rendering with sharper de- tails and user-friendly editing. The above results demon- strate that our method is comprehensive in terms of both rendering and editing and has great application potentials.
Xu_Unsupervised_3D_Shape_Reconstruction_by_Part_Retrieval_and_Assembly_CVPR_2023	Abstract Representing a 3D shape with a set of primitives can aid perception of structure, improve robotic object manipula- tion, and enable editing, stylization, and compression of 3D shapes. Existing methods either use simple paramet- ric primitives or learn a generative shape space of parts. Both have limitations: parametric primitives lead to coarse approximations, while learned parts offer too little control over the decomposition. We instead propose to decompose shapes using a library of 3D parts provided by the user, giving full control over the choice of parts. The library can contain parts with high-quality geometry that are suit- able for a given category, resulting in meaningful decom- positions with clean geometry. The type of decomposition can also be controlled through the choice of parts in the li- brary. Our method works via a unsupervised approach that iteratively retrieves parts from the library and refines their placements. We show that this approach gives higher recon- struction accuracy and more desirable decompositions than existing approaches. Additionally, we show how the decom- position can be controlled through the part library by using different part libraries to reconstruct the same shapes.	1. Introduction The ability to compactly represent a 3D shape as a com- bination of primitive elements has applications in multiple domains. In computer vision, the ability to automatically decompose a shape into parts can aid machine perception of the 3D structure of objects, which can in turn help au-tonomous agents plan how to manipulate such objects. In computer graphics, a combination of primitives can be used as a compressed geometry representation, as a way to abstract, stylize, or edit a 3D shape by allowing users to alter the underlying primitive library. Ideally, a system that performs this kind of shape decomposition should be able to do so without supervision in the form of ground-truth decompositions, as such data is rarely available at scale. Past research in vision and graphics has studied this unsupervised shape decomposition problem. Initially, re- searchers sought methods for decomposing 3D shapes into sets of simple parametric primitives, such as cuboids or su- perquadric surfaces [18,21,23,26]. These methods produce clean, parametric geometry as output, and the choice of primitive type allows a small degree of user control over the decomposition. However, parametric primitives produce only a coarse approximation of the input shape, which may not be desirable in all applications. Thus, more recent work has investigated unsupervised decomposition of shapes into arbitrarily-shaped primitives whose geometries are deter- mined by a neural network [4, 10, 17]. These methods pro- duce a set of “neural primitives” whose union closely ap- proximates the input shape. However, the geometries of these primitives may contain artifacts (e.g. bumps, blobs). Further, these methods offer little to no control over the type of decomposition they produce – the network outputs what- ever primitives it thinks are best to reconstruct the input shape since it lacks access to a supervised part prior. Is it possible to obtain a decomposition of a 3D shape whose primitives exhibit clean geometry and closely recon- struct the input shape, while also providing more control 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. 8559 over the type of decomposition produced? This is possible if, rather than using simple parametric primitives or arbi- trary neural primitives, one chooses a middle point between these two extremes: reconstruct an input shape by retrieving and assembling primitives from a library of pre-defined 3D parts . This retrieve-and-assemble approach has several ad- vantages. First, the parts in the library can be high-quality 3D meshes, guaranteeing clean geometry as output. Sec- ond, a large part library can contain parts that are good geometric matches for different regions of various shapes, meaning that accurate reconstructions of input shapes are possible. Finally, this approach offers a high degree of con- trollability, as the user can change the part library to produce different decompositions of the same input shape. In this paper, we present a method for unsupervised de- composition of 3D shapes using a user-defined library of parts. Finding a subset of parts from a large part library which best reconstructs an input shape is a large-scale com- binatorial search problem. To make this problem tractable, we represent the library of parts on a continuous manifold by training a part autoencoder. This continuous representa- tion of the part library allows us jointly optimize for the identities and poses of parts which reconstruct the input shape. To escape the many worst of local optimas in this optimization landscape, the algorithm periodically uses its current predicted set of parts to segment the input shape; these segments are then re-encoded into the part feature manifold to produce a new estimate of the parts that best reconstruct the input shape. This data-driven, discontinu- ous jump in the optimization state is similar to stages from other non-gradient-based algorithms for global optimiza- tion or latent variable estimation, including the mean shift step from the mean shift algorithm and the E-step from ex- pectation maximization [2]. Our algorithm can be run independently for any individ- ual target shape, allowing it to work in a “zero-shot” set- ting. When a larger dataset of related shapes is available, we can also optimize for their part decomposition in advance (a “training” phase) and then perform fast decomposition of a new shape from that category by initializing its decomposi- tion using its nearest neighbor from the “training” set. We evaluate our algorithm by using it to reconstruct shapes from point clouds, using parts from the PartNet dataset. We compare to the recent Neural Parts unsuper- vised decomposition system [17] and show that our al- gorithm produces qualitatively more desirable decomposi- tions that also achieve higher reconstruction accuracy. We demonstrate the control offered by our method by show- ing how it is possible to reconstruct shapes from one cat- egory using parts from another (e.g. make a chair out of lamp parts). This also has application for 3D graphics con- tent creation, which we demonstrate by reconstructing tar- get shapes using parts from a modular 3D asset library.In summary, our contribution is an unsupervised algo- rithm which retrieves and poses 3D parts to reconstruct in- put 3D shapes. We will release our code upon publication.
Yu_Graphics_Capsule_Learning_Hierarchical_3D_Face_Representations_From_2D_Images_CVPR_2023	Abstract The function of constructing the hierarchy of objects is important to the visual process of the human brain. Previ- ous studies have successfully adopted capsule networks to decompose the digits and faces into parts in an unsuper- vised manner to investigate the similar perception mecha- nism of neural networks. However, their descriptions are restricted to the 2D space, limiting their capacities to imi- tate the intrinsic 3D perception ability of humans. In this paper, we propose an Inverse Graphics Capsule Network (IGC-Net) to learn the hierarchical 3D face representations from large-scale unlabeled images. The core of IGC-Net is a new type of capsule, named graphics capsule, which rep- resents 3D primitives with interpretable parameters in com- puter graphics (CG), including depth, albedo, and 3D pose. Speciﬁcally, IGC-Net ﬁrst decomposes the objects into a set of semantic-consistent part-level descriptions and then as- sembles them into object-level descriptions to build the hier- archy. The learned graphics capsules reveal how the neural networks, oriented at visual perception, understand faces as a hierarchy of 3D models. Besides, the discovered parts can be deployed to the unsupervised face segmentation task to evaluate the semantic consistency of our method. Moreover, the part-level descriptions with explicit physical meanings provide insight into the face analysis that originally runs in a black box, such as the importance of shape and texture for face recognition. Experiments on CelebA, BP4D, and Multi-PIE demonstrate the characteristics of our IGC-Net. Corresponding author.	1. Introduction A path toward autonomous machine intelligence is to en- able machines to have human-like perception and learning abilities [19]. As humans, by only observing the objects, we can easily decompose them into a set of part-level com- ponents and construct their hierarchy even though we have never seen these objects before. This phenomenon is sup- ported by the psychological studies that the visual process of the human brain is related to the construction of the hi- erarchical structural descriptions [11,22,23,29]. To investi- gate the similar perception mechanism of neural networks, previous studies [18, 35] incorporate the capsule networks, which are designed to present the hierarchy of objects and describe each entity with interpretable parameters. After observing a large-scale of unlabeled images, these meth- ods successfully decompose the digits or faces into a set of parts, which provide insight into how the neural networks understand the objects. However, their representations are limited in the 2D space. Speciﬁcally, these methods follow the analysis-by-synthesis strategy in model training and try to reconstruct the image by the decomposed parts. Since the parts are represented by 2D templates, the reconstruction becomes estimating the afﬁne transformations to warp the templates and put them in the right places, which is just like painting with stickers. This strategy performs well when the objects are intrinsically 2D, like handwritten digits and frontal faces, but has difﬁculty in interpreting 3D objects in the real world, especially when we want a view-independent representation like humans [2]. How to represent the perceived objects is the core re- search topic in computer vision [3, 25]. One of the most popular theories is the Marr’s theory [22, 23]. He believed that the purpose of the vision is to build the descriptions 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. 20981 of shapes and positions of things from the images and con- struct hierarchical 3D representations of objects for recog- nition. In this paper, we try to materialize Marr’s the- ory on human faces and propose an Inverse Graphics Cap- sule Network (IGC-Net), whose primitive is a new type of capsule (i.e., graphics capsule) that is deﬁned by com- puter graphics (CG), to learn the hierarchical 3D represen- tations from large-scale unlabeled images. Figure 1 shows an overview of the proposed method. Speciﬁcally, the hi- erarchy of the objects is described with the part capsules and the object capsules, where each capsule contains a set of interpretable parameters with explicit physical meanings, including depth, albedo, and pose. During training, the in- put image is ﬁrst encoded to a global shape and albedo em- beddings, which are sent to a decomposition module to get the spatially-decoupled part-level graphics capsules. Then, these capsules are decoded by a shared capsule decoder to get explicit 3D descriptions of parts. Afterward, the parts are assembled by their depth to generate the object capsules as the object-centered representations, naturally construct- ing the part-object hierarchy. Finally, the 3D objects em- bedded in the object capsules are illuminated, posed, and rendered to ﬁt the input image, following the analysis-by- synthesis manner. When an IGC-Net is well trained, the learned graphics capsules naturally build hierarchical 3D representations. We apply IGC-Net to human faces, which have been widely used to investigate human vision system [31] due to the similar topology structures and complicated appear- ances. Thanks to the capacity of the 3D descriptions, IGC- Net successfully builds the hierarchy of in-the-wild faces that are captured under various illuminations and poses. We evaluate the IGC-Net performance on the unsupervised face segmentation task, where the silhouettes of the discovered parts are regarded as segment maps. We also incorporate the IGC-Net into interpretable face analysis to uncover the mechanism of neural networks when recognizing faces. The main contributions of this paper are summarized as: This paper proposes an Inverse Graphics Capsule Net- work (IGC-Net) to learn the hierarchical 3D face repre- sentations from unlabeled images. The learned graph- ics capsules in the network provide insight into how the neural networks, oriented at visual perception, un- derstand faces as a hierarchy of 3D models. A Graphics Decomposition Module (GDM) is pro- posed for part-level decomposition, which incorpo- rates shape and albedo information as cues to ensure that each part capsule represents a semantically con- sistent part of objects. We execute the interpretable face analysis based on the part-level 3D descriptions of graphics capsules. Be- sides, the silhouettes of 3D parts are deployed to theunsupervised face segmentation task. Experiments on CelebA, BP4D, and Multi-PIE show the effectiveness of our method.
Yin_3D_GAN_Inversion_With_Facial_Symmetry_Prior_CVPR_2023	Abstract Recently, a surge of high-quality 3D-aware GANs have been proposed, which leverage the generative power of neu- ral rendering. It is natural to associate 3D GANs with GAN inversion methods to project a real image into the gener- ator’s latent space, allowing free-view consistent synthesis and editing, referred as 3D GAN inversion. Although with the facial prior preserved in pre-trained 3D GANs, recon- structing a 3D portrait with only one monocular image is still an ill-pose problem. The straightforward application of 2D GAN inversion methods focuses on texture similarity only while ignoring the correctness of 3D geometry shapes. It may raise geometry collapse effects, especially when re- constructing a side face under an extreme pose. Besides, the synthetic results in novel views are prone to be blurry. In this work, we propose a novel method to promote 3D GAN inversion by introducing facial symmetry prior. We design a pipeline and constraints to make full use of the pseudo auxiliary view obtained via image flipping, which helps obtain a view-consistent and well-structured geome- try shape during the inversion process. To enhance texture fidelity in unobserved viewpoints, pseudo labels from depth- guided 3D warping can provide extra supervision. We de- sign constraints to filter out conflict areas for optimization in asymmetric situations. Comprehensive quantitative and qualitative evaluations on image reconstruction and editing demonstrate the superiority of our method.	1. Introduction Recent 3D-aware generative adversarial networks (3D GANs) have seen immense progress. By incorporating a neural rendering engine into the generator network ar- chitecture, 3D GANs can synthesize view-consistent im- ages. To increase the generation resolution, existing meth- ods [ 5,12,25,30,31,36±38,41] boost the 3D inductive bias Work done during an internship at Tencent AI Lab. ²Corresponding Author. Original View ImageSource Image Source Image Novel View ImageNovel View ShapePTI Ours PTI OursFigure 1. Visual examples of our inversion method. Direct apply- ing 2D GAN inversion methods (PTI [ 28]) to the 3D GAN suffers from inaccurate geometry in novel views. Our method excels in synthesizing consistent geometry and high-fidelity texture in dif- ferent views, even reconstructing a face under an extreme pose. with an additional 2D CNN-based upsampler or an efficient 3D representation modeling method. With tremendous ef- fort, 3D GANs can produce photorealistic images while en- forcing strong 3D consistency across different views. We are interested in the task of reconstructing a human face with 3D geometry and texture given only one monocu- lar image. It is an ill-posed problem and close to the harsh condition of real scenarios. With the power of 3D GANs, it seems achievable via projecting a target image onto the manifold of a pre-trained generator. The process is referred as 3D GAN inversion. A straightforward path is to follow the 2D GAN inversion method [ 28],i.e.,optimizing the la- tent code and the network parameters of the generator to overfit the specific portrait. 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. 342 However, since the ground truth 3D geometry is absent given one monocular image, the inversion result is far from satisfactory. The process of fitting a 3D GAN to one im- age would sacrifice geometric correctness in order to make the synthetic texture as close as possible to the input, even destroying the original semantic-rich latent space. As the optimization process goes, the face geometry tends to de- generate into a flattened shape, due to the absence of geom- etry supervision, e.g., images from other views. Besides, there exist quality issues in texture synthesis under novel views. The rendered images of unseen views tend to be blurry and inconsistent with the original image, especially when reconstructing a side face under an extreme pose. Be- cause there is no texture supervision for unseen views given only one monocular image. The failure cases of directly applying [ 28] are illustrated in Fig. 1. In this work, to alleviate the issue caused by missing ge- ometry and texture supervision under multiple views, we propose a novel 3D GAN inversion approach by taking full advantage of facial symmetry prior to construct pseudo su- pervision of different views. Intuitively, we note that human faces are almost symmetric. Assuming the given portrait is symmetric, we can obtain an additional perspective of the portrait by simply mirroring the image. The images of two distinct views can provide geometric relations between the 3D points and their 2D projections based on epipolar geom- etry. Motivated by this, we seek to leverage facial symmetry as the geometric prior constraining the inversion. The sym- metry prior is also employed in a traditional 3D reconstruc- tion work [ 35]. We leverage the mirrored image as extra supervision of another view when performing the inversion, which prevents the geometry collapse. A rough geometry can be obtained by the inversion with the original and mir- ror images. To further enhance texture quality and geometry in novel views, we employ depth-guided 3D warping to generate the pseudo images of the views surrounding the input and sym- metric camera pose. The depth is inferred from the rough 3D volume. The original image along with the pseudo im- ages are used to fine-tune the generator’s parameters for the joint promotion of texture and geometry. To prevent the op- timized geometry from deviating too much from the rough geometry, we design a geometry regularization term as a constraint. However, human faces are never fully symmet- ric in practice, neither in shape nor appearance. Therefore, we design several constraints to extract meaningful infor- mation adaptively from the mirror image without compro- mising the original reconstruction quality. Our main contributions are as follows: • We propose a novel 3D GAN inversion method by in- corporating facial symmetry prior. It enables a high- quality reconstruction while preserving the multi-view consistency in geometry and texture.• We conduct comprehensive experiments to demon- strate the effectiveness of our method and compare it with many state-of-the-art inversion methods. We also apply our method to various downstream applications.
Yang_PVT-SSD_Single-Stage_3D_Object_Detector_With_Point-Voxel_Transformer_CVPR_2023	Abstract Recent Transformer-based 3D object detectors learn point cloud features either from point- or voxel-based rep- resentations. However, the former requires time-consuming sampling while the latter introduces quantization errors. In this paper, we present a novel Point-Voxel Transformer for single-stage 3D detection (PVT-SSD) that takes advan- tage of these two representations. Specifically, we first use voxel-based sparse convolutions for efficient feature encod- ing. Then, we propose a Point-Voxel Transformer (PVT) module that obtains long-range contexts in a cheap manner from voxels while attaining accurate positions from points. The key to associating the two different representations is our introduced input-dependent Query Initialization mod- ule, which could efficiently generate reference points and content queries. Then, PVT adaptively fuses long-range contextual and local geometric information around refer- ence points into content queries. Further, to quickly find the neighboring points of reference points, we design the Virtual Range Image module, which generalizes the native range image to multi-sensor and multi-frame. The experi- ments on several autonomous driving benchmarks verify the effectiveness and efficiency of the proposed method. Code will be available.	1. Introduction 3D object detection from point clouds has become in- creasingly popular thanks to its wide applications, e.g., autonomous driving and virtual reality. To process un- ordered point clouds, Transformer [51] has recently at- tracted great interest as the self-attention is invariant to the permutation of inputs. However, due to the quadratic ∗This work was done when Honghui was an intern at Shanghai Arti- ficial Intelligence Laboratory. †Corresponding authorcomplexity of self-attention, it involves extensive compu- tation and memory budgets when processing large point clouds. To overcome this problem, some point-based meth- ods [29, 36, 37] perform attention on downsampled point sets, while some voxel-based methods [10, 33, 64] employ attention on local non-empty voxels. Nevertheless, the for- mer requires farthest point sampling (FPS) [41] to sam- ple point clouds, which is time-consuming on large-scale outdoor scenes [19], while the latter inevitably introduces quantization errors during voxelization, which loses accu- rate position information. In this paper, we propose PVT-SSD that absorbs the ad- vantages of the above two representations, i.e., voxels and points, while overcoming their drawbacks. To this end, in- stead of sampling points directly, we convert points to a small number of voxels through sparse convolutions and sample non-empty voxels to reduce the runtime of FPS. Then, inside the PVT-SSD, voxel features are adaptively fused with point features to make up for the quantization er- ror. In this way, both long-range contexts from voxels and accurate positions from points are preserved. Specifically, PVT-SSD consists of the following components: Firstly, we propose an input-dependent Query Initial- ization module inspired by previous indoor Transformer- based detectors [29, 36], which provides queries with bet- ter initial positions and instance-related features. Un- like [29, 36], our queries originate from non-empty voxels instead of points to reduce the sampling time. Concretely, with the 3D voxels generated by sparse convolutions, we firstcollapse 3D voxels into 2D voxels by merging voxels along the height dimension to further reduce the number of voxels. The sample operation is then applied to select a rep- resentative set of voxels. We finally liftsampled 2D voxels to generate 3D reference points. Subsequently, the corre- sponding content queries are obtained in an efficient way by projecting reference points onto a BEV feature map and indexing features at the projected locations. Secondly, we introduce a Point-Voxel Transformer 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. 13476 module that captures long-range contextual features from voxel tokens and extracts fine-grained point features from point tokens. To be specific, the voxel tokens are obtained from non-empty voxels around reference points to cover a large attention range. In contrast, the point tokens are gener- ated from neighboring points near reference points to retain fine-grained information. These two different tokens are adaptively fused by the cross-attention layer based on the similarity with content queries to complement each other. Furthermore, we design a Virtual Range Image mod- ule to accelerate the neighbor querying process in the point- voxel Transformer. With the constructed range image, refer- ence points can quickly find their neighbors based on range image coordinates. Unlike the native range image captured by LiDAR sensors, we can handle situations where multiple points overlap on the same pixel in the range image. There- fore, it can be used for complex scenarios, such as multiple sensors and multi-frame fusion. Extensive experiments have been conducted on several detection benchmarks to verify the efficacy and efficiency of our approach. PVT-SSD achieves competitive results on KITTI [13], Waymo Open Dataset [48], and nuScenes [3].
Zeng_SceneComposer_Any-Level_Semantic_Image_Synthesis_CVPR_2023	Abstract We propose a new framework for conditional image syn- thesis from semantic layouts of any precision levels, ranging from pure text to a 2D semantic canvas with precise shapes. More specifically, the input layout consists of one or more semantic regions with free-form text descriptions and ad- justable precision levels, which can be set based on the desired controllability. The framework naturally reduces to text-to-image (T2I) at the lowest level with no shape informa- tion, and it becomes segmentation-to-image (S2I) at the high- est level. By supporting the levels in-between, our framework is flexible in assisting users of different drawing expertise and at different stages of their creative workflow. We intro- duce several novel techniques to address the challenges com- ing with this new setup, including a pipeline for collecting training data; a precision-encoded mask pyramid and a text feature map representation to jointly encode precision level, semantics, and composition information; and a multi-scale guided diffusion model to synthesize images. To evaluatethe proposed method, we collect a test dataset containing user-drawn layouts with diverse scenes and styles. Experi- mental results show that the proposed method can generate high-quality images following the layout at given precision, and compares favorably against existing methods. Project pagehttps://zengxianyu.github.io/scenec/	1. Introduction Recently, deep generative models such as StyleGAN [24, 25] and diffusion models [9, 19, 49] have made a signif- icant breakthrough in generating high-quality images. Im- age generation and editing technologies enabled by these models have become highly appealing to artists and design- ers by helping their creative workflows. To make image generation more controllable, researchers have put a lot of effort into conditional image synthesis and introduced models using various types and levels of semantic input such as object categories, text prompts, and segmentation 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. 22468 Table 1. Difference from related conditional image synthesis works. T2I: text to image, S2I: segmentation to image, ST2I: Scene-based text to image [13], Box2I: bounding box layout to image [50]. Setting Open-domain layout Shape control Sparse layout Coarse shape Level control T2I ✓ ✗ ✗ ✗ ✗ S2I ✗ ✓ ✗ ✗ ✗ ST2I ✗ ✓ ✗ ✗ ✗ Box2I ✗ ✗ ✓ ✗ ✗ Ours ✓ ✓ ✓ ✓ ✓ maps etc. [23, 35, 36, 43, 44, 67]. However, existing models are not flexible enough to sup- port the full creative workflow. They mostly consider fixed- level semantics as the input [63], e.g. image-level text de- scriptions in text-to-image generation (T2I) [35, 39, 41, 43, 44], or pixel-level segmentation maps in segmentation-to- image generation (S2I) [23,36,67]. Recent breakthroughs on T2I such as DALLE2 [39] and StableDiffusion [1,43] demon- strate extraordinary capabilities of generating high-quality results. They can convert a rough idea into visual messages to provide inspirations at the beginning of the creative pro- cess, but provide no further control over image composition. On the other hand, S2I allows users to precisely control the image composition. As it is extremely challenging to draw a detailed layout directly, S2I is more useful for later cre- ative stages given initial designs. For real-world use cases, it is highly desirable to have a model which can generate images from not only pure text or segmentation maps, but also intermediate-level layouts with coarse shapes. To this end, we propose a new unified conditional image synthesis framework to generate images from a semantic lay- out at any combination of precision levels. It is inspired by the typical coarse-to-fine workflow of artists and designers: they first start from an idea, which can be expressed as a text prompt or a set of concepts (Fig. 1 (a)), then tend to draw the approximate outlines and refine each object (Fig. 1 (a)- (d)). More specifically, we model a semantic layout as a set of semantic regions with free-form text descriptions. The layout can be sparse and each region can have a precision level to control how well the generated object should fit to the specified shape. The framework reduces to T2I when the layout is the coarsest (Fig. 1 (a)), and it becomes S2I when the layout is a segmentation map (Fig. 1 (d)). By adjusting the precision level, users can achieve their desired controlla- bility (Fig. 1 (a)-(d)). This framework is different from the existing works in many aspects, as summarized in Table 1. This new setup comes with several challenges. First, it is non-trivial to encode open-domain layouts in image syn- thesis frameworks. Second, to handle hand-drawn layouts of varying precision, we need an effective and robust way to inject the precision information into the layout encod- ing. Third, there is no large-scale open-domain layout/image dataset. To generate high-quality images and generalize to novel concepts, a large and diverse training dataset is crucial. We introduce several novel ideas to address these chal-lenges. First, we propose a text feature map representation for encoding a semantic layout. It can be seen as a spatial ex- tension of text embedding or generalization of segmentation masks from binary to continuous space. Second, we intro- duce a precision-encoded mask pyramid to model layout precision. Inspired by the classical image pyramid mod- els [2, 6, 47, 62], we relate shape precision to levels in a pyramid representation and encode precision by dropping out regions of lower precision levels. In other words, the l-th level of the mask pyramid is a sub-layout (subset of regions) consisting of semantic regions with precision level no less thanl. By creating a text feature map for each sub-layout, we obtain a text feature pyramid as a unified representation of semantics, composition, and precision. Finally, we feed the text feature pyramid to a multi-scale guided diffusion model to generate images. We fulfill the need for training data by collecting them from two sources: (1) large-scale image- text pairs; (2) a relatively small pseudo layout/image dataset using text-based object detection and segmentation. With this multi-source training strategy, both text-to-image and layout-to-image can benefit from each other synergistically. Our contributions are summarized as follows: •A unified framework for diffusion-based image syn- thesis from semantic layouts with any combination of precision control. •Novel ideas to build the model, including precision- encoded mask pyramid and pyramid text feature map representation, and multi-scale guided diffusion model, and training with multi-source data. •A new real-world user-drawn layout dataset and ex- tensive experiments showing the effectiveness of our model for text-to-image and layout-to-image generation with precision control.
Zhang_Hyperspherical_Embedding_for_Point_Cloud_Completion_CVPR_2023	Abstract Most real-world 3D measurements from depth sensors are incomplete, and to address this issue the point cloud completion task aims to predict the complete shapes of objects from partial observations. Previous works often adapt an encoder-decoder architecture, where the encoder is trained to extract embeddings that are used as inputs to generate predictions from the decoder. However, the learned embeddings have sparse distribution in the feature space, which leads to worse generalization results during testing. To address these problems, this paper proposes a hyperspherical module, which transforms and normal- izes embeddings from the encoder to be on a unit hyper- sphere. With the proposed module, the magnitude and direc- tion of the output hyperspherical embedding are decoupled and only the directional information is optimized. We the- oretically analyze the hyperspherical embedding and show that it enables more stable training with a wider range of learning rates and more compact embedding distributions. Experiment results show consistent improvement of point cloud completion in both single-task and multi-task learn- ing, which demonstrates the effectiveness of the proposed method.	1. Introduction The continual improvement of 3D sensors has made point clouds much more accessible, which drives the de- velopment of algorithms to analyze them. Thanks to deep learning techniques, state of the art algorithms for point cloud analysis have achieved incredible performance [9,20– 22,25] by effectively learning representations from large 3D datasets [3, 6, 26] and have many applications in robotics, autonomous driving, and 3D modeling. However, point clouds in the real-world are often incomplete and sparse due to many reasons, such as occlusions, low resolution, and the limited view of 3D sensors. So it is critical to have an algorithm that is capable of predicting complete shapes of objects from partial observations. Given the importance of point cloud completion, it is Unconstrained Embedding Encoder Completion Decoder Other Decoder Hyperspherical Embedding Completion Decoder Other Decoder MLP Norm Hyper Module Encoder 00.51.0 00.51.0Figure 1. An illustration of the architecture proposed in this paper. The up- per subfigure shows the general point cloud analysis structure, where the embedding is directly output from the encoder without constraints. The lower subfigure shows the structure of the model with the proposed hy- perspherical module. The figures under the embeddings illustrate the co- sine similarity distribution between embeddings, which indicates a more compact embedding distribution achieved by the proposed method and im- proves point cloud completion. unsurprising that various methods have been proposed to address this challenge [18, 27, 31, 34, 37, 41]. Most exist- ing methods adapt encoder-decoder structures, in which the encoder takes a partial point cloud as input and outputs an embedding vector, and then it is taken by the decoder which predicts a complete point cloud. The embedding space is designed to be high-dimensional as it must have large enough capacity to contain all information needed for downstream tasks. However, the learned high-dimensional embeddings, as shown in this paper, tend to have a sparse distribution in the embedding space, which increases the possibility that unseen features at testing are not captured by the representation learned at training and leads to worse generalizability of models. Usually, one real-world application requires predictions from multiple different tasks. For example, to grasp an object in space the robot arm would need the informa- tion about the shape, category, and orientation of the tar- get object. In contrast to training all tasks individually from scratch, a more numerically efficient approach would be to train all relevant tasks jointly by sharing parts of networks between different tasks [11,13,28]. However, existing point cloud completion methods lack the analysis of accomplish- 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. 5323 ing point cloud completion jointly with other tasks. We show that training existing point cloud completion methods with other semantic tasks together leads to worse perfor- mance when compared to learning each individually. To address the above limitations, this paper proposes a hyperspherical module which outputs hyperspherical em- beddings for point cloud completion. The proposed hy- perspherical module can be integrated into existing ap- proaches with encoder-decoder structures as shown in Fig- ure 1. Specifically, the hyperspherical module transforms and constrains the output embedding onto the surface of a hypersphere by normalizing the embedding’s magnitude to unit, so only the directional information is kept for later use. We theoretically investigate the effects of hyperspher- ical embeddings and show that it improves the point cloud completion models by more stable training with large learn- ing rate and more generalizability by learning more com- pact embedding distributions. We also demonstrate the proposed hyperspherical embedding in multi-task learning, where it helps reconcile the learning conflicts between point cloud completion and other semantic tasks at training. The reported improvements of the existing state-of-the-art ap- proaches on several public datasets illustrate the effective- ness of the proposed method. The main contributions of this paper are summarized as follows: • We propose a hyperspherical module that outputs hy- perspherical embeddings, which improves the perfor- mance of point cloud completion. • We theoretically investigate the effects of hyperspher- ical embeddings and demonstrate that the point cloud completion benefits from them by stable training and learning a compact embedding distribution. • We analyze training point cloud completion with other tasks and observe conflicts between them, which can be reconciled by the hyperspherical embedding.
Yi_MIME_Human-Aware_3D_Scene_Generation_CVPR_2023	Abstract Generating realistic 3D worlds occupied by moving hu- mans has many applications in games, architecture, and synthetic data creation. But generating such scenes is ex- pensive and labor intensive. Recent work generates human poses and motions given a 3D scene. Here, we take the opposite approach and generate 3D indoor scenes given 3D human motion. Such motions can come from archival motion capture or from IMU sensors worn on the body, effec- tively turning human movement into a “scanner” of the 3D world. Intuitively, human movement indicates the free-space in a room and human contact indicates surfaces or objects that support activities such as sitting, lying or touching. We propose MIME (Mining Interaction and Movement to infer 3D Environments), which is a generative model of indoor scenes that produces furniture layouts that are consistent with the human movement. MIME uses an auto-regressive transformer architecture that takes the already generated objects in the scene as well as the human motion as input, and outputs the next plausible object. To train MIME, we build a dataset by populating the 3D FRONT scene dataset with 3D humans. Our experiments show that MIME pro- duces more diverse and plausible 3D scenes than a recent generative scene method that does not know about human movement. Code and data are available for research at https://mime.is.tue.mpg.de . *This work was performed when C.P. H. was at the MPI-IS.	1. Introduction Humans constantly interact with their environment. They walk through a room, touch objects, rest on a chair, or sleep in a bed. All these interactions contain information about the scene layout and object placement. In fact, a mime is a performer who uses our understanding of such interactions to convey a rich, imaginary, 3D world using only their body motion. Can we train a computer to take human motion and, similarly, conjure the 3D scene in which it belongs? Such a method would have many applications in synthetic data generation, architecture, games, and virtual reality. For example, there exist large datasets of 3D human motion like AMASS [ 38] and such data rarely contains information about the 3D scene in which it was captured. Could we take AMASS and generate plausible 3D scenes for all the motions? If so, we could use AMASS to generate training data containing realistic human-scene interaction. To answer such questions, we train a new method called MIME (Mining Interaction and Movement to infer 3D Envi- ronments) that generates plausible indoor 3D scenes based on 3D human motion. Why is this possible? The key in- tuitions are that (1) A human’s motion through free space indicates the lack of objects, effectively carving out regions of the scene that are free of furniture. And (2), when they are in contact with the scene, this constrains both the type and placement of 3D objects; e.g., a sitting human must be sitting on something, such as a chair, a sofa, a bed, etc. To make these intuitions concrete, we develop MIME, 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. 12965 which is a transformer-based auto-regressive 3D scene gen- eration method that, given an empty floor plan and a human motion sequence, predicts the furniture that is in contact with the human. It also predicts plausible objects that have no contact with the human but that fit with the other objects and respect the free-space constraints induced by the human motion. To condition the 3D scene generation with human motion, we estimate possible contact poses using POSA [ 23] and divide the motion in contact and non-contact snippets (Fig. 2). The non-contact poses define free-space in the room, which we encode as 2D floor maps, by projecting the foot vertices onto the ground plane. The contact poses and cor- responding 3D human body models are represented by 3D bounding boxes of the contact vertices predicted by POSA. We use this information as input to the transformer and auto- regressively predict the objects that fulfill the contact and free-space constraints; see Fig. 1. To train MIME, we built a new dataset called 3D-FRONT HUMAN that extends the large-scale synthetic scene dataset 3D-FRONT [ 18]. Specifically, we automatically populate the 3D scenes with humans; i.e., non-contact humans (a sequence of walking motion and standing humans) as well as contact humans (sitting, touching, and lying humans). To this end, we leverage motion sequences from AMASS [38], as well as static contact poses from RenderPeople [ 47] scans. At inference time, MIME generates a plausible 3D scene layout for the input motion, represented as 3D bounding boxes. Based on this layout, we select 3D models from the 3D-FUTURE dataset [ 19] and refine their 3D placement based on geometric constraints between the human poses and the scene. In comparison to pure 3D scene generation approaches like ATISS [ 46], our method generates a 3D scene that sup- ports human contact and motion while putting plausible ob- jects in free space. In contrast to Pose2Room [ 43] which is a recent pose-conditioned generative model, our method en- ables the generation of objects that are not in contact with the human, thus, predicting the entire scene instead of isolated objects. We demonstrate that our method can directly be applied to real captured motion sequences such as PROX-D [22] without finetuning . In summary, we make the following contributions: •a novel motion-conditioned generative model for 3D room scenes that auto-regressively generates objects that are in contact with the human and do not occupy free-space defined by the motion. •a new 3D scene dataset with interacting humans and free space humans which is constructed by populating 3D FRONT with static contact/standing poses from RenderPeople and motion data of AMASS. Figure 2. We divide input humans into two parts: contact humans and free-space humans. We extract the 3D bounding boxes for each contact human, and use non-maximum suppression on the 3D IoU to aggregate multiple humans in the same 3D space into a single contact 3D bounding box (orange boxes). We project the foot vertices of free-space humans on the floor plane, to get the 2D free-space mask (dark blue).
Zhang_Starting_From_Non-Parametric_Networks_for_3D_Point_Cloud_Analysis_CVPR_2023	Abstract We present a Non-parametric Network for 3D point cloud analysis, Point-NN , which consists of purely non- learnable components: farthest point sampling (FPS), k- nearest neighbors ( k-NN), and pooling operations, with trigonometric functions. Surprisingly, it performs well on various 3D tasks, requiring no parameters or training, and even surpasses existing fully trained models. Start- ing from this basic non-parametric model, we propose two extensions. First, Point-NN can serve as a base archi- tectural framework to construct Parametric Networks by simply inserting linear layers on top. Given the supe- rior non-parametric foundation, the derived Point-PN ex- hibits a high performance-efficiency trade-off with only a few learnable parameters. Second, Point-NN can be re- garded as a plug-and-play module for the already trained 3D models during inference. Point-NN captures the comple- mentary geometric knowledge and enhances existing meth- ods for different 3D benchmarks without re-training. We hope our work may cast a light on the community for un- derstanding 3D point clouds with non-parametric meth- ods. Code is available at https://github.com/ ZrrSkywalker/Point-NN .	1. Introduction Point cloud 3D data processing is a foundational opera- tion in autonomous driving [4, 12, 21], scene understand- ing [1, 3, 33, 44], and robotics [5, 20, 26]. Point clouds contain unordered points discretely depicting object sur- faces in 3D space. Unlike grid-based 2D images, they are distribution-irregular and permutation-invariant, which leads to non-trivial challenges for algorithm designs. Since PointNet++ [23], the prevailing trend has been †Corresponding author Non-ParametricEncoderPoint-MemoryBank+No Parameter Non-Parametric Components FPS k-NNPooling TrigonometricFunctions ClassificationFew-Shot Cls. Segmentation Detection81.8% Acc.90.9%Acc.70.4%mIoU33.3%AP!" NoTraining Figure 1. The Pipeline of Non-Parametric Networks. Point-NN is constructed by the basic non-parametric components without any learnable operators. Free from training, Point-NN can achieve favorable performance on various 3D tasks. adding advanced local operators and scaled-up learnable pa- rameters. Instead of max pooling for feature aggregation, mechanisms are proposed to extract local geometries, e.g., adaptive point convolutions [14, 16, 30, 37, 38] and graph- like message passing [11, 34, 43]. The performance gain also rises from scaling up the number of parameters, e.g., KPConv [30]’s 14.3M and PointMLP [17]’s 12.6M, is much larger than PointNet++’s 1.7M. This trend has increased network complexity and computational resources. Instead, the non-parametric framework underlying all the learnable modules remains nearly the same since Point- Net++, including farthest point sampling (FPS), k-Nearest Neighbors ( k-NN), and pooling operations. Given that few 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. 5344 works have investigated their efficacy, we ask the question: can we achieve high 3D point cloud analysis performance using only these non-parametric components? We present a Non-parametric Network, termed Point- NN, which is constructed by the aforementioned non- learnable components. Point-NN, as shown in Figure 1, consists of a non-parametric encoder for 3D feature ex- traction and a point-memory bank for task-specific recogni- tion. The multi-stage encoder applies FPS, k-NN, trigono- metric functions, and pooling operations to progressively aggregate local geometries, producing a high-dimensional global vector for the point cloud. We only adopt simple trigonometric functions to reveal local spatial patterns at every pooling stage without learnable operators. Then, we adopt the non-parametric encoder of Point-NN to extract the training-set features and cache them as the point-memory bank. For a test point cloud, the bank outputs task-specific predictions via naive feature similarity matching, which val- idates the encoder’s discrimination ability. Free from an
Yang_GD-MAE_Generative_Decoder_for_MAE_Pre-Training_on_LiDAR_Point_Clouds_CVPR_2023	Abstract Despite the tremendous progress of Masked Autoen- coders (MAE) in developing vision tasks such as image and video, exploring MAE in large-scale 3D point clouds re- mains challenging due to the inherent irregularity. In con- trast to previous 3D MAE frameworks, which either design a complex decoder to infer masked information from main- tained regions or adopt sophisticated masking strategies, we instead propose a much simpler paradigm. The core idea is to apply a Generative Decoder for MAE (GD-MAE) to automatically merges the surrounding context to restore the masked geometric knowledge in a hierarchical fusion manner. In doing so, our approach is free from introducing the heuristic design of decoders and enjoys the flexibility of exploring various masking strategies. The correspond- ing part costs less than 12% latency compared with con- ventional methods, while achieving better performance. We demonstrate the efficacy of the proposed method on several large-scale benchmarks: Waymo, KITTI, and ONCE. Con- sistent improvement on downstream detection tasks illus- trates strong robustness and generalization capability. Not only our method reveals state-of-the-art results, but remark- ably, we achieve comparable accuracy even with 20% of the labeled data on the Waymo dataset. Code will be released.	1. Introduction We have witnessed great success in 3D object detec- tion [44, 47, 64, 68, 71, 78], due to the numerous applica- tions in autonomous driving, robotics, and navigation. De- spite the impressive performance, most methods count on large amounts of carefully labeled 3D data, which is often ∗Equal contribution. This work was done when Honghui was an intern at Shanghai Artificial Intelligence Laboratory. †Corresponding author … … …Head … … …HeadTransformer Encoder Transformer Decoder (a) MAE, Point-MAE, etc. Transformer EncoderTransformer Decoder (b) ConvMAE, Point-M2AE, etc. …Head Transformer Encoder (c) GD-MAE (Ours)Generative DecoderFigure 1. Comparisons. Previous MAE-style pre-training archi- tectures of (a) single-scale [18, 19, 38] and (b) multi-scale [12, 73] take as inputs the visible tokens and learnable tokens for decoders. In contrast, (c) the proposed framework avoids such a process. of high cost and time-consuming. Such a fully supervised manner hinders the possibility of using massive unlabeled data and can be vulnerable when applied in different scenes. Mask Autoencoder (MAE) [18], serving as one of the ef- fective ways for pre-training, has demonstrated great po- tential in learning holistic representations. This is achieved by encouraging the method to learn a semantically consis- tent understanding of the input beyond low-level statistics. Although MAE-based methods have shown effectiveness in 2D image [18] and video [52], how to apply it in large-scale point clouds remains an open problem. Due to the large variation of the visible extent of ob- jects, learning hierarchical representation is of great signif- 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. 9403 icance in 3D supervised learning [40, 46, 62]. To enable MAE-style pre-training on the hierarchical structure, previ- ous approaches [12, 73] introduce either complex decoders or elaborate masking strategies to learn robust latent repre- sentations. For example, ConvMAE [12] adopts a block- wise masking strategy that first obtains a mask for the late stage of the encoder and then progressively upsamples the mask to larger resolutions in early stages to maintain mask- ing consistency. Point-M2AE [73] proposes a hierarchi- cal decoder to gradually incorporate low-level features into learnable tokens for reconstruction. Meanwhile, it needs a multi-scale masking strategy that backtracks unmasked po- sitions to all preceding scales to ensure coherent visible re- gions and avoid information leakage. The minimum size of masking granularity is highly correlated to output tokens of the last stage, which inevitably poses new challenges, espe- cially to objects with small sizes, e.g., pedestrians. To alleviate the issue, we present a much simpler paradigm dubbed GD-MAE for pre-training, as shown in Figure 1. The key is to use a generative decoder to automat- ically expand the visible regions to the underlying masked area. In doing so, it eliminates the need for designing complex decoders, in which masked regions are presented as learnable tokens. It also allows for the unification of multi-scale features into the same scale, thus enabling flex- ible masking strategies, e.g., point- and patch-wise mask- ing, while avoiding intricate operations such as backtrack- ing in [12, 73] to keep masking consistency. Specifically, it consists of the following components: Firstly, we propose the Sparse Pyramid Transformer (SPT) as the multi-scale encoder. Following [9,22,43], SPT takes pillars as input due to the compact and regular repre- sentation. Unlike PointPillars [22] that uses traditional con- volutions for feature extraction, we use the sparse convo- lution [62] to downsample the tokens and the sparse trans- former [9] to enlarge the receptive field of the visible tokens when deploying extensive masking. Secondly, we introduce the Generative Decoder (GD) to simplify MAE-style pre-training on multi-scale backbones. GD consists of a series of transposed convolutions used to upsample multi-scale features and a convolution utilized to expand the visible area, as shown in Figure 2. The expanded features are then directly indexed according to the coordi- nates of the masked tokens for the geometric reconstruction. Extensive experiments have been conducted on Waymo Open Dataset [49], KITTI [13], and ONCE [33] to ver- ify the efficacy. On the Waymo dataset, GD-MAE sets new state-of-the-art detection results compared to previ- ously published methods. Our contributions are summarized as follows: • We introduce a simpler MAE framework that avoids complex decoders and thus simplifies pre-training. … Masking Multi-Scale Encoder DecoderFigure 2. Illustration of area expansion. The input point cloud (i.e., the orange curve) is voxelized and fed into the multi-scale encoder. The generative decoder can automatically expand visible features to potentially masked areas. • The proposed decoder enables flexible masking strate- gies on LiDAR point clouds, while costing less than 12% latency compared with conventional methods. • Extensive experiments are conducted to verify the ef- fectiveness of the proposed model.
Zhang_Layout-Based_Causal_Inference_for_Object_Navigation_CVPR_2023	Abstract Previous works for ObjectNav task attempt to learn the association (e.g. relation graph) between the visual inputs and the goal during training. Such association contains the prior knowledge of navigating in training environments, which is denoted as the experience. The experience per- forms a positive effect on helping the agent infer the likely location of the goal when the layout gap between the un- seen environments of the test and the prior knowledge ob- tained in training is minor. However, when the layout gap is significant, the experience exerts a negative effect on nav- igation. Motivated by keeping the positive effect and re- moving the negative effect of the experience, we propose the layout-based soft Total Direct Effect (L-sTDE) frame- work based on the causal inference to adjust the predic- tion of the navigation policy. In particular, we propose to calculate the layout gap which is defined as the KL diver- gence between the posterior and the prior distribution of the object layout. Then the sTDE is proposed to appropri- ately control the effect of the experience based on the lay- out gap. Experimental results on AI2THOR, RoboTHOR, and Habitat demonstrate the effectiveness of our method. The code is available at https://github.com/sx- zhang/Layout-based-sTDE.git .	1. Introduction The visual object-oriented navigation task (i.e. Object- Nav) [3] requires the agent to navigate to a user-specified goal (e.g. laptop) based on the egocentric visual observa- tions. A typical challenge is navigating in unseen environ- ments, where the goal is invisible most of the time, i.e. the partial observable problem, which typically results in the agent’s meaningless actions (e.g. back-tracking or getting lost at dead-ends). Although encouraging the exploration in the unseen environment (until the goal is visible) is an in- tuitive solution, the lack of environment layout information still limits the efficiency of goal-oriented navigation. : ObservationS(a) The factual prediction(b) The counterfactual prediction : GoalG : ActionA : ExperienceZZGS¯S¯GZ¯ASGZAInterventionInterventionCounterfactualFigure 1. The proposed causal graph. (a) represents the fact pre- diction a, i.e. the original prediction of the trained model. (b) refers to the counter-fact prediction ¯a, i.e. the prediction is only affected by the experience Z. (b) is realized by applying the inter- vention and counterfactual operations to the original model. Recently, the learning-based methods attempt to model the prior knowledge of the spatial relationships among the objects, so the agent could infer the likely locations of the goal based on the current observation (which objects are observed currently) and the prior knowledge (the spatial relationships between the goal and the observed objects) learned in the training stage. Some works utilize additional modules to construct the objects graph [15, 59, 60], the re- gion graph [63] and the attention mechanism [32], while others [16, 56] employ a network that implicitly learns the spatial relationships end-to-end. All these methods attempt to establish prior knowledge in training environments, so that the agent would utilize the prior knowledge to associate the real-time observations with the goal, and infer the likely locations of the goal during the inference. The underlying assumption of these methods is that all of the object lay- outs in unseen environments should be exactly consistent with those in training environments. However, the layout consistency assumption is only partially correct due to the limited training data. Thus, those methods typically suffer from poor generalization [31] in unseen environments. To reveal the causes of poor generalization, we propose to use the casual graph (i.e. Structural Causal Model, SCM 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. 10792 [38]) to analyze these navigation works. As illustrated in Fig. 1 (a), the navigation model takes the observation Sand the goal Gas the input, and predicts the action Aat each timestamp. The causal links S→ZandG→Zrepresent that the observation and the goal are embedded by the pre- constructed modules [15, 32, 59, 60, 63] or the pre-trained network [16, 56]. The embedding vector is defined as the experience Zin the causal graph, which introduces the prior knowledge to influence action prediction ( Z→A). Mean- while, the real-time observation and the goal also indepen- dently affect the prediction without being encoded by the prior knowledge module, which is represented as S→A andG→A, respectively. The causal links S→Aand G→Arepresent the exploration-based effect (only related to the current episode) on the action prediction, which is dif- ferent from the experience-based effect Z→A. Consider two cases of the layout gap between the current environment and the prior knowledge: 1) the layout gap is minor and 2) the layout gap is significant. In the former case, the object layout is consistent in the current environment and the prior knowledge. Thus, the experience Zexerts a positive effect on the prediction of action A. However, the effect of experi- enceZin the latter case could be negative. If the agent still relies on the “negative” experience to predict actions, it will suffer from poor generalization. Therefore, wisely utilizing the experience is essential to the ObjectNav task. Motivated by wisely utilizing the learned experience, we propose the soft Total Direct Effect (sTDE) framework based on the Total Direct Effect analysis in causal inference. Our sTDE improves the generalization of the trained model in inference by eliminating the prediction bias brought by the experience. To decouple the effect of experience, we construct the counter-fact prediction ¯a: the prediction is only affected by the experience Zwhile ignoring the Sand G, as shown in Fig. 1 (b). Then we propose the object layout estimator that calculates whether the effect of the ex- perience is positive, by measuring the layout gap between the current environment and the prior knowledge. Further- more, our sTDE will remove the counter-fact prediction ¯a from the fact prediction awhen the layout gap is large. In this paper, we propose the layout-based soft TDE framework for the ObjectNav task. Specifically, we adopt the Dirichlet-Multinomial distribution [22] to formulate the contextual relationship between objects, which represents the object layout of the environment. Before training, the agent learns prior layout distribution (i.e. the prior parame- ters of Dirichlet-Multinomial distribution) by randomly ex- ploring the training environments. In the training stage, based on the Bayesian inference, the agent estimates the posterior layout distribution with the prior distribution and newly obtained observations. Then the constantly updated posterior layout is encoded into the navigation model and utilized to learn the environment-adaptive experience. Theentire model is trained with RL by maximizing the reward of reaching the goal. In the test stage, our agent will not di- rectly use the trained policy as most previous works do. The agent first calculates the layout gap and the counter-fact pre- diction. The layout gap is determined by calculating the KL divergence between the posterior and prior distribution of object layouts and serves as a weight to determine whether to remove the counter-fact prediction (i.e. experience ef- fect) from the original prediction. The experimental results on AI2THOR [27], RoboTHOR [12] and Habitat [48] in- dicate that our layout-based sTDE (L-sTDE) can be a plug- and-play method to boost existing methods to achieve better navigation performances.
Ye_Partial_Network_Cloning_CVPR_2023	Abstract In this paper, we study a novel task that enables par- tial knowledge transfer from pre-trained models, which we term as Partial Network Cloning (PNC). Unlike prior meth- ods that update all or at least part of the parameters in the target network throughout the knowledge transfer process, PNC conducts partial parametric “cloning” from a source network and then injects the cloned module to the target, without modifying its parameters. Thanks to the transferred module, the target network is expected to gain additional functionality, such as inference on new classes; whenever needed, the cloned module can be readily removed from the target, with its original parameters and competence kept intact. Specifically, we introduce an innovative learning scheme that allows us to identify simultaneously the com- ponent to be cloned from the source and the position to be inserted within the target network, so as to ensure the opti- mal performance. Experimental results on several datasets demonstrate that, our method yields a significant improve- ment of 5%in accuracy and 50% in locality when com- pared with parameter-tuning based methods. Our code is available at https://github.com/JngwenYe/PNCloning.	1. Introduction With the recent advances in deep learning, an increas- ingly number of pre-trained models have been released online, demonstrating favourable performances on various computer vision applications. As such, many model-reuse approaches have been proposed to take advantage of the pre-trained models. In practical scenarios, users may re- quest to aggregate partial functionalities from multiple pre- trained networks, and customize a target network whose competence differs from any network in the model zoo. A straightforward solution to the functionality dynamic changing is to re-train the target network using the origi- nal training dataset, or to conduct finetuning together with regularization strategies to alleviate catastrophic forget- †Corresponding author. … 𝐼𝑛𝑠𝑒𝑟𝑡(4)Source Models 𝐿𝑜𝑐𝑎𝑙(4) Target ModelTransferableModules 𝐿𝑜𝑐𝑎𝑙(4)𝐿𝑜𝑐𝑎𝑙(4) … FunctionalAddition𝐼𝑛𝑠𝑒𝑟𝑡(4) 𝐼𝑛𝑠𝑒𝑟𝑡(4)… …Figure 1. Illustration of partial network cloning. Given a set of pre-trained source models, we “clone” the transferable modules from the source, and insert them into the target model (left) while preserving the functionality (right). ting [3,19,39], which is known as continual learning. How- ever, direct re-training is extremely inefficient, let alone the fact that original training dataset is often unavailable. Con- tinual learning, on the other hand, is prone to catastrophic forgetting especially when the amount of data for finetun- ing is small, which, unfortunately, often occurs in practice. Moreover, both strategies inevitably overwrite the original parameters of the target network, indicating that, without explicitly storing original parameters of the target network, there is no way to recover its original performance or com- petence when this becomes necessary. In this paper, we investigate a novel task, termed as Partial Network Cloning (PNC), to migrate knowledge from the source network, in the form of a transferable mod- ule, to the target one. Unlike prior methods that rely on updating parameters of the target network, PNC attempts to clone partial parameters from the source network and then directly inject the cloned module into the target, as shown in Fig. 1. In other words, the cloned module is transferred to the target in a copy-and-paste manner. Meanwhile, 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. 20137 original parameters of the target network remain intact, in- dicating that whenever necessary, the newly added module can be readily removed to fully recover its original function- ality. Notably, the cloned module per se is a fraction of the source network, and therefore requirements no additional storage expect for the lightweight adapters. Such flexibil- ity to expand the network functionality and to detach the cloned module without altering the base of the target or al- locating extra storage, in turn, greatly enhances the utility of pre-trained model zoo and largely enables plug-and-play model reassembly. Admittedly, the ambitious goal of PNC comes with sig- nificant challenges, mainly attributed to the black-box na- ture of the neural networks, alongside our intention to pre- serve the performances on both the previous and newly- added tasks of the target. The first challenge concerns the localization of the to-be-cloned module within the source network, since we seek discriminant representations and good transferability to the downstream target task. The sec- ond challenge, on the other hand, lies in how to inject the cloned module to ensure the performance. To solve these challenges, we introduce an innovative strategy for PNC, through learning the localization and in- sertion in an intertwined manner between the source and tar- get network. Specifically, to localize the transferable mod- ule in the source network, we adopt a local-performance- based pruning scheme for parameter selection. To adap- tively insert the module into the target network, we utilize a positional search method in the aim to achieve the optimal performance, which, in turn, optimizes the localization op- eration. The proposed PNC scheme achieves performances significantly superior to those of the continual learning set- ting ( 5%∼10%), while reducing data dependency to 30%. Our contributions are therefore summarized as follows. • We introduce a novel yet practical model re-use setup, termed as partial network cloning (PNC). In contrast to conventional settings the rely on updating all or part of the parameters in the target network, PNC migrates parameters from the source in a copy-and-paste man- ner to the target, while preserving original parameters of the target unchanged. • We propose an effective scheme towards solving PNC, which conducts learnable localization and insertion of the transferable module jointly between the source and target network. The two operations reinforce each other and together ensure the performance of the tar- get network. • We conduct experiments on four widely-used datasets and showcase that the proposed method consis- tently achieves results superior to the conventional knowledge-transfer settings, including continual learn- ing and model ensemble.
Yu_TOPLight_Lightweight_Neural_Networks_With_Task-Oriented_Pretraining_for_Visible-Infrared_Recognition_CVPR_2023	Abstract Visible-infrared recognition (VI recognition) is a chal- lenging task due to the enormous visual difference across heterogeneous images. Most existing works achieve promis- ing results by transfer learning, such as pretraining on the ImageNet, based on advanced neural architectures like ResNet and ViT. However, such methods ignore the neg- ative influence of the pretrained colour prior knowledge, as well as their heavy computational burden makes them hard to deploy in actual scenarios with limited resources. In this paper, we propose a novel task-oriented pretrained lightweight neural network (TOPLight) for VI recognition. Specifically, the TOPLight method simulates the domain conflict and sample variations with the proposed fake do- main loss in the pretraining stage, which guides the network to learn how to handle those difficulties, such that a more general modality-shared feature representation is learned for the heterogeneous images. Moreover, an effective fine- grained dependency reconstruction module (FDR) is devel- oped to discover substantial pattern dependencies shared in two modalities. Extensive experiments on VI person re- identification and VI face recognition datasets demonstrate the superiority of the proposed TOPLight, which signifi- cantly outperforms the current state of the arts while de- manding fewer computational resources.	1. Introduction Identity recognition technologies have provided numer- ous reliable solutions for monitoring systems, which strive to match the face (face recognition [6, 7]) or pedestrian (person re-identification [42]) images of the same identity. However, the majority of previous efforts only consider vis- ible images. In real-life practice, many surveillance cam- eras can switch to infrared imaging mode at night. Thus, the essential cross-modality visible-infrared recognition (VI *Corresponding Author (Email: [email protected]) : mAP55.157.6ImageNet -mini + Generic variation ImageNet-1k pretrained networksDual-path trainingLearn to represent ''heterogenous'' features Learn to embed ''heterogenous'' featuresDomain Conflict SP TOP MobileNetV3-L(a) TOP ShuffleNetV2-1.5TOP GhostNet-1.347.862.4 47.360.1 50.961.8 47.159.0 47.8 47.9 (b)+ Colour variation + T exture variation : Rank-1 SP SPFigure 1. (a) The task-oriented pretraining strategy; (b) Per- formance comparison of the standard ImageNet-1k pretraining scheme (SP) and the proposed task-oriented pretraining scheme (TOP) on the SYSU-MM01 dataset [35] (All-Search mode). recognition) technology has been developed to match the visible and infrared photographs of the same people. Recently, visible-infrared person re-identification (VI- ReID) [26, 38] and visible-infrared face recognition [8, 13, 44] have been widely studied. The key issue is identi- fying the modality-shared patterns. To this end, several works [29, 30] use generative adversarial networks (GANs) to implement cross-modality alignment at the pixel and fea- ture levels. Others [4, 26, 38] design the dual-path feature extraction network, coupled with inter-feature constraints, to close the embedding space of two modalities. However, these methods utilize at least one pretrained ResNet-50 [12] backbone to extract solid features, which makes them un- suitable for edge monitoring devices. Recent works [4, 38] employ auxiliary models (e.g., pose estimation, graph rea- soning) to relieve the modality discrepancy, which enhances the performance on academic benchmarks but reduces the real-time inference speed. Compared with conventional deep networks (e.g., ResNet, ViT), lightweight networks [11,15,24] can extract basal features rapidly. In VI recogni- 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. 3541 tion tasks, however, the vast modality discrepancy renders the performance of lightweight networks significantly infe- rior to that of conventional deep networks. The main rea- son is that lightweight networks lack the ability to identify modality-shared patterns from heterogeneous images. To address this issue, we present an effective task- oriented pretraining (TOP) strategy. As shown in Fig. 1(a), we first train a lightweight network on the ImageNet-1k dataset to learn vision prior knowledge. After that, the trained network is transformed into the dual-path network and further trained by using task-oriented data augmenta- tions, identity consistency loss and fake domain loss on the ImageNet-mini dataset [18]. The task-oriented pretrain- ing (TOP) strategy simulates the sample differences in VI scenes and teaches the network how to represent and em- bed discrepant features. Fig. 1(b) reports the performance of three lightweight networks in the VI-ReID task. Com- pared with the ImageNet-1k pretraining, our TOP strategy can remarkably improve the baseline performance. Another weakness of lightweight networks is that few feature maps are learned from raw images for rapid infer- ence. In the VI recognition scene, it is challenging to dis- cover modality-shared patterns with so few learned feature maps. In practice, the network can focus on a group of ag- gregated modality-specific patterns that offer the most gra- dient for identity classification. In contrast, the fine-grained and modality-shared patterns, which are crucial for achiev- ing robust cross-modality matching, are neglected. Based on the above observations, we present a novel fine-grained dependency reconstruction (FDR) module to help lightweight networks learn modality-shared and fine- grained patterns. Specifically, inspired by the horizontal slice scheme [1], we first slice feature maps horizontally and vertically to extract fine-grained patterns from diver- sified local regions. Then, the original spatial dependen- cies of these patterns are eliminated by using pooling oper- ations. Further, the cross-modality dependencies are built by using up-sampling layers to amplify the modality-shared parts from these patterns. At last, to avoid overfitting, the shuffle attention is designed to re-weight the channel depen- dencies of all the feature maps, which spreads attention to local patterns as much as possible. In general, the major contributions of this paper can be summarized as follows. • We propose an effective task-oriented pretrained lightweight neural network (TOPLight) for VI recog- nition. To the best of our knowledge, it is the first work to develop a paradigm for VI recognition on edge de- vices with an extremely low computation budget. • An effective task-oriented pretraining strategy is pro- posed to enhance the heterogeneous feature learning capacity of lightweight networks with task-oriented augmentations and the proposed fake domain loss.• A fine-grained dependency reconstruction module is designed to mine cross-modality dependencies. • Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art methods on mainstream VI-ReID and VI face recognition datasets by a remarkable margin and extremely low complexity.
Yu_Adaptive_Spot-Guided_Transformer_for_Consistent_Local_Feature_Matching_CVPR_2023	Abstract Local feature matching aims at finding correspondences between a pair of images. Although current detector-free methods leverage Transformer architecture to obtain an im- pressive performance, few works consider maintaining lo- cal consistency. Meanwhile, most methods struggle with large scale variations. To deal with the above issues, we propose Adaptive Spot-Guided Transformer (ASTR) for lo- cal feature matching, which jointly models the local consis- tency and scale variations in a unified coarse-to-fine archi- tecture. The proposed ASTR enjoys several merits. First, we design a spot-guided aggregation module to avoid in- terfering with irrelevant areas during feature aggregation. Second, we design an adaptive scaling module to adjust the size of grids according to the calculated depth information at fine stage. Extensive experimental results on five stan- dard benchmarks demonstrate that our ASTR performs fa- vorably against state-of-the-art methods. Our code will be released on https://astr2023.github.io . *Equal Contribution †Corresponding Author	1. Introduction Local feature matching (LFM) is a fundamental task in computer vision, which aims to establish correspondence for local features across image pairs. As a basis for many 3D vision tasks, local feature matching can be applied in Structure-from-Motion (SfM) [49], 3D reconstruction [13], visual localization [48, 51], and pose estimation [18, 41]. Because of its broad applications, local feature matching has attracted substantial attention and facilitated the devel- opment of many researches [14, 27, 42, 44, 50]. However, finding consistent and accurate matches is still difficult due to various challenging factors such as illumination varia- tions, scale changes, poor textures, and repetitive patterns. To deal with the above challenges, numerous matching methods have been proposed, which can be generally cat- egorized into two major groups, including detector-based matching methods [2, 14, 15, 39, 42, 47] and detector-free matching methods [9, 23, 27, 43, 44, 50]. Detector-based matching methods require to first design a keypoint de- tector to extract the keypoints between two images, and then establish matches between these extracted keypoints. The quality of detected keypoints will significantly af- fect the performance of detector-based matching methods. Therefore, many works aim to improve keypoint detection through multi-scale detection [36], repeatable and reliable 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. 21898 verification [42]. Thanks to the high-quality keypoints de- tected, these methods can achieve satisfactory performance while maintaining high computational and memory effi- ciency. However, these detector-based matching methods may have difficulty in finding reliable matches in textureless areas, where keypoints are challenging to detect. Differ- ently, detector-free matching methods do not need to detect keypoints and try to establish pixel-level matches between local features. In this way, it is possible to establish matches in the texture-less areas. Due to the power of attention in capturing long-distance dependencies, many Transformer- based methods [9, 50, 52, 57] have emerged in recent years. As a representative work, considering the computation and memory costs, LoFTR [50] applies Linear Transformer [25] to aggregate global features at the coarse stage and then crops fixed-size grids for further refinement. To alleviate the problem caused by scale changes, COTR [24] calculate the co-visible area iteratively through attention mechanism. The promising performance of Transformer-based methods proves that attention mechanism is effective on local feature matching. Nevertheless, some recent works [28, 60] indi- cate Transformer lacks spatial inductive bias for continuous dense prediction tasks, which may cause inconsistent local matching results. By studying the previous matching methods, we sum up two issues that are imperative for obtaining the dense cor- respondence between images. (1) How to maintain local consistency. The correct matching result usually satisfies the local matching consistency, i.e., for two similar adja- cent pixels, their matching points are also extremely close to each other. Existing methods [24,50,57] utilize global at- tention in feature aggregation, introducing many irrelevant regions that affect feature updates. Some pixels are dis- turbed by noisy or similar areas and aggregate information from wrong regions, leading to false matching results. As shown in Figure 1 (b), for two adjacent similar pixels, high- lighted regions of global linear attention are decentralized and inconsistent with each other. The inconsistency is also present in vanilla attention (see Figure 1 (c)). Therefore, it is necessary to utilize local consistency to focus the attention area on the correct place. (2) How to handle scale vari- ation. In a coarse-to-fine architecture, since the attention mechanism at the coarse stage is not sensitive to scale vari- ations, we should focus on the fine stage. Previous meth- ods [9, 27, 50, 57] select fixed-size grids for matching at the fine stage. However, when the scale varies too much across images, the correct match point may be out of the range of the grid, resulting in matching failure. Hence, the scheme of cropping grids should be adaptively adjusted according to scale variation across views. To deal with the above issues, we propose a novel Adap- tive Spot-guided Transformer (ASTR) for consistent local feature matching, including a spot-guided aggregation mod-ule and an adaptive scaling module. In the spot-guided ag- gregation module , towards the goal of maintaining local consistency, we design a novel attention mechanism called spot-guided attention: each point is guided by similar high- confidence points around it, focusing on a local candidate region at each layer. Here, we also adopt global features to enhance the matching ability of the network in the can- didate regions. Specifically, for any point p, we pick the points with high feature similarity and matching confidence in the local area. Their corresponding matching regions are used for the next attention of point p. In addition, global features are applied to help the network to make judgments. The coarse feature maps are iteratively updated in the above way. With our spot-guided aggregation module, the red and green pixels are guided to the correct area, avoiding the in- terference of repetitive patterns (see Figure 1 (d)). In Fig- ure 1 (e), our ASTR produces more accurate matching re- sults, which maintains local matching consistency. In the adaptive scaling module , to fully account of possible scale variations, we attempt to adaptively crop different sizes of grids for alignment. In detail, we compute the correspond- ing depth map using the coarse matching result and leverage the depth information to crop adaptive size grids from the high-resolution feature maps for fine matching. The contributions of our method could be summarized into three-fold: (1) We propose a novel Adaptive Spot- guided Transformer (ASTR) for local feature matching, in- cluding a spot-guided aggregation module and an adap- tive scaling module. (2) We design a spot-guided aggre- gation module that can maintain local consistency and be unaffected by irrelevant regions while aggregating features. Our adaptive scaling module is able to leverage depth in- formation to adaptively crop different size grids for refine- ment. (3) Extensive experimental results on five challeng- ing benchmarks show that our proposed method performs favorably against state-of-the-art image matching methods.
Zeng_Learning_Transferable_Spatiotemporal_Representations_From_Natural_Script_Knowledge_CVPR_2023	Abstract Pre-training on large-scale video data has become a common recipe for learning transferable spatiotemporal representations in recent years. Despite some progress, ex- isting methods are mostly limited to highly curated datasets (e.g., K400) and exhibit unsatisfactory out-of-the-box rep- resentations. We argue that it is due to the fact that they only capture pixel-level knowledge rather than spatiotem- poral semantics, which hinders further progress in video understanding. Inspired by the great success of image- text pre-training ( e.g., CLIP), we take the first step to ex- ploit language semantics to boost transferable spatiotem- poral representation learning. We introduce a new pre- text task, Turning to Video for Transcript Sorting (TVTS), which sorts shuffled ASR scripts by attending to learned video representations. We do not rely on descriptive cap- tions and learn purely from video, i.e., leveraging the natu- ral transcribed speech knowledge to provide noisy but use- ful semantics over time. Our method enforces the vision model to contextualize what is happening over time so that it can re-organize the narrative transcripts, and can seam- lessly apply to large-scale uncurated video data in the real world. Our method demonstrates strong out-of-the-box spa- tiotemporal representations on diverse benchmarks, e.g., +13.6% gains over VideoMAE on SSV2 via linear prob- ing. The code is available at https://github.com/ TencentARC/TVTS .	1. Introduction The aspiration of representation learning is to encode general-purpose representations that transfer well to di- verse downstream tasks, where self-supervised methodolo- gies [9, 25] dominate due to their advantage in exploitinglarge-scale unlabeled data. Despite significant progress in learning representations of still images [23, 43], the real world is dynamic and requires reasoning over time. In this paper, we focus on out-of-the-box spatiotemporal represen- tation learning , a more challenging but practical task to- wards generic video understanding, which aims to capture hidden representations that can be further used to conduct reasoning on broader tasks, e.g., classification and retrieval. There have been various attempts at self-supervised pre- training on video data from discriminative learning ob- jectives [5, 8, 27] to generative ones [17, 49], where the core is context capturing in spatial and temporal dimen- sions. Though promising results are achieved when trans- ferring the pre-trained models to downstream video recog- nition [22, 33, 48] via fine-tuning, the learned representa- tions are still far away from out-of-the-box given the poor linearly probing results (see Figure 1(a)). Moreover, exist- ing works mostly develop video models on the highly cu- rated dataset with particular biases, i.e., K400 [31]. Their applicability in the real world is questioned given the ob- served performance drops when training on a larger but uncurated dataset, YT-Temporal [55]. We argue that, to address the above issue, the rich spatiotemporal semantics contained in the video itself should be fully exploited. But current video models generally exploit visual-only percep- tion ( e.g., pixels) without explicit semantics. Recently, the success of CLIP [43] has inspired the com- munity to learn semantically aware image representations that are better transferable to downstream tasks and scal- able to larger uncurated datasets. It provides a feasible so- lution for improving spatiotemporal representation learning but remains two key problems. (1) The vision-language contrastive constraints in CLIP mainly encourage the un- derstanding of static objects (noun contrast) and simple mo- tions (verb contrast), while how to enable long-range tem- 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. 23079 17.9 SOTA (K400) Ours (YT) 31.515.9 SOTA (YT)+15.6%+13.6% SSV2N/A SOTA (K400) Ours (YT) 55.120.4 SOTA (YT)+34.7% Kine?cs-400 52.7 SOTA (K400) Ours (YT) 83.444.4 SOTA (YT)+39.0%+30.7% UCF-10130.9 SOTA (K400) Ours (YT) 58.423.9 SOTA (YT)+34.5%+27.5% HMDB-51 (a) Linear Probe Accuracy (%) What’s the order of transcripts? Reasoning from the video! 2⃣ I can’t ride… 1⃣ I take a picture…Out-of-order ASR Transcripts (b) Our Mo? va?on ⋯Figure 1. (a) We evaluate the transferability of spatiotemporal representations via linear probing on four video recognition datasets [22, 31, 33, 48], where the state-of-the-art method [49] underperforms. It performs even worse when pre-trained with a large-scale uncurated dataset, YT-Temporal [55]. (b) We encourage complex temporal understanding and advanced spatiotemporal representation learning with a new pretext task of sorting transcripts. poral understanding with language supervision needs to be studied. (2) The quality of language supervision [47] is crit- ical to the final performance of CLIP, however, it is hard to collect large-scale video data with literal captions that care- fully describe the dynamic content over time. The ideal way for self-supervised learning is to learn useful knowledge purely from the data itself, which is also the philosophy followed by previous video pre-training methods [17, 49]. Fortunately, video data is naturally multi-modal with tran- scribed speech knowledge in the form of text (ASR), pro- viding time-dependent semantics despite some noise. To facilitate spatiotemporal understanding in large-scale uncurated data under the supervision of inherent script knowledge, we introduce a new pretext task for video pre- training, namely, Turning to Video for Transcript Sorting (TVTS). Intuitively, people sort out the order of events by temporal reasoning. As illustrated in Figure 1(b), given several unordered transcripts, it is difficult to reorganize the narrative by merely understanding the literal semantics. When the corresponding video is provided, it will be much easier to sort the transcripts by contextualizing what is hap- pening over time. Whereas in neural networks, the tem- poral inference is embedded in spatiotemporal representa- tions. Thus we believe that if the chronological order of transcripts can be correctly figured out via resorting to the correlated video representations, the video has been well understood. We realize the pretext task of TVTS by performing joint attention among the encoded video spatiotemporal repre- sentations and the extracted ASR transcript representations. Specifically, given an input video and its successive tran- scripts, we randomly shuffle the order of the sentences.Subsequently, we concatenate the encoded script repre- sentations and the video representations and perform self- attention to predict the actual orders of the shuffled tran- scripts by fully understanding the spatiotemporal seman- tics in the video. The order prediction is cast as a K-way classification task, where Kis the number of transcripts. The pretext task indirectly regularizes our model to prop- erly capture contextualized spatiotemporal representations to provide enough knowledge for transcript ordering. The usage of language supervision is related to video- text alignment [4, 20] and multimodal representation learn- ing [18, 55] methods, however, we are completely differ- ent. (1) Video-text alignment methods focus on retrieval tasks and are devoted to associating the vision patterns with language concepts. They are generally single-frame bi- ased [34] and fail to encode strong out-of-the-box tempo- ral representations. (2) Multimodal representation learning methods aim to learn fused representations across modali- ties rather than vision-only spatiotemporal representations in our work. Moreover, different from our pretext task that aims to optimize spatiotemporal video representations, [55] sorts video frames by taking the features of individual frames as inputs without temporal modeling, i.e., learning video representations only at the image level. As [55] points out, its ordering pretext task is not critical for downstream tasks (performance even drops) and primarily serves as an interface to query the model about temporal events. To summarize, our contributions are three-fold. ( i) We exploit the rich semantics from script knowledge which is naturally along with the video, rendering a flexible pre- training method that can easily apply to uncurated video data in the real world. ( ii) We introduce a novel pre- 23080 text task for video pre-training, namely, Turning to Video for Transcript Sorting (TVTS). It promotes the capability of the model in learning transferable spatiotemporal video representations. ( iii) We conduct comprehensive compar- isons with advanced methods. Our pre-trained model ex- hibits strong out-of-the-box spatiotemporal representations on downstream action recognition tasks, especially the rel- atively large-scale and the most challenging SSV2 [22]. We also achieve state-of-the-art performances on eight common video datasets in terms of fine-tuning.
Zhang_Boosting_Video_Object_Segmentation_via_Space-Time_Correspondence_Learning_CVPR_2023	Abstract Current top-leading solutions for video object segmen- tation (VOS) typically follow a matching-based regime: for each query frame, the segmentation mask is inferred accor- ding to its correspondence to previously processed and the ﬁrst annotated frames. They simply exploit the supervisory signals from the groundtruth masks for learning mask pre- diction only, without posing any constraint on the space-time correspondence matching, which, however, is the fundamen- tal building block of such regime. To alleviate this crucial yet commonly ignored issue, we devise a correspondence-aware training framework, which boosts matching-based VOS so- lutions by explicitly encouraging robust correspondence ma- tching during network learning. Through comprehensively exploring the intrinsic coherence in videos on pixel and ob- ject levels, our algorithm reinforces the standard, fully su- pervised training of mask segmentation with label-free, con- trastive correspondence learning. Without neither requiring extra annotation cost during training, nor causing speed de- lay during deployment, nor incurring architectural modiﬁ- cation, our algorithm provides solid performance gains on four widely used benchmarks, i.e., DAVIS2016&2017, and YouTube-VOS2018&2019, on the top of famous matching- based VOS solutions.	1. Introduction In this work, we address the task of (one-shot) video ob- jectsegmentation(VOS)[5,73,96].Givenaninputvideowith groundtruth object masks in the ﬁrst frame, VOS aims at ac- curately segmenting the annotated objects in the subsequent frames. As one of the most challenging tasks in computer vision, VOS beneﬁts a wide range of applications including augmented reality and interactive video editing [72]. Modern VOS solutions are built upon fully supervised deep learning techniques and the top-performing ones [10, 12] largely follow a matching-based paradigm, where the object masks for a new coming frame ( i.e., query frame) are The ﬁrst two authors contribute equally to this work. yCorresponding author. [10] Figure 1. (a-b) shows some correspondences between a reference frame and a query frame. (c) gives mask prediction. XMem [10], even a top-leading matching-based VOS solution, still suffers from unreliable correspondence. In contrast, with our correspondence- aware training strategy, robust space-time correspondence can be established, hence leading to better mask-tracking results. generated according to the correlations between the query frame and the previously segmented as well as ﬁrst anno- tated frames ( i.e., reference frames), which are stored in an outside memory. It is thus apparent that the module for cross- frame matching ( i.e., space-time correspondence modeling) plays the central role in these advanced VOS systems. Nev- ertheless, these matching-based solutions are simply trained under the direct supervision of the groundtruth segmenta- tion masks. In other words, during training, the whole VOS system is purely optimized towards accurate segmentation mask prediction, yet without taking into account any ex- plicit constraint/regularization on the central component — space-time correspondence matching. This comes with a le- gitimate concern for sub-optimal performance, since there is no any solid guarantee of truly establishing reliable cross- frame correspondence during network learning. Fig. 1(a) of- fers a visual evidence for this viewpoint. XMem [10], the latest state-of-the-art matching-based VOS solution, tends to struggle at discovering valid space-time correspondence; indeed,somebackgroundpixels/patchesareincorrectlyreco- gnized as highly correlated to the query foreground. The aforementioned discussions motivate us to propose a new, space-time correspondence-aware training framework which addresses the weakness of existing matching-based 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. 2246 VOS solutions in an elegant and targeted manner. The core idea is to empower the matching-based solutions with en- hanced robustness of correspondence matching, through mi- ning complementary yet freesupervisory signals from the inherent nature of space-time continuity of training video sequences. In more detail, we comprehensively investigate the coherence nature of videos on both pixel and object lev- els: i)pixel-level consistency : spatiotemporally proximate pixels/patches tend to be consistent; and ii)object-level co- herence : visual semantics of same object instances at differ- ent timesteps tend to retain unchanged. By accommodating these two properties to an unsupervised learning scheme, we give more explicit direction on the correspondence match- ing process, hence promoting the VOS model to learn dense discriminative and object-coherent visual representation for robust, matching-based mask tracking (see Fig. 1 (b-c)). It is worth mentioning that, beyond boosting the segmen- tation performance, our space-time correspondence-aware training framework enjoys several compelling facets. First , our algorithm supplements the standard, fully supervised training paradigm of matching-based VOS with self-training of space-time correspondence. As a result, it does not cause any extra annotation burden. Second , our algorithm is fully compatible with current popular matching-based VOS solu- tions [10, 12], without particular adaption to the segmenta- tion network architecture. This is because the learning of the correspondence matching only happens in the visual embed- ding space. Third , as a training framework, our algorithm does not produce additional computational budget to the ap- plied VOS models during the deployment phase. We make extensive experiments on various gold-standard VOS datasets, i.e., DA VIS2016&2017 [52], and YouTube- VOS2018&2019 [86]. We empirically prove that, on the top of recent matching-based VOS models, i.e., STCN [12] and XMem [10], our approach gains impressive results, surpass- ing all existing state-of-the-arts. Concretely, in multi-object scenarios, it improves STCN by 1.2%,2.3%, and 2.3%, and XMem by 1.5%,1.2%, and 1.1% on DA VIS2017 val,Youtube- VOS2018val, as well as Youtube-VOS2019 val, respectively, in terms ofJ&F. Besides, it respectively promotes STCN and XMem by 0.4% and 0.7% on single-object benchmark dataset DA VIS2016 val.
Yin_NeRFInvertor_High_Fidelity_NeRF-GAN_Inversion_for_Single-Shot_Real_Image_Animation_CVPR_2023	Abstract Nerf-based Generative models have shown impressive capacity in generating high-quality images with consistent 3D geometry. Despite successful synthesis of fake identity images randomly sampled from latent space, adopting these models for generating face images of real subjects is still a challenging task due to its so-called inversion issue. In this paper, we propose a universal method to surgically fine-tune these NeRF-GAN models in order to achieve high-fidelity animation of real subjects only by a single image. Given the optimized latent code for an out-of-domain real image, we employ 2D loss functions on the rendered image to re- duce the identity gap. Furthermore, our method leverages explicit and implicit 3D regularizations using the in-domain neighborhood samples around the optimized latent code to remove geometrical and visual artifacts. Our experiments confirm the effectiveness of our method in realistic, high- fidelity, and 3D consistent animation of real faces on multi- ple NeRF-GAN models across different datasets.	1. Introduction Animating a human with a novel view and expression sequence from a single image opens the door to a wide range of creative applications, such as talking head synthe- sis [22, 34], augmented and virtual reality (AR/VR) [19], image manipulation [24, 32, 45], as well as data augmenta- tion for training of deep models [25,42,43]. Early works of image animation mostly employed either 2D-based image generation models [14, 26, 31, 37], or 3D parametric mod- els [4, 11, 40, 41] ( e.g. 3DMM [6]), but they mostly suffer from artifacts, 3D inconsistencies or unrealistic visuals. Representing scenes as Neural Radiance Fields (NeRF) [23] has recently emerged as a breakthrough Project page: https : / / yuyin1 . github . io / NeRFInvertor_Homepage/ NovelViews&ExpressionsNovelExpressionsNovelViewsInputFigure 1. Image animation results of our method. NeRFInver- torachieves 3D-consistent and ID-preserving animation ( i.e. novel views and expressions) of real subjects given only a single image. approach for generating high-quality images of a scene in novel views. However, the original NeRF models [5,33,44] only synthesize images of a static scene and require exten- sive multi-view data for training, restricting its application to novel view synthesis from a single image. Several studies have shown more recent advances in NeRFs by extending it to generate multi-view face images with single-shot data even with controllable expressions [7, 8, 12, 27, 38, 46]. These Nerf-based Generative models (NeRF-GANs) are able to embed attributes of training samples into their latent 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. 8539 variables, and synthesize new identity face images with different expressions and poses by sampling from their latent space. While animatable synthesis of fake identity images is im- pressive, it is still challenging to generate 3D-consistent and identity-preserving images of real faces. Specifically, cur- rent Nerf-GANs have difficulties to accurately translate out- of-domain images into their latent space, and consequently change identity attributes and/or introduce artifacts when applied to most real-world images. In order to synthesize real faces, the conventional method applies optimization algorithms to invert the input image to a latent code in a smaller ( i.e.W) or an extended ( i.e.W+) NeRF-GAN la- tent space. However, they both either have ID-preserving or artifacts issues as shown in Figure 2. The Wspace inver- sion, in particular, generates realistic novel views and clean 3D geometries, but suffers from the identity gap between the real and synthesized images. In contrast, the W+space inversion well preserves the identity but commonly gener- ates an inaccurate 3D geometry, resulting in visual artifacts when exhibited from new viewpoints. Hence, it remains as a trade-off to have a 3D-consistent geometry or preserve identity attributes when inverting face images out of latent space distribution. In this paper, we present NeRFInvertor as a universal inversion method for NeRF-GAN models to achieve high- fidelity, 3D-consistent, and identity-preserving animation of real subjects given only a single image. Our method is ap- plicable to most of NeRF-GANs trained for a static ordy- namic scenes, and hence accomplishes synthesis of real im- ages with both novel views and novel expressions (see Fig- ure 1). Since the real images are mostly out of the domain of NeRF-GANs latent space, we surgically fine-tune their generator to enrich the latent space by leveraging the single input image without degrading the learned geometries. In particular, given an optimized latent code for the in- put image, we first use image space supervision to narrow the identity gap between the synthesized and input images. Without a doubt, the fine-tuned model can be overfitted on the input image and well reconstruct the input in the origi- nal view. However, fine-tuning with just image space super- vision produces erroneous 3D geometry due to the insuffi- cient geometry and content information in a single image, resulting in visual artifacts in novel views. To overcome this issue, we introduce regularizations using the surround- ing samples in the latent space, providing crucial guidance for the unobserved part in the image space. By sampling la- tent codes from the neighborhood of optimized latent vari- ables with different poses and expressions, we enforce a novel geometric constraint on the density outputs of fine- tuned and original pretrained generators. We also further add regularizations on the rendered images of neighborhood samples obtained from the fine-tuned and pretrained genera- NovelviewInput Rec. 𝒲inversion𝒲+inversionID:0.99ID:0.363Dshapes ArtifacesIDgap ArtifacesInaccurateAttributes:•Haircut•BeardIssuesFigure 2. Trade-off between ID-preserving and removing ar- tifacts. Optimizing latent variables of Nerf-GANs for synthesis of a real face leads to a trade-off between identity-preserving and geometrical and visual artifacts. Specifically, Wspace inversion results in clean geometry but identity gap between real and gener- ated images, and W+space inversion causes preserving of iden- tity attributes but inaccurate geometry and visual artifacts. tors. These regularizations help us to leverage the geometry and content information of those in-domain neighborhood samples around the input. Our experiments validate the ef- fectiveness of our method in realistic, high-fidelity, and 3D consistent animating of real face images. The main contributions of this paper are as follows: 1. We proposed a universal method for inverting NeRF- GANs to achieve 3D-consistent, high-fidelity, and identity-preserving animation of real subjects given only a single image. 2. We introduce a novel geometric constraint by leverag- ing density outputs of in-domain samples around the input to provide crucial guidance for the unobserved part in the 2D space. 3. We demonstrate the effusiveness of our method on multiple NeRF-GAN models across different datasets.
Zhang_PeakConv_Learning_Peak_Receptive_Field_for_Radar_Semantic_Segmentation_CVPR_2023	Abstract The modern machine learning-based technologies have shown considerable potential in automatic radar scene un- derstanding. Among these efforts, radar semantic segmen- tation (RSS) can provide more refined and detailed infor- mation including the moving objects and background clut- ters within the effective receptive field of the radar. Moti- vated by the success of convolutional networks in various visual computing tasks, these networks have also been in- troduced to solve RSS task. However, neither the regular convolution operation nor the modified ones are specific to interpret radar signals. The receptive fields of existing convolutions are defined by the object presentation in opti- cal signals, but these two signals have different perception mechanisms. In classic radar signal processing, the object signature is detected according to a local peak response, i.e., CFAR detection. Inspired by this idea, we redefine the receptive field of the convolution operation as the peak re- ceptive field (PRF) and propose the peak convolution oper- ation (PeakConv) to learn the object signatures in an end- to-end network. By incorporating the proposed PeakConv layers into the encoders, our RSS network can achieve bet- ter segmentation results compared with other SoTA meth- ods on a multi-view real-measured dataset collected from an FMCW radar. Our code for PeakConv is available at https://github.com/zlw9161/PKC .	1. Introduction Radar is a remote sensor, which usually uses modu- lated electromagnetic signals to detect the objects of interest through directional transmitting antennas in a specific effec- tive working field [22]. As an active detection device, radar is more robust to extreme weather ( e.g., haze, rain or snow) than other active detection device such as LiDARs [2], and it is also not susceptible to dim light condition and sun glare, *Equal contribution.†Corresponding author. This research is sup- ported by Young Science Foundation of National Natural Science Foun- dation of China (No.62206258).as the passive optical sensors are [19]. In addition to the real-world location information, it can also tell the velocity of the moving objects thanks to the Doppler effects. Due to these advantages, radar sensors have played an irreplaceable role for many automotive security and defense applications, e.g., autonomous safety driving or UA V early warning. Conventional radar detection mostly relies on the peak detection algorithm following constant false alarm rate (CFAR) [22, 23] principle. Taking frequency modulated continuous wave (FMCW) radar as example, the raw radar echos are first converted as multi-domain united frequency representations, e.g., range-Doppler (RD) and range-angle (RA) maps, through a series of cascading fast Fourier trans- formations (FFTs). Then for each cell under test (CUT) in the input RD/RA map, the CFAR detector will determine whether it contains moving object information according to an estimated detection threshold, which fully considers the characteristics of the radar signal itself. However, to ob- tain good effect in practical application, it is necessary to manually fine-tune various hyper-parameters including the thresholding factor, sizes and shapes of the local scope ( i.e., the bandwidths of reference and guard units). Beyond that, conventional radar detection cannot give category informa- tion of the object. These two inconveniences hinder the con- ventional detection method from automatic semantic radar scene understanding. Encouraged by the success of modern deep learning techniques in computational perception, especially the ob- ject detection [8, 15, 20, 21, 29] and semantic segmenta- tion [5, 11, 16, 24, 28] in computer vision, some efforts had been made recently for better automatic radar scene inter- pretation. These efforts evolve the target-clutter binary hy- pothesis of conventional radar testing into target semantic characterization of modern machine learning, i.e., radar ob- ject detection (ROD) [10, 17, 27] and radar semantic seg- mentation (RSS) [3, 13, 18]. Most of these methods used convolution networks as backbone models, which take radar frequency representations as input, and then make predic- tions on RA or RD view or both two views. For example, a multi-view RSS (MVRSS) network [18] was proposed 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. 17577 Figure 1. An examplar illustration of moving object signa- tures/presentations in (a) the 2D RD map and the corresponding (b) RD-amplitude 3D representation of radar signals, and their (c) synchronized camera image. to take better advantage of radar localization capability by making “unit-wise” predictions on both RD and RA fre- quency domains. To support the sufficient training of these deep models, a few large-scale radar datasets were also col- lected and created, e.g., OxfordRobotCar [9], nuScenes [4], CRUW [26] and CARRADA [19]. However, the electromagnetic object signatures received by radar are not as intuitively understood as the optical ones captured by the cameras as shown in Fig. 1. With rich tex- ture and color information in the image, the convolution op- eration can learn useful semantic information from a rectan- gular local spatial receptive field (RF). And by introducing some intuitive priors of human vision, more efficient learn- ing mechanisms for convolution had been proposed, e.g., multi-scale fusion [12,15,25], dilation [5,28] and deforma- tion [8, 29]. So far, these mechanisms are also introduced into radar data processing, such as the inception or pyramid pooling for multi-scale information, atrous convolution for larger dilated RF and deformable convolution for irregular object signature in ROD-Net [27] and MVRSS [18]. De- spite the multi-scale mechanism, which is more of a modu- lar idea, i.e., the computation is decoupled from the convo- lution itself, other variants are actually changing the RF it- self. One conclusion might be summed up that, the RF sam- pling/selection manner plays a very important role in convo- lution. While none of these RF selection manners including the regular one is proposed specifically for the radar data, thus they might not fully exploit the potential of convolu- tional networks in radar scene understanding. This concern motivates us to rethink the internal relation between convo- lution and the conventional radar detection mechanism, and try to find a more efficient and specific convolution mecha-nism for radar data. To achieve our goal, we take a look inside of the con- ventional radar detection method and the convolution op- eration in deep learning. As aforementioned, the conven- tional detection method is a kind of CFAR-based peak de- tection, e.g., commonly used cell averaging-CFAR (CA- CFAR) [22]. For a CUT, xc, of the input RD representation, CA-CFAR detection can be divided into three steps: (i) av- eraging aggregation from reference cells {x(i) r}N i=1around CUT, excluding the guard cells; (ii) threshold computing, Θ = ξ·1 NPN i=1x(i) r; (iii) decision-making by comparing xcandΘ. It can be seen that, the decision-making basis is the difference between CUT and its threshold, i.e., the weighted summation of {x(i) r}N i=1with a shared weight,ξ N. In another word, the key to determine whether the CUT has object for CA-CFAR is the denoised peak frequency response from an RF consisted of the CUT and its refer- ence cells. Yet none of the convolution operators mentioned above can explicitly possess such property, i.e., each output unit is actually a weighted summation of the units in a local dense/dilated rectangular or deformable RF, which does not strictly follow the guard-reference policy. Therefore, in this work we redefine the RF of the con- volution operator as the guard-reference style, and call such new type RF the peak receptive field (PRF), which consists of the center unit and its reference neighbors. Then with some simple computational designs, we present two novel conv olution operations to explicitly learn the peak response from PRF, i.e., PeakConvs. Compared with other convolu- tion operations, PeakConvs explicitly possess the advantage of the conventional radar detection methods. In comparison with the conventional CA-CFAR, adaptive peak response with learnable weights and high-level semantic representa- tion via task-driven learning paradigm can be achieved since PeakConvs maintain the computational compatibility of the regular convolution operation. The main contributions are: •A novel convolution computing paradigm for radar data processing . Instead of extracting radar signature directly from RF, we propose learning peak response from redefined PRF, which is more suitable for learn- ing tasks related to radar data. •Two implementations of the proposed PeakConv . According to the participation of center unit dur- ing interference ( e.g., device noises and background clutters) estimation, there are two approaches of PeakConv, including vanilla-PeakConv (PKC), and response d ifference a ware PeakConv (ReDA-PKC). •Well-performed multi-view RSS frameworks based on PeakConvs : by introducing PeakConvs into encoders of the convolutional automatic-encoder- decoder (CAED) framework, two RSS networks with 17578 multi-input and multi-output (MIMO) style are pre- sented. Our networks can achieve SoTA performance on both RD and RA views.
Yi_Generating_Holistic_3D_Human_Motion_From_Speech_CVPR_2023	Abstract This work addresses the problem of generating 3D holistic body motions from human speech. Given a speech record- ing, we synthesize sequences of 3D body poses, hand ges- tures, and facial expressions that are realistic and diverse. To achieve this, we first build a high-quality dataset of 3D holistic body meshes with synchronous speech. We then define a novel speech-to-motion generation framework in which the face, body, and hands are modeled separately. The separated modeling stems from the fact that face artic- ulation strongly correlates with human speech, while body poses and hand gestures are less correlated. Specifically, we employ an autoencoder for face motions, and a composi- tional vector-quantized variational autoencoder (VQ-VAE) for the body and hand motions. The compositional VQ- VAE is key to generating diverse results. Additionally, we propose a cross-conditional autoregressive model that gener- ates body poses and hand gestures, leading to coherent and *Equal Contribution. †Joint Corresponding Authors.realistic motions. Extensive experiments and user studies demonstrate that our proposed approach achieves state-of- the-art performance both qualitatively and quantitatively. Our dataset and code are released for research purposes at https://talkshow.is.tue.mpg.de/ .	1. Introduction From linguistics and psychology we know that humans use body language to convey emotion and use gestures in communication [ 22,28]. Motion cues such as facial expres- sion, body posture and hand movement all play a role. For instance, people may change their gestures when shifting to a new topic [ 52], or wave their hands when greeting an audience. Recent methods have shown rapid progress on modeling the translation from human speech to body mo- tion, and can be roughly divided into rule-based [ 38] and learning-based [ 20,21,23,32,33,55] methods. Typically, the body motion in these methods is represented as the mo- tion of a 3D mesh of the face/upper-body [ 5,15,26,43,44], or 2D/3D landmarks of the face with 2D/3D joints of 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. 469 hands and body [ 21,23,55]. However, this is not sufficient to understand human behavior. Humans communicate with their bodies, hands and facial expressions together. Captur- ing such coordinated activities as well as the full 3D surface in tune with speech is critical for virtual agents to behave realistically and interact with listeners meaningfully. In this work, we focus on generating the expressive 3D motion of person, including their body, hand gestures, and facial expressions, from speech alone; see Fig. 1. To do this, we must learn a cross-modal mapping between audio and 3D holistic body motion, which is very challenging in practice for several reasons. First, datasets of 3D holistic body meshes and synchronous speech recordings are scarce. Acquiring them in the lab is expensive and doing so in the wild has not been possible. Second, real humans often vary in shape, and their faces and hands are highly deformable. It is not trivial to generate both realistic and stable results of 3D holistic body meshes efficiently. Lastly, as different body parts correlate differently with speech signals, it is difficult to model the cross-modal mapping and generate realistic and diverse holistic body motions. We address the above challenges and learn to model the conversational dynamics in a data-driven way. Firstly, to overcome the issue of data scarcity, we present a new set of 3D holistic body mesh annotations with synchronous audio from in-the-wild videos. This dataset was previously used for learning 2D/3D gesture modeling with 2D body key- point annotations [ 21] and 3D keypoint annotations of the holistic body [ 23] by applying existing models separately. Apart from facilitating speech and motion modeling, our dataset can also support broad research topics like realistic digital human rendering. Then, to support our data-driven approach to modeling speech-to-motion translation, an ac- curate holistic body mesh is needed. Existing methods have focused on capturing either the body shape and pose isolated from the hands and face [ 8,17,25,34,47,57,58,61], or the different parts together, which often produces unreal- istic or unstable results, especially when applied to video sequences [ 18,40,62]. To solve this, we present SHOW, which stands for “Synchronous Holistic Optimization in the Wild”. Specifically, SHOW adapts SMPLify-X [ 40] to the videos of talking persons, and further improves it in terms of stability, accuracy, and efficiency through careful design choices. Figure 2 shows example reconstruction results. Lastly, we investigate the translation from audio to 3D holistic body motion represented as a 3D mesh (Fig. 1). We propose TalkSHOW, the first approach to autoregressively synthesize realistic and diverse 3D body motions, hand ges- tures and facial expression of a talking person from speech. Motivated by the fact that the face (i.e. mouth region) is strongly correlated with the audio signal, while the body and hands are less correlated, or even uncorrelated, TalkSHOW designs separate motion generators for different parts andgives each part full play. For the face part, to model the highly correlated nature of phoneme-to-lip motion, we de- sign a simple encoder-decoder based face generator that encodes rich phoneme information by incorporating the pre- trained wav2vec 2.0 [ 6]. On the other hand, to predict the non-deterministic body and hand motions, we devise a novel VQ-V AE [ 50] based framework to learn a compositional quantized space of motion, which efficiently captures a di- verse range of motions. With the learned discrete represen- tation, we further propose a novel autoregressive model to predict a multinomial distribution of future motion, cross- conditioned between existing motions. From this, a wide range of motion modes representing coherent poses can be sampled, leading to realistic looking motion generation. We quantitatively evaluate the realism and diversity of our synthesized motion compared to ground truth and baseline methods and ablations. To further corroborate our qualitative results, we evaluate our approach through an extensive user study. Both quantitative and qualitative studies demonstrate the state-of-the-art quality of our speech-synthesized full expressive 3D character animations.
Yang_Bootstrap_Your_Own_Prior_Towards_Distribution-Agnostic_Novel_Class_Discovery_CVPR_2023	Abstract Novel Class Discovery (NCD) aims to discover unknown classes without any annotation, by exploiting the transfer- able knowledge already learned from a base set of known classes. Existing works hold an impractical assumption that the novel class distribution prior is uniform, yet neglect the imbalanced nature of real-world data. In this paper, we relax this assumption by proposing a new challenging task: distribution-agnostic NCD, which allows data drawn from arbitrary unknown class distributions and thus ren- ders existing methods useless or even harmful. We tackle this challenge by proposing a new method, dubbed “Boot- strapping Your Own Prior (BYOP)”, which iteratively es- timates the class prior based on the model prediction it- self. At each iteration, we devise a dynamic temperature technique that better estimates the class prior by encour- aging sharper predictions for less-confident samples. Thus, BYOP obtains more accurate pseudo-labels for the novel samples, which are beneficial for the next training itera- tion. Extensive experiments show that existing methods suf- fer from imbalanced class distributions, while BYOP1out- performs them by clear margins, demonstrating its effec- tiveness across various distribution scenarios.	1. Introduction With the ever-increasing growth of massive unlabeled data, our community is interested in mining and leveraging the “dark” knowledge therein [2, 7, 28]. To this end, Novel Class Discovery (NCD) [14] is considered as a pivotal step, which aims to automatically recognize novel classes by par- titioning the unlabeled data into different clusters with the knowledge learned from a labeled base class set. Note that the base knowledge is indispensable because clustering without a prior is known as an ill-posed problem [20]—data *Corresponding author. 1Code: https://github.com/muliyangm/BYOP . Model Model Model ModelT ransferLabeled base data “dog” Labeled base data “cat” Unlabeled data ( imbalanced )With uniform prior With uniform prior Novel class 1 Novel class 2With no prior Unlabeled data ( balanced ) Unlabeled data ( balanced ) Novel class 1 Novel class 2Novel class 1 Novel class 2(a) (b) (c)Figure 1. Novel Class Discovery (NCD) in different scenarios. (a) NCD with no prior onbalanced unlabeled data. (b)NCD with theuniform prior onbalanced unlabeled data. (c)NCD with the uniform prior onimbalanced unlabeled data. can always be clustered w.r.t. any feature dimension, e.g., color and background. Hence, the base set provides a pre- liminary prior for defining class vs.non-class features, e.g., the object background feature is removed for discovering new classes. Yet, clustering is still ambiguous to other features not re- moved by the base knowledge. As shown in Fig. 1(a), if we do not specify the class distribution prior, i.e., #sample per class, the two clusters may be considered as red vs.other color , but not the desired moose vs.cow. Therefore, clustering with such a specified prior is a common practice in existing NCD methods [11,34,47]. However, they hold a na¨ıve assumption that the class distribution in the unlabeled data is balanced, i.e., the prior is uniform. This is imprac- tical because the nature of data distribution—especially for 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. 3459 (a) Clustering with Class Prior (b) Class Distribution Prediction New class prior 0.55 0.33 0.11 1 2 3(c) Class Prior Estimation100010001 pseudo-labels 1 2 3 1 2 3 1 2 3Classiﬁer with dynamic temperature 001100 5 1max 0103 132 1235 13predictions 123 5 3 1training samples For next iterationFigure 2. The training pipeline of our proposed BYOP for distribution-agnostic NCD. (a)At each iteration, BYOP clusters the unlabeled data using the class prior coming from the previ- ous iteration to generate pseudo-labels (Sec. 3.1). (b)BYOP is trained to predict the novel class distributions using the generated pseudo-labels, where we devise a dynamic temperature technique to encourage more confident predictions (Sec. 3.2). (c)The class prior is estimated by calculating the proportion of each class as- signment, which is ready to use for the next iteration (Sec. 3.3). large-scale data—is imbalanced [23, 31, 33]. As shown in the comparison between Figs. 1(b) and (c), if the data is imbalanced, the uniform prior is misleading. In this paper, we relax such an impractical assump- tion by allowing novel data drawn from an arbitrary un- known class distribution. We term this new challenging task distribution-agnostic NCD , which renders existing meth- ods useless or even harmful when the novel data is highly- imbalanced. The crux of the problem is the prior itself—on one hand, it is a critical ingredient against cluster ambigui- ties; on the other hand, it becomes misleading when it mis- matches with the true class distribution. This gives rise to a chicken-egg problem in distribution-agnostic NCD, as the class distribution is no longer known as a priori. We pro- pose to address this dilemma by “ Bootstrapping YourOwn Prior” (BYOP /baI"6p/ )—iteratively estimating the class distribution based on the model prediction itself, which can be used as a prior to obtain more accurate pseudo-labels that help the next training iteration. The BYOP pipeline is summarized in Fig. 2. Given a batch of unlabeled data with an arbitrary unknown class distribution, we deploy a clustering method [1] that parti- tions the data subject to the current class prior. At each iteration, the current class prior estimation is not yet ac- curate ( e.g., we initialize by the uniform prior), and thus may result in ambiguous clusters for the minority classes if the true class distribution is highly-imbalanced (Fig. 2(a)).The cluster assignments are used as pseudo-labels to train a classifier to discover novel classes. However, due to the imperfections in pseudo-labels, the predicted class distribu- tions are inevitably ambiguous, especially for those minor- ity classes. To this end, we propose a dynamic tempera- ture technique that can be integrated into the classifier to output more confident distribution predictions (Fig. 2(b)). The main idea is to encourage sharper predicted distribu- tions for less-confident data by a per-sample temperature adjustment. In particular, we call it “adaptive” because it won’t hurt the prediction for the samples which are already confident, while significantly disambiguating those who are less confident, as later discussed in Fig. 3. To estimate the class prior, we gather the predicted novel class distributions as the class assignments for the training samples, and calculate the proportion of each class assign- ment (Fig. 2(c)), so that we can derive a new class prior that is beneficial for the next training iteration. Note that the higher prediction accuracy for majority classes guar- antees to estimate a preliminary prior that helps generate more accurate pseudo-labels, which in turn promotes the reliability of the prior estimation for other classes via more accurate model predictions. We benchmark our proposed BYOP and the current state-of-the-art methods in the chal- lenging distribution-agnostic NCD task on several standard datasets. While current methods suffer from imbalanced class distributions, BYOP outperforms them by large mar- gins, demonstrating its effectiveness across different class distributions, including the conventionally balanced one. To sum up, our contributions are three-fold: • A new challenging distribution-agnostic NCD task that relaxes the impractical uniform class distribution as- sumption in current NCD works. • A novel training paradigm dubbed BYOP to handle ar- bitrary unknown class distributions in NCD by itera- tively estimating and utilizing the class prior. • Extensive experiments that benchmark the current state-of-the-art methods as well as the superiority of the proposed BYOP in distribution-agnostic NCD.
Yi_Towards_Artistic_Image_Aesthetics_Assessment_A_Large-Scale_Dataset_and_a_CVPR_2023	Abstract Image aesthetics assessment (IAA) is a challenging task due to its highly subjective nature. Most of the current stud- ies rely on large-scale datasets (e.g., AVA and AADB) to learn a general model for all kinds of photography images. However, little light has been shed on measuring the aes- thetic quality of artistic images, and the existing datasets only contain relatively few artworks. Such a defect is a great obstacle to the aesthetic assessment of artistic images. To fill the gap in the field of artistic image aesthetics assess- ment (AIAA), we first introduce a large-scale AIAA dataset: Boldbrush Artistic Image Dataset (BAID), which consists of 60,337 artistic images covering various art forms, with more than 360,000 votes from online users. We then pro- pose a new method, SAAN (Style-specific Art Assessment Network), which can effectively extract and utilize style- specific and generic aesthetic information to evaluate artis- tic images. Experiments demonstrate that our proposed approach outperforms existing IAA methods on the pro- posed BAID dataset according to quantitative comparisons. We believe the proposed dataset and method can serve as a foundation for future AIAA works and inspire more re- search in this field. Dataset and code are available at: https://github.com/Dreemurr-T/BAID.git	1. Introduction With the ever-growing scale of online visual data, image aesthetic assessment (IAA) shows great potential in a vari- ety of applications such as photo recommendation, image ranking and image search [6]. In recent years, image style transfer [9, 14, 19, 20, 26] and AI painting [15, 39] have be- come high-profile research areas. Users can easily generate artworks of numerous styles from websites and online ap- plications, which has led to the explosion of artistic images online and the drastic increase in demand for automatically evaluating artwork aesthetics. We refer to this problem as *Corresponding author. Figure 1. Samples from the proposed BAID dataset. BAID covers a wide range of artistic styles and painting themes. artistic image aesthetic assessment (AIAA) . The artistic image aesthetic assessment task is similar to IAA for being extremely challenging due to its highly subjective nature, as different individuals may have distinct visual and art preferences. Existing datasets related to this task can be summarized into three categories, but none of them meets the requirements of the AIAA task: (1) IAA datasets : modern IAA methods [13, 21, 23, 30, 32, 34] are data-driven, usually trained and evaluated on large-scale IAA datasets, e.g., A V A [25], AADB [17] and CUHK- PQ [22]. However, these datasets only contain real-world 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. 22388 photos and do not include artistic images like oil paintings or pencil sketches. This deficiency of artistic images is prevalent in existing IAA datasets [4, 16, 17, 22, 28], which means that given an artwork, existing IAA methods evalu- ate it based on perceptions learned from photography, and the evaluation is likely to be inaccurate since the perceptual rules of photography and art are not the same. (2) Artis- tic datasets without aesthetic labels : existing large-scale artistic image datasets [1, 29, 36] are mainly used to train style transfer, artistic style classification or text to image models, but they lack score annotations indicating image aesthetic level. (3) Small-scale AIAA datasets : efforts into building public AIAA datasets are scarce and the existing datasets [3,8] contain relatively few number of images (less than 2,000). Based on the above observations, we conclude thatthe lack of a large-scale AIAA dataset is the biggest obstacle towards developing AIAA approaches. To solve the problem, we first introduce a large-scale dataset specifically constructed for the AIAA task: the Boldbrush Artistic Image Dataset (BAID), which consists of 60,337 artworks annotated with more than 360,000 votes. The proposed BAID is, to our knowledge, the largest AIAA dataset, which far exceeds existing IAA and AIAA datasets in the quantity and quality of artworks. Furthermore, we propose a baseline model, called the Style-specific Art Assessment Network (SAAN), which can effectively exploit the style features and the generic aes- thetic features of the given artwork. Our model consists of three modules: 1) Generic Aesthetic Feature Extrac- tion Branch: inspired by the studies [27, 31], we adopt a self-supervised learning scheme to train a Generic Aes- thetic Branch to extract aesthetics-aware features. The self- supervised scheme is based on the correlation between the aesthetic quality of the images and degradation editing op- erations. This essentially provides data augmentation such that the model can better learn the quality of different art- works. 2) Style-specific Aesthetic Feature Extraction Branch: observing that the style of the artwork is critical when assessing its aesthetic value and different styles need to extract different style-related aesthetic features, we pro- pose a Style-specific Aesthetic Branch to incorporate style information into aesthetic features and extract style-specific aesthetic features via adaptive instance normalization [14]. 3)Spatial Information Fusion: we also add a non-local block [35] into the proposed method to fuse spatial infor- mation into the extracted aesthetic features. The main contributions of our work are three-fold: • We address the problem of artistic image aesthetics assessment, and introduce a new large-scale dataset BAID consisting of 60,337 artworks annotated with more than 360,000 votes to facilitate research in this direction. • We propose a style-specific artistic image assessmentTable 1. Summary of IAA/AIAA datasets and our proposed BAID dataset. BAID provides a significantly larger number of artistic images and has user subjective votes. Dataset Number of images Number of artistic images DP Challenge [4] 16,509 – Photo.Net [16] 20,278 – CUHK-PQ [22] 17,673 – A V A [25] 255,530 – AADB [17] 10,000 – FLICKR-AES [28] 40,000 – PARA [37] 31,220 – TAD66K [12] 66,327 1,200 JenAesthetic [3] 1,628 1,628 V APS [8] 999 999 BAID (Ours) 60,337 60,337 network called SAAN, which combines style-specific and generic aesthetic features to evaluate artworks. • We evaluate the state-of-the-art IAA approaches and our proposed method on the proposed BAID dataset. Our model achieves promising results on all the met- rics, which clearly demonstrates the validity of our model.
Yan_Two-Shot_Video_Object_Segmentation_CVPR_2023	Abstract Previous works on video object segmentation (VOS) are trained on densely annotated videos. Nevertheless, ac- quiring annotations in pixel level is expensive and time- consuming. In this work, we demonstrate the feasibility of training a satisfactory VOS model on sparsely annotated videos—we merely require two labeled frames per train- ing video while the performance is sustained. We term this novel training paradigm as two-shot video object segmen- tation, or two-shot VOS for short. The underlying idea is to generate pseudo labels for unlabeled frames during train- ing and to optimize the model on the combination of la- beled and pseudo-labeled data. Our approach is extremely simple and can be applied to a majority of existing frame- works. We first pre-train a VOS model on sparsely an- notated videos in a semi-supervised manner, with the first frame always being a labeled one. Then, we adopt the pre- trained VOS model to generate pseudo labels for all un- labeled frames, which are subsequently stored in a pseudo- label bank. Finally, we retrain a VOS model on both labeled and pseudo-labeled data without any restrictions on the first frame. For the first time, we present a general way to train VOS models on two-shot VOS datasets. By using 7.3%and 2.9%labeled data of YouTube-VOS and DAVIS benchmarks, our approach achieves comparable results in contrast to the counterparts trained on fully labeled set. Code and models are available at https://github.com/yk- pku/Two-shot-Video-Object-Segmentation .	1. Introduction Video object segmentation (VOS), also known as mask tracking, aims to segment the target object in a video given the annotation of the reference (or first) frame. Existing approaches [7, 9, 21, 30, 37, 46, 52] are trained on densely annotated datasets such as DA VIS [33, 34] and YouTube- VOS [50]. However, acquiring dense annotations, partic- *Corresponding authors. …… Previous Methods (DenselyAnnotated Videos) Ours (2Labeled Frames per Video)UnlabeledUnlabeled… (a) Previous works on video object segmentation rely on densely annotated videos. We present two-shot video object segmentation, which merely ac- cesses two labeled frames per video. 87.981.0 80.8 80.691.585.283.0 82.791.385.182.982.7 6065707580859095 DAVIS 2016DAVIS 2017YouTube-VOS 2018YouTube-VOS 2019Score 2-shot STCNFull-set STCN2-shot STCN w/ Ours (b) Comparison among naive 2-shot STCN, STCN trained on full set and 2-shot STCN equipped with our approach on DA VIS 2016/2017 and YouTube-VOS 2018/2019. Figure 1. (a) Problem formulation. (b) Comparison among STCN variants on various datasets. ularly at the pixel level, is laborious and time-consuming. For instance, the DA VIS benchmark consists of 60 videos, each with an average of 70 labeled frames; the YouTube- VOS dataset has an even larger amount of videos, and every fifth frame of each video is labeled to lower the annotation cost. It is necessary to develop data-efficient VOS models to reduce the dependency on labeled data. In this work, we investigate the feasibility of training a 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. 2257 satisfactory VOS model on sparsely annotated videos. For the sake of convenience, we use the term N-shot to denote thatNframes are annotated per training video. Note that 1-shot is meaningless since it degrades VOS to the task of image-level segmentation. We use STCN [9] as our base- line due to its simplicity and popularity. Since at least two labeled frames per video are required for VOS train- ing, we follow the common practice to optimize a naive 2-shot STCN model on the combination of YouTube-VOS and DA VIS, and evaluate on YouTube-VOS 2018/2019 and DA VIS 2016/2017, respectively. We compare the native 2- shot STCN with its counterpart trained on full set in Fig. 1b. Surprisingly, 2-shot STCN still achieves decent results, for instance, only a −2.1%performance drop is observed on YouTube-VOS 2019 benchmark, demonstrating the practi- cality of 2-shot VOS. So far, the wealth of information present in unlabeled frames is yet underexplored. In the last decades, semi- supervised learning, which combines a small amount of labeled data with a large collection of unlabeled data dur- ing training, has achieved considerable success on vari- ous tasks such as image classification [3, 39], object detec- tion [40, 49] and semantic segmentation [14, 17]. In this work, we also adopt this learning paradigm to promote 2- shot VOS (see Fig. 1a). The underlying idea is to generate credible pseudo labels for unlabeled frames during training and to optimize the model on the combination of labeled and pseudo-labeled data. Here we continue to use STCN [9] as an example to illustrate our design principle, neverthe- less, our approach is compatible with most VOS models. Concretely, STCN takes a randomly selected triplet of la- beled frames as input but the supervisions are only applied to the last two—VOS requires the annotation of the first frame as reference to segment the object of interest that ap- peared in subsequent frames. This motivates us to utilize the ground-truth for the first frame to avoid error propa- gation during early training. Each of the last two frames, nevertheless, can be either a labeled frame or an unlabeled frame with a high-quality pseudo label. Although the per- formance is improved with this straightforward paradigm, the capability of semi-supervised learning is still underex- plored due to the restriction of employing the ground truth as the starting frame. We term the process described above asphase-1 . To take full advantage of unlabeled data, we lift the re- striction placed on the starting frame, allowing it to be either a labeled or pseudo-labeled frame. To be specific, we adopt the VOS model trained in phase-1 to infer the unlabeled frames for pseudo-labeling. After that, each frame is as- sociated with a pseudo label that approximates the ground- truth. The generated pseudo labels are stored in a pseudo- label bank for the convenience of access. The VOS model is then retrained without any restrictions—similar to howit is trained through supervised learning, but each frame has either a ground-truth or a pseudo-label attached to it. It is worth noting that, as training progresses, the predic- tions become more precise, yielding more reliable pseudo labels—we update the pseudo-label bank once we identify such pseudo labels. The above described process is named asphase-2 . As shown in Fig. 1b, our approach assembled onto STCN, achieves comparable results ( e.g. 85.2% v.s 85.1% on DA VIS 2017, and 82.7% v.s 82.7% on YouTube- VOS 2019) in contrast to its counterpart, STCN trained on full set, though our approach merely accesses 7.3% and 2.9% labeled data of YouTube-VOS and DA VIS bench- mark, respectively. Our contributions can be summarized as follows: • For the first time, we demonstrate the feasibility of two-shot video object segmentation: two labeled frames per video are almost sufficient for training a decent VOS model, even without the use of unlabeled data. • We present a simple yet efficient training paradigm to exploit the wealth of information present in unlabeled frames. This novel paradigm can be seamlessly ap- plied to various VOS models, e.g., STCN [9], RDE- VOS [21] and XMem [7] in our experiments. • Though we only access a small amount of labeled data (e.g.7.3%for YouTube-VOS and 2.9%for DA VIS), our approach still achieves competitive results in con- trast to the counterparts trained on full set. For example, 2-shot STCN equipped with our approach achieves 85.1 %/82.7%on DA VIS 2017/YouTube-VOS 2019, which is +4.1 %/+2.1 %higher than the naive 2- shot STCN while -0.1 %/-0.0%lower than the STCN trained on full set.
Yang_Context_De-Confounded_Emotion_Recognition_CVPR_2023	Abstract Context-Aware Emotion Recognition (CAER) is a cru- cial and challenging task that aims to perceive the emo- tional states of the target person with contextual informa- tion. Recent approaches invariably focus on designing so- phisticated architectures or mechanisms to extract seem- ingly meaningful representations from subjects and con- texts. However, a long-overlooked issue is that a con- text bias in existing datasets leads to a significantly unbal- anced distribution of emotional states among different con- text scenarios. Concretely, the harmful bias is a confounder that misleads existing models to learn spurious correlations based on conventional likelihood estimation, significantly limiting the models’ performance. To tackle the issue, this paper provides a causality-based perspective to disentan- gle the models from the impact of such bias, and formu- late the causalities among variables in the CAER task via a tailored causal graph. Then, we propose a Contextual Causal Intervention Module (CCIM) based on the backdoor adjustment to de-confound the confounder and exploit the true causal effect for model training. CCIM is plug-in and model-agnostic, which improves diverse state-of-the-art ap- proaches by considerable margins. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our CCIM and the significance of causal insight.	1. Introduction As an essential technology for understanding human in- tentions, emotion recognition has attracted significant at- tention in various fields such as human-computer interac- tion [1], medical monitoring [28], and education [40]. Pre- vious works have focused on extracting multimodal emo- tion cues from human subjects, including facial expres- sions [9, 10, 49], acoustic behaviors [2, 50, 52], and body §Corresponding Author. Engagement Engagement Excitement Happiness PleasureAffection Happiness PleasureDisapproval Disconnection Disquietment Doubt/Confusion Engagement Sadness Testing phase GT： GT： GT： GT：GT：Training phase Prediction Kosti et al. Affection Anticipation Engagement Excitement Pleasure Kosti et al. + CCIM(ours) Similar Context Disapproval Disconnection Disquietment Doubt/Confusion Engagement SadnessGrass,trees, outdoors,etc Figure 1. Illustration of the context bias in the CAER task. GT means the ground truth. Most images contain similar contexts in the training data with positive emotion categories. In this case, the model learns the spurious correlation between specific contexts and emotion categories and gives wrong results. Thanks to CCIM, the simple baseline [19] achieves more accurate predictions. postures [25, 53], benefiting from advances in deep learn- ing algorithms [6, 7, 21, 26, 27, 43, 44, 46, 47, 54, 55, 59]. Despite the impressive improvements achieved by subject- centered approaches, their performance is limited by natu- ral and unconstrained environments. Several examples in Figure 1 (left) show typical situations on a visual level. In- stead of well-designed visual contents, multimodal repre- sentations of subjects in wild-collected images are usually indistinguishable ( e.g., ambiguous faces or gestures), which forces us to exploit complementary factors around the sub- ject that potentially reflect emotions. Inspired by psychological study [3], recent works [19,22, 23, 29, 56] have suggested that contextual information con- tributes to effective emotion cues for Context-Aware Emo- tion Recognition (CAER). The contexts are considered to include the place category, the place attributes, the objects, or the actions of others around the subject [20]. The major- ity of such research typically follows a common pipeline: (1) Obtaining the unimodal/multimodal representations of the recognized subject; (2) Building diverse contexts and extracting emotion-related representations; (3) Designing 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. 19005 (0.8, 1.0] [0, 0.2]020406080100120 (0.2, 0.4] (0.4, 0.6] (0.6, 0.8]Sc ene Categories Conditional Entropy (Anger)(0.8, 1.0] [0, 0.2]020406080100120 (0.2, 0.4] (0.4, 0.6] (0.6, 0.8]Sc ene Categories Conditional Entropy (Happy) (0.8, 1.0] [0, 0.2]020406080100120 (0.2, 0.4] (0.4, 0.6] (0.6, 0.8]Sc ene Categories Conditional Entropy (Anger)(0.8, 1.0] [0, 0.2]020406080100120 (0.2, 0.4] (0.4, 0.6] (0.6, 0.8]Sc ene Categories Conditional Entropy (Happy)(a) EMOTIC Dataset (b) CAER -S DatasetFigure 2. We show a toy experiment on the EMOTIC [20] and CAER-S [22] datasets for scene categories of angry and happy emotions. More scene categories with normalized zero- conditional entropy reveal a strong presence of the context bias. fusion strategies to combine these features for emotion la- bel predictions. Although existing methods have improved modestly through complex module stacking [12,23,51] and tricks [16, 29], they invariably suffer from a context bias of the datasets, which has long been overlooked. Recall- ing the process of generating CAER datasets, different an- notators were asked to label each image according to what they subjectively thought people in the images with diverse contexts were feeling [20]. This protocol makes the prefer- ence of annotators inevitably affect the distribution of emo- tion categories across contexts, thereby leading to the con- text bias. Figure 1 illustrates how such bias confounds the predictions. Intrigued, most of the images in training data contain vegetated scenes with positive emotion categories, while negative emotions in similar contexts are almost non- existent. Therefore, the baseline [19] is potentially misled into learning the spurious dependencies between context- specific features and label semantics. When given test im- ages with similar contexts but negative emotion categories, the model inevitably infers the wrong emotional states. More intrigued, a toy experiment is performed to ver- ify the strong bias in CAER datasets. This test aims to ob- serve how well emotions correlate with contexts ( e.g., scene categories). Specifically, we employ the ResNet-152 [15] pre-trained on Places365 [58] to predict scene categories from images with three common emotion categories ( i.e., “anger”, “happy”, and “fear”) across two datasets. The top 200 most frequent scenes from each emotion category are selected, and the normalized conditional entropy of each scene category across the positive and negative set of a spe- cific emotion is computed [30]. While analyzing correla- tions between scene contexts and emotion categories in Fig- ure 2 ( e.g., “anger” and “happy”), we find that more scene categories with the zero conditional entropy are most likely to suggest the significant context bias in the datasets, asit shows the presence of these scenes only in the positive or negative set of emotions. Concretely, for the EMOTIC dataset [20], about 40% of scene categories for anger have zero conditional entropy while about 45% of categories for happy ( i.e., happiness) have zero conditional entropy. As an intuitive example, most party-related scene contexts are present in the samples with the happy category and almost non-existent in the negative categories. These observations confirm the severe context bias in CAER datasets, leading to distribution gaps in emotion categories across contexts and uneven visual representations . Motivated by the above observation, we attempt to em- brace causal inference [31] to reveal the culprit that poi- sons the CAER models, rather than focusing on beating them. As a revolutionary scientific paradigm that facili- tates models toward unbiased prediction, the most impor- tant challenge in applying classical causal inference to the modern CAER task is how to reasonably depict true causal effects and identify the task-specific dataset bias. To this end, this paper attempts to address the challenge and rescue the bias-ridden models by drawing on human instincts, i.e., looking for the causality behind any association. Specifi- cally, we present a causality-based bias mitigation strategy. We first formulate the procedure of the CAER task via a proposed causal graph. In this case, the harmful context bias in datasets is essentially an unintended confounder that misleads the models to learn the spurious correlation between similar contexts and specific emotion semantics. From Figure 3, we disentangle the causalities among the input images X, subject features S, context features C, confounder Z, and predictions Y. Then, we propose a simple yet effective Contextual Causal Intervention Module (CCIM) to achieve context-deconfounded training and use thedo-calculus P(Y\|do(X))to calculate the true causal ef- fect, which is fundamentally different from the conventional likelihood P(Y\|X). CCIM is plug-in and model-agnostic, with the backdoor adjustment [14] to de-confound the con- founder and eliminate the impact of the context bias. We comprehensively evaluate the effectiveness and superiority of CCIM on three standard and biased CAER datasets. Nu- merous experiments and analyses demonstrate that CCIM can significantly and consistently improve existing base- lines, achieving a new state-of-the-art (SOTA). The main contributions can be summarized as follows: • To our best knowledge, we are the first to investigate the adverse context bias of the datasets in the CAER task from the causal inference perspective and iden- tify that such bias is a confounder, which misleads the models to learn the spurious correlation. • We propose CCIM, a plug-in contextual causal inter- vention module, which could be inserted into most CAER models to remove the side effect caused by the 19006 confounder and facilitate a fair contribution of diverse contexts to emotion understanding. • Extensive experiments on three standard CAER datasets show that the proposed CCIM can facilitate existing models to achieve unbiased predictions.
Yang_Towards_Effective_Adversarial_Textured_3D_Meshes_on_Physical_Face_Recognition_CVPR_2023	Abstract Face recognition is a prevailing authentication solution in numerous biometric applications. Physical adversarial attacks, as an important surrogate, can identify the weak- nesses of face recognition systems and evaluate their ro- bustness before deployed. However, most existing physical attacks are either detectable readily or ineffective against commercial recognition systems. The goal of this work is to develop a more reliable technique that can carry out an end- to-end evaluation of adversarial robustness for commercial systems. It requires that this technique can simultaneously deceive black-box recognition models and evade defensive mechanisms. To fulfill this, we design adversarial textured 3D meshes ( AT3D ) with an elaborate topology on a human face, which can be 3D-printed and pasted on the attacker’s face to evade the defenses. However, the mesh-based op- timization regime calculates gradients in high-dimensional mesh space, and can be trapped into local optima with un- satisfactory transferability. To deviate from the mesh-based space, we propose to perturb the low-dimensional coeffi- cient space based on 3D Morphable Model, which signifi- cantly improves black-box transferability meanwhile enjoy- ing faster search efficiency and better visual quality. Exten- sive experiments in digital and physical scenarios show that our method effectively explores the security vulnerabilities of multiple popular commercial services, including three recognition APIs, four anti-spoofing APIs, twoprevailing mobile phones and twoautomated access control systems.	1. Introduction Face recognition has become a prevailing authentication solution in biometric applications, ranging from financial payment to automated surveillance systems. Despite its †Corresponding authors. üüFace Recognition Face Anti-spoofing Owner (Victim) AttackerFigure 1. Demonstration of physical black-box attacks for unlock- ing one prevailing mobile phone. The attacker wearing the 3D- printed adversarial mesh can successfully mislead the face recog- nition model to be recognized as the victim, meanwhile evading face anti-spoofing. More results are shown in Sec. 4. blooming development [4, 26, 33], recent research in adver- sarial machine learning has revealed that face recognition models based on deep neural networks are highly vulnera- ble to adversarial examples [10,41], leading to serious con- sequences or security problems in real-world applications. Due to the imperative need of evaluating model robust- ness [30, 45], extensive attempts have been devoted to ad- versarial attacks on face recognition models. Adversarial at- tacks in the digital world [8,28,39,45] are characterized by adding minimal perturbations to face images in the digital space, aiming to evade being recognized or to impersonate another identity. Since an adversary usually cannot access the digital input of practical systems, physical adversarial examples wearable for real human faces are more feasible for evaluating their adversarial robustness. Some studies have shown the success of physical attacks against popular recognition models by adopting different attack types, such as eyeglass frames [27, 28], hats [17] and stickers [29]. In spite of the remarkable progress, it is challenging to launch practical andeffective physical attack methods on automatic face recognition systems. First, the defen- 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. 4119 Frames [28] AdvHat [17] FaceAdv [29] PadvFace [50] AdvMask [52] Face3DAdv [40] RHDE [35] Ours 3D attack types No Partially Partially No Yes Yes Partially Yes Commercial recognition Yes No No No No No Yes Yes Commercial defenses No No No No No Yes No Yes Number of physical tests 10 3 10 10 30 10 3 50 Table 1. A comparison among different methods regarding whether using 3D attack types, commercial face recognition models, commercial defenses, and the number of physical evaluation. Partially indicates that this method involved some geometric transformations to make 2D patch approximately approach the realistic 3D patch. sive mechanism [14, 42, 43, 46, 48] on face recognition, i.e., face anti-spoofing, has achieved impressive performance among the academic and industry communities. Some pop- ular defenses [18, 34, 49] have injected more sensors (such as depth, multi-spectral and infrared cameras) to provide more effective defenses. However, most of the physical at- tacks have not evaluated the passing rates against practi- cal defensive mechanisms, as reported in Table. 1. Second, these methods cannot perform satisfactorily for imperson- ation attacks against diverse commercial black-box recog- nition models due to the limited black-box transferability. The goal of this work is to develop practical andeffective physical adversarial attacks that can simultaneously deceive black-box recognition models and evade defensive mecha- nisms in commercial face recognition systems, e.g., unlock- ing mobile phones, as demonstrated in Fig. 1. Evading the defensive mechanisms. Recent research has found that high-fidelity 3D masks [19, 21] can better fool the prevailing face anti-spoofing methods by 3D print- ing techniques. It becomes an appealing and feasible way to apply a 3D adversarial mask for evading defensive mecha- nisms in face recognition systems. To achieve this goal, we first design adversarial textured 3D meshes ( AT3D ) with an elaborate topology on a human face, which can be us- able by standard graphics software such as Blender [9] and Maya [22]. As a primary 3D representation, textured meshes can be immediately 3D-printed and pasted on real faces for physical adversarial attacks, which have geometric details, complex topology and high-quality textures. Exper- imentally, AT3D can be more conducive to steadily passing commercial face anti-spoofing services, such as FaceID and Tencent anti-spoofing APIs, two mobile phones and two ac- cess control systems with multiple sensors . Misleading the black-box recognition models. The typical 3D mesh attacks [23, 36, 47] proposed to optimize adversarial examples in mesh representation space. Thus, high complexity is virtually inevitable for calculating gradi- ents in such high-dimensional search space due to the thou- sands of triangle faces on each human face. The procedures are also costly and probably trapped into overfitting [20] with unsatisfactory transferability. Therefore, we aim to perform the optimization trajectory in a low-dimensional manifold as a regularization aiming for escaping from over- fitting. The low-dimensional manifold should possess a sufficient capacity that encodes any 3D face in this low-dimensional feature space, thus successfully achieving the white-box adversarial attack against a substitute model. A principled way of spanning such a subspace is considered by leveraging 3D Morphable Model (3DMM) [31] that ef- fectively achieves dimensionality reduction of any high- dimensional mesh data. Based on this, we are capable of generating an adversarial mesh by perturbing the low- dimensional coefficients of 3DMM, making it constrained on the data manifold of realistic 3D faces. Therefore, the crafted mesh can obtain a strong semantic feature of a 3D face, which can achieve well-generalizing performance among the white-box and black-box models due to knowl- edgable semantic pattern characteristics [37, 38, 44]. In ad- dition, low-dimensional optimization can also avoid self- intersection and flying vertices problems in mesh-based op- timization [47], resulting in better visual appearance. Experimentally, we have effectively explored the secu- rity vulnerabilities of multiple popular commercial services, including 1) recognition APIs—Amazon, Face++, and Ten- cent; 2) anti-spoofing APIs—FaceID, SenseID, Tencent, and Aliyun; 3) twoprevailing mobile phones and twoauto- mated access control systems that incorporate multiple sen- sors. Our main contributions can be summarized as: • We propose effective and practical adversarial textured 3D meshes with elaborate topology and effective opti- mization, which can simultaneously evade black-box recognition models and defensive mechanisms. • Extensive physical experiments demonstrate that our method can consistently mislead multiple commer- cial systems, including unlocking prevailing mobile phones and automated access control systems. • We present a reliable technique to evaluate the robust- ness of face recognition systems, which can be further leveraged as an effective data augmentation strategy to improve defensive ability.
Yu_ANetQA_A_Large-Scale_Benchmark_for_Fine-Grained_Compositional_Reasoning_Over_Untrimmed_CVPR_2023	Abstract Building benchmarks to systemically analyze different capabilities of video question answering (VideoQA) models is challenging yet crucial. Existing benchmarks often use non-compositional simple questions and suffer from language biases, making it difﬁcult to diagnose model weaknesses incisively. A recent benchmark AGQA [8] poses a promising paradigm to generate QA pairs automatically from pre-annotated scene graphs, enabling it to measure diverse reasoning abilities with granular control. However, its questions have limitations in reasoning about the ﬁne- grained semantics in videos as such information is absent in its scene graphs. To this end, we present ANetQA, a large-scale benchmark that supports ﬁne-grained compo- sitional reasoning over the challenging untrimmed videos from ActivityNet [4]. Similar to AGQA, the QA pairs in ANetQA are automatically generated from annotated video scene graphs. The ﬁne-grained properties of ANetQA are reﬂected in the following: (i) untrimmed videos with ﬁne-grained semantics; (ii) spatio-temporal scene graphs with ﬁne-grained taxonomies; and (iii) diverse questions generated from ﬁne-grained templates. ANetQA attains 1.4 billion unbalanced and 13.4 million balanced QA pairs, which is an order of magnitude larger than AGQA with a similar number of videos. Comprehensive experiments are performed for state-of-the-art methods. The best model achieves 44.5% accuracy while human performance tops out at 84.5%, leaving sufﬁcient room for improvement.	1. Introduction Recent advances in deep learning have enabled machines to tackle complicated video-language tasks that involve Jun Yu is the corresponding author A: diving gearQ1: Is the diver jumping into the water before sea turtles graze and swim in the ocean ?Example QA pairs A: yes Q1: Did the person twist the bottle after taking a picture ? A: yesExample QA pairs Q2: What did the person hold after putting a phone down ? A: bottleaction object attribute relationship Q4: What is the occupation of the person with black headwear ? A: diver a manta ray swims in the ocean over a reef. various fish are seen swimming through the reefsea turtles graze and swim in the ocean beginning of the video end of the videopeople suit up in diving gear then jump into waterblue diver black headwearperson diving gearwearingjumping intowater fish swimming reefcrossing greenblack and whitechasing onmanta ray manta ray reefocean in sea turtle greenswimmingwrackgreenblue eating sea turtleinocean blue swimming identical Q3: What color is the swimming sea turtle before it is eating the wrack? A: green Q2: What is the black object that the person is wearing before various fish are seen swimming through the reef ?black AGQAANetQA (ours)similar templatesFigure 1. Comparisons of ANetQA and AGQA [8]. The QA pairs in both benchmarks are automatically generated from spatio- temporal scene graphs by using handcrafted question templates. Beneﬁting from the untrimmed long videos and ﬁne-grained scene graphs, our questions require more ﬁne-grained reasoning abilities than those in AGQA when similar templates are applied. Moreover, the newly introduced attribute annotations allow us to design many ﬁne-grained question templates that are not supported in AGQA ( e.g., “what color ” and “ what is the occupation ”). both video and language clues, e.g., video-text retrieval, video captioning, video temporal grounding, and video question answering. Among these tasks, video question answering (VideoQA) is one of the most challenging tasks as it veriﬁes multiple skills simultaneously. Taking the question “ What is the black object that the person is wearing before various ﬁsh are seen swimming through the reef? ” in Figure 1 as an example, it requires a synergistic understanding of both the video and question, together with spatio-temporal reasoning to predict an accurate answer. 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. 23191 To comprehensively evaluate the capabilities of existing VideoQA models, several prominent benchmarks have been established [11, 21, 29, 33, 38, 42, 43]. Despite their useful- ness, they also have distinct shortcomings. Some bench- marks use simulated environments to synthesize video con- tents [29, 42], which provides controllable diagnostics over different reasoning skills. However, the synthetic videos lack visual diversity and the learned models on the bench- marks cannot generalize to real-world scenarios directly. Some real-world benchmarks generate QA pairs from off- the-shelf video captions [38, 48] or human annotations [11, 21, 33, 43], which suffer from simple question expressions and biased answer distributions. These weaknesses may be exploited by models to make educated guesses to obtain the correct answers without seeing video contents [24, 40]. One recent VideoQA benchmark AGQA poses a promis- ing paradigm to address the above limitations [8]. AGQA is built upon the real-world videos from Charades [32]. In contrast to previous benchmarks, AGQA adopts a two- stage paradigm instead. For each video, a spatio-temporal scene graph over representative frames is ﬁrst annotated by humans, which consists of spatially-grounded object- relationship triplets and temporally-grounded actions. After that, different types of questions are generated on top of the scene graph using corresponding question templates, enabling it to measure various reasoning abilities with granular control. Despite the comprehensiveness of AGQA, we argue that its foundation—the spatio-temporal scene graph—has limitations in representing the ﬁne-grained se- mantics of videos. Speciﬁcally, their scene graphs encode objects and relationships from limited taxonomies, which arenotﬁne-grained enough for generating questions that require reasoning about the detailed video semantics. To this end, we introduce ANetQA1, a new benchmark that supports ﬁne-grained compositional reasoning over complex web videos from ActivityNet [4]. Similar to the strategy of AGQA, the QA pairs in ANetQA are automati- cally generated from pre-annotated scene graphs. As shown in Figure 1, we claim that ANetQA is more ﬁne-grained than AGQA in terms of the following: (i) The benchmark is built upon untrimmed long videos with ﬁne-grained semantics. Each video may involves multiple indoor or outdoor scenarios, containing com- plicated interactions between persons and objects. (ii) The spatio-temporal scene graph consists of ﬁne- grained objects ( e.g., “manta ray ”, “diving gear ”), relationships ( e.g., “jumping into ”, “chasing ”), at- tributes ( e.g., “swimming ”, “black and white ”), and actions in natural language ( e.g., “a manta ray swims in the ocean over a reef ”). 1Note that there is a VideoQA benchmark ActivityNet-QA [43] whose QA pairs are fully annotated by humans. To avoid confusion, we name our benchmark ANetQA.(iii) Beneﬁting from the ﬁne-grained scene graphs, we are able to design diverse question templates that requires ﬁne-grained compositional reasoning ( e.g., “what color ... ” and “ what is the occupation ... ”). Beneﬁting from the above ﬁne-grained characteristics, ANetQA obtains 1.4B unbalanced and 13.4M balanced QA pairs. To the best of our knowledge, ANetQA is the largest VideoQA benchmark in terms of the number of questions. Compared with the previous largest benchmark AGQA, ANetQA is an order of magnitude larger than it with a similar number of videos. We conduct comprehensive experiments and intensive analyses on ANetQA for the state-of-the-art VideoQA models, including HCRN [19], ClipBERT [20], and All-in-One [35]. The best model delivers 44.5% accuracy while human performance tops out at 84.5%, showing sufﬁcient room for future improvement. The benchmark is available at here2.
Zhang_Dimensionality-Varying_Diffusion_Process_CVPR_2023	Abstract Diffusion models, which learn to reverse a signal de- struction process to generate new data, typically require the signal at each step to have the same dimension. We argue that, considering the spatial redundancy in image signals, there is no need to maintain a high dimension- ality in the evolution process, especially in the early generation phase. To this end, we make a theoretical generalization of the forward diffusion process via signal decomposition. Concretely, we manage to decompose an image into multiple orthogonal components and control the attenuation of each component when perturbing the image. That way, along with the noise strength increasing, we are able to diminish those inconsequential components and thus use a lower-dimensional signal to represent the source, barely losing information. Such a reformulation allows to vary dimensions in both training and inference of diffusion models. Extensive experiments on a range of datasets suggest that our approach substantially reduces the computational cost and achieves on-par or even better synthesis performance compared to baseline methods. We also show that our strategy facilitates high-resolution image synthesis and improves FID of diffusion model trained on FFHQ at 1024×1024 resolution from 52.40 to 10.46. Code is available at https://github.com/damo-vilab/dvdp. ∗corresponding author. †Work performed at Alibaba DAMO Academy.	1. Introduction Diffusion models [2, 6, 9, 15, 21, 24, 28] have recently shown great potential in image synthesis. Instead of directly learning the observed distribution, it constructs a multi- step forward process through gradually adding noise onto the real data ( i.e., diffusion). After a sufficiently large number of steps, the source signal could be considered completely destroyed, resulting in a pure noise distribution that naturally supports sampling. In this way, starting from sampled noises, we can expect new instances after reversing the diffusion process step by step. As it can be seen, the above pipeline does not change the dimension of the source signal throughout the entire diffusion process [6,26,28]. It thus requires the reverse pro- cess to map a high-dimensional input to a high-dimensional output at every single step, causing heavy computation overheads [10, 22]. However, images present a measure of spatial redundancy [4] from the semantic perspective ( e.g., an image pixel could usually be easily predicted according to its neighbours). Given such a fact, when the source signal is attenuated to some extent along with the noise strength growing, it should be possible to get replaced by a lower-dimensional signal. We therefore argue that there is no need to follow the source signal dimension along the entire distribution evolution process, especially at early steps ( i.e., steps close to the pure noise distribution) for coarse generation. 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. 14307 Forward process Reverse processDVDPDDPM Figure 2. Conceptual comparison between DDPM [6] and our proposed DVDP, where our approach allows using a varying dimension in the diffusion process. In this work, we propose dimensionality-varying diffu- sion process (DVDP), which allows dynamically adjusting the signal dimension when constructing the forward path. The varying dimensionality concept is shown in Fig. 2. For this purpose, we first decompose an image into multiple orthogonal components, each of which owns dimension lower than the original data. Then, based on such a decomposition, we theoretically generalize the conventional diffusion process such that we can control the attenuation of each component when adding noise. Thanks to this refor- mulation, we manage to drop those inconsequential com- ponents after the noise strength reaches a certain level, and thus represent the source image using a lower-dimensional signal with little information lost. The remaining diffusion process could inherit this dimension and apply the same technique to further reduce the dimension. We evaluate our approach on various datasets, including objects, human faces, animals, indoor scenes, and outdoor scenes. Experimental results suggest that DVDP achieves on-par or even better synthesis performance than baseline models on all datasets. More importantly, DVDP relies on much fewer computations, and hence speeds up both train- ing and inference of diffusion models. We also demonstrate the effectiveness of our approach in learning from high- resolution data. For example, we are able to start from a 64×64noise to produce an image under 1024×1024 resolu- tion. With FID [5] as the evaluation metric, our 1024×1024 model trained on FFHQ improves the baseline [28] from 52.40 to 10.46. All these advantages benefit from using a lower-dimensional signal, which reduces the computational cost and mitigates the optimization difficulty.
Zhang_Regularized_Vector_Quantization_for_Tokenized_Image_Synthesis_CVPR_2023	Abstract Quantizing images into discrete representations has been a fundamental problem in unified generative modeling. Predominant approaches learn the discrete representation either in a deterministic manner by selecting the best- matching token or in a stochastic manner by sampling from a predicted distribution. However, deterministic quantiza- tion suffers from severe codebook collapse and misalign- ment with inference stage while stochastic quantization suf- fers from low codebook utilization and perturbed recon- struction objective. This paper presents a regularized vec- tor quantization framework that allows to mitigate above issues effectively by applying regularization from two per- spectives. The first is a prior distribution regularization which measures the discrepancy between a prior token dis- tribution and the predicted token distribution to avoid code- book collapse and low codebook utilization. The second is a stochastic mask regularization that introduces stochastic- ity during quantization to strike a good balance between in- ference stage misalignment and unperturbed reconstruction objective. In addition, we design a probabilistic contrastive loss which serves as a calibrated metric to further miti- gate the perturbed reconstruction objective. Extensive ex- periments show that the proposed quantization framework outperforms prevailing vector quantization methods con- sistently across different generative models including auto- regressive models and diffusion models.	1. Introduction With the prevalence of multi-modal image synthesis [3, 23, 37, 39] and Transformers [31], unifying data mod- eling regardless of data modalities has attracted increas- ing interest from the research communities. Aiming for a generic data representation across different data modalities, discrete representation learning [21, 25] plays a significant role in the unified modeling. In particular, vector quantiza- tion models (e.g., VQ-V AE [21] and VQ-GAN [8]) emerge as a promising family for learning generic image represen- tations by discretizing images into discrete tokens. With the *Corresponding author, E-mail: [email protected] VQ-GANGumbel-VQRegularized Quantization Figure 1. Visualization of codebook (first row) and illustration of codebook utilization (second row) on ADE20K dataset [42]. VQ- GAN [8] severely suffers from codebook collapse as most code- book embeddings are invalid values. Gumbel-VQ [2] learns valid values for all codebook embeddings, while only a small number of embeddings are actually used for quantization as illustrated in codebook utilization. As a comparison, the proposed regularized quantization prevents codebook collapse and achieves full code- book utilization. The codebook visualization method is provided in the supplementary file. tokenized representation, generative models such as auto- regressive model [8, 9] and diffusion model [6, 12] can be applied to accommodate the dependency of the sequential tokens for image generation, which is referred as tokenized image synthesis under this context. Vector quantization models can be broadly grouped into deterministic quantization and stochastic quantization ac- cording to the selection of discrete tokens. Specifically, typical deterministic methods like VQ-GAN [8] directly se- lect the best-matching token via Argmin or Argmax, while stochastic methods like Gumbel-VQ [2] select a token by stochastically sampling from a predicted token distribution. On the other hand, deterministic quantization suffers from codebook collapse [26], a well-known problem where large portion of codebook embeddings are invalid with near-zero values as shown in Fig. 1 (first row). In addition, determin- istic quantization is misaligned with the inference stage of generative modeling, where the tokens are usually randomly 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. 18467 sampled instead of selecting the best matching one. In- stead, stochastic quantization samples tokens according to a predicted token distribution with Gumbel-Softmax [2, 14], which allows to avoid codebook collapse and mitigate infer- ence misalignment. However, although most codebook em- beddings are valid values in stochastic quantization, only a small part is actually utilized for vector quantization as shown in Fig. 1 (second row), which is dubbed as low code- book utilization. Besides, as stochastic methods randomly sample tokens from a distribution, the image reconstructed from the sampled tokens is usually not well aligned with the original image, leading to perturbed reconstruction ob- jective and unauthentic image reconstruction. In this work, we introduce a regularized quantization framework that allows to prevent above problems effec- tively via regularization from two perspectives. Specifi- cally, to avoid codebook collapse and low codebook uti- lization where only a small number of codebook embed- dings are valid or used for quantization, we introduce a prior distribution regularization by assuming a uniform distribution as the prior for token distribution. As the pos- terior token distribution can be approximated by the quanti- zation results, we can measure the discrepancy between the prior token distribution and posterior token distribution. By minimizing the discrepancy during training, the quantiza- tion process is regularized to use all the codebook embed- dings, which prevents the predicted token distribution from collapse into a small number of codebook embeddings. As deterministic quantization suffers from inference stage misalignment and stochastic quantization suffers from perturbed reconstruction objective, we introduce a stochas- tic mask regularization to strike a good balance between them. Specifically, the stochastic mask regularization ran- domly masks certain ratio of regions for stochastic quanti- zation, while leaving the unmasked regions for determinis- tic quantization. This introduces uncertainty for the selec- tion of tokens and results of quantization, which narrows the gap with the inference stage of generative modelling where tokens are selected randomly. We also conduct thorough and comprehensive experiments to analyze the selection of masking ratio for optimal image reconstruction and genera- tion. On the other hand, with the randomly sampled tokens, the stochastically quantized region will suffer from per- turbed reconstruction objective. The perturbed reconstruc- tion objective mainly results from the target for perfect re- construction of the original image from randomly sampled tokens. Instead of naively enforcing a perfect image re- construction with L1 loss, we introduce a contrastive loss for elastic image reconstruction, which mitigates the per- turbed reconstruction objective significantly. Similar to PatchNCE [22, 41], the contrastive loss treats the patch at the same spatial location as positive pairs and others as neg-ative pairs. By pushing the positive pairs closer and pulling negative pairs away, the elastic image reconstruction can be achieved. Another issue with the randomly sampled tokens is that they tend to introduce perturbation of different scales in the reconstruction objective, We thus introduce a Proba- bilistic Contrastive Loss (PCL) that adjusts the pulling force of different regions according to the discrepancy between the sampled token embedding and the best-matching token embedding. The contributions of this work can be summarized in three aspects. First, we present a regularized quantization framework that introduces a prior distribution regulariza- tion to prevent codebook collapse and low codebook utiliza- tion. Second, we propose a stochastic mask regularization which mitigates the misalignment with the inference stage of generative modelling. Third, we design a probabilistic contrastive loss that achieves elastic image reconstruction and mitigates the perturbed objective adaptively for differ- ent regions with stochastic quantization.
Zhang_PointDistiller_Structured_Knowledge_Distillation_Towards_Efficient_and_Compact_3D_Detection_CVPR_2023	Abstract The remarkable breakthroughs in point cloud represen- tation learning have boosted their usage in real-world ap- plications such as self-driving cars and virtual reality. How- ever, these applications usually have a strict requirement for not only accurate but also efficient 3D object detec- tion. Recently, knowledge distillation has been proposed as an effective model compression technique, which trans- fers the knowledge from an over-parameterized teacher to a lightweight student and achieves consistent effectiveness in 2D vision. However, due to point clouds’ sparsity and irregularity, directly applying previous image-based knowl- edge distillation methods to point cloud detectors usually leads to unsatisfactory performance. To fill the gap, this paper proposes PointDistiller, a structured knowledge dis- tillation framework for point clouds-based 3D detection. Concretely, PointDistiller includes local distillation which extracts and distills the local geometric structure of point clouds with dynamic graph convolution and reweighted learning strategy, which highlights student learning on the crucial points or voxels to improve knowledge distillation efficiency. Extensive experiments on both voxels-based and raw points-based detectors have demonstrated the effective- ness of our method over seven previous knowledge distilla- tion methods. For instance, our 4 ×compressed PointPillars student achieves 2.8 and 3.4 mAP improvements on BEV and 3D object detection, outperforming its teacher by 0.9 and 1.8 mAP , respectively. Codes are available in https: //github.com/RunpeiDong/PointDistiller .	1. Introduction The growth in large-scale lidar datasets [14] and the achievements in end-to-end 3D representation learning [46, 47] have boosted the developments of point cloud based seg- mentation, generation, and detection [25, 48]. As one of the essential tasks of 3D computer vision, 3D object detection †The first two authors contribute equally. This work is done during the internship of L. Zhang in DIDI. K. Ma is the corresponding author. 60 62 64 66 68 70 72 74 76 78 Teacher (n/a)Student (4X)Student (16X)Teacher (n/a)Student (4X)Student (16X)Teacher (n/a)Student (8X)Student (16X)BEV Detection PointPillars SECOND PointRCNN 51 56 61 66 71 76 Teacher (n/a)Student (4X)Student (16X)Teacher (n/a)Student (4X)Student (16X)Teacher (n/a)Student (8X)Student (16X)3D Detection PointPillars SECOND PointRCNNmAP on KITTI mAP on KITTI Figure 1. Experimental results (mAP of moderate difficulty) of our methods on 4 ×, 8×, and 16 ×compressed students on KITTI. The area of dash lines indicates the benefits of knowledge distillation. plays a fundamental role in real-world applications such as autonomous driving cars [3, 6, 14] and virtual reality [43]. However, recent research has shown a growing discrepancy between cumbersome 3D detectors that achieve state-of-the- art performance and lightweight 3D detectors which are affordable in real-time applications on edge devices. To ad- dress this problem, sufficient model compression techniques have been proposed, such as network pruning [18,35,37,73], quantization [8,12,40], lightweight model design [21,38,51], and knowledge distillation [20]. Knowledge distillation, which aims to improve the per- formance of a lightweight student model by training it to mimic a pre-trained and over-parameterized teacher model, has evolved into one of the most popular and effective model compression methods in both computer vision and natu- ral language processing [20, 50, 52, 66]. Sufficient theoret- ical and empirical results have demonstrated its effective- ness in image-based visual tasks such as image classifica- tion [20, 50], semantic segmentation [33] and object detec- tion [1, 5, 28, 71]. However, compared with images, point clouds have their properties: (i) Point clouds inherently lack topological information, which makes recovering the local topology information crucial for the visual tasks [26, 39, 65]. (ii) Different from images that have a regular structure, point clouds are irregularly and sparsely distributed in the metric space [13, 15]. These differences between images and point clouds have hindered the image-based knowledge distillation methods from achieving satisfactory performance on point clouds and also raised the requirement to design specific knowledge dis- tillation methods for point clouds. Recently, a few methods have been proposed to apply knowledge distillation to 3D 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. 21791 ㊫࡛ 1 23 45 6 7 8 9 10 11 12 13 14 14 15 16 17+68 % 16 % 7% 3% Number of Points in the Voxel Ratio of Voxels 25 %50 %75 %Figure 2. Distribution of the voxels with different number of points inside them. V oxels with no points are not included here. detection [17, 53]. However, most of these methods focus on the choice of student-teacher in a multi-modal setting, e.g., teaching point clouds-based student detectors with an images-based teacher or vice versa, and still ignore the pe- culiar properties of point clouds. To address this problem, we propose a structured knowledge distillation framework named PointDistiller, which involves local distillation to distill teacher knowledge in the local geometric structure of point clouds, and reweighted learning strategy to handle the sparsity of point clouds by highlight student learning on the relatively more crucial voxels or points. Local Distillation Sufficient recent studies show that captur- ing and making use of the semantic information in the local geometric structure of point clouds have a crucial impact on point cloud representation learning [47, 64]. Hence, instead of directly distilling the backbone feature of teacher detec- tors to student detectors, we propose local distillation, which firstly clusters the local neighboring voxels or points with KNN (K-Nearest Neighbours), then encodes the semantic information in local geometric structure with dynamic graph convolutional layers [64], and finally distill them from teach- ers to students. Hence, the student detectors can inherit the teacher’s ability to understand point clouds’ local geometric information and achieve better detection performance. Reweighted Learning Strategy One of the mainstream meth- ods for processing point clouds is to convert them into volu- metric voxels and then encode them as regular data. How- ever, due to the sparsity and the noise in point clouds, most of these voxels contain only a single point. For instance, as shown in Figure 2, on the KITTI dataset, around 68% voxels in point clouds contain only one point, which has a high probability of being a noise point. Hence, the representative features in these single-point voxels have relatively lower im- portance in knowledge distillation compared with the voxels which contain multiple points. Motivated by this observation, we propose a reweighted learning strategy, which highlights student learning on the voxels with multiple points by giv- ing them larger learning weights. Besides, the similar idea can also be easily extended to raw points-based detectors to highlight knowledge distillation on the points with more considerable influence on the prediction.Extensive experiments on both voxels-based and raw- points based detectors have been conducted to demonstrate the effectiveness of our method over the previous seven knowledge distillation methods. As shown in Figure 1, on PointPillars and SECOND detectors, our method leads to 4 × compression and 0.9 ∼1.8 mAP improvements at the same time. On PointRCNN, our method leads to 8 ×compression with only 0.2 BEV mAP drop. Our main contributions be summarized as follows. •We propose local distillation , which firstly encodes the local geometric structure of point clouds with dynamic graph convolution and then distills them to students. •We propose reweighted learning strategy to handle the sparsity and noise in point clouds. It highlights stu- dent learning on the voxels, which have more points inside them, by giving them higher learning weights in knowledge distillation. •Extensive experiments on both voxels-based and raw points-based detectors have been conducted to demon- strate the performance of our method over seven previ- ous methods. Besides, we have released our codes to promote future research.
Yang_Vid2Seq_Large-Scale_Pretraining_of_a_Visual_Language_Model_for_Dense_CVPR_2023	Abstract In this work, we introduce Vid2Seq, a multi-modal single-stage dense event captioning model pretrained on narrated videos which are readily-available at scale. The Vid2Seq architecture augments a language model with spe- cial time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output se- quence. Such a unified model requires large-scale training data, which is not available in current annotated datasets. We show that it is possible to leverage unlabeled narrated videos for dense video captioning, by reformulating sen- tence boundaries of transcribed speech as pseudo event boundaries, and using the transcribed speech sentences as pseudo event captions. The resulting Vid2Seq model pre- trained on the YT-Temporal-1B dataset improves the state of the art on a variety of dense video captioning bench- marks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the tasks of video para- graph captioning and video clip captioning, and to few-shot settings. Our code is publicly available at [1].	1. Introduction Dense video captioning requires the temporal localiza- tion and captioning of all events in an untrimmed video [45, 98, 127]. This differs from standard video captioning [62, 69, 79], where the goal is to produce a single caption for a given short video clip. Dense captioning is significantly more difficult, as it raises the additional complexity of lo- calizing the events in minutes-long videos. However, it also *This work was done when the first author was an intern at Google.benefits from long-range video information. This task is potentially highly useful in applications such as large-scale video search and indexing, where the video content is not segmented into clips. Existing methods mostly resort to two-stage ap- proaches [36, 45, 96], where events are first localized and then captioned. To further enhance the inter-task interac- tion between event localization and captioning, some ap- proaches have introduced models that jointly solve the two tasks [19,98,127]. However, often these approaches still re- quire task-specific components such as event counters [98]. Furthermore, they exclusively train on manually annotated datasets of limited size [34, 45, 126], which makes it diffi- cult to effectively solve the task. To address these issues, we take inspiration from recent sequence-to-sequence mod- els pretrained on Web data which have been successful on a wide range of vision and language tasks [3,10,12,101,113]. First, we propose a video language model, called Vid2Seq. We start from a language model trained on Web text [77] and augment it with special time tokens that repre- sent timestamps in the video. Given video frames and tran- scribed speech inputs, the resulting model jointly predicts all event captions and their corresponding temporal bound- aries by generating a single sequence of discrete tokens, as illustrated in Figure 1 (right). Such a model therefore has the potential to learn multi-modal dependencies between the different events in the video via attention [90]. However this requires large-scale training data, which is not avail- able in current dense video captioning datasets [34,45,126]. Moreover, collecting manual annotations of dense captions for videos is expensive and prohibitive at scale. 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. 10714 Hence we propose to pretrain Vid2Seq by leveraging un- labeled narrated videos which are readily-available at scale. To do this, we reformulate sentence boundaries of tran- scribed speech as pseudo event boundaries, and use the tran- scribed speech sentences as pseudo event captions. We then pretrain Vid2Seq with a generative objective, that requires predicting the transcribed speech given visual inputs, and a denoising objective, which masks spans of transcribed speech. Note that transcribed speech may not describe the video content faithfully, and is often temporally misaligned with the visual stream [31, 42, 70]. For instance, from the example in Figure 1 (left), one can understand that the grey skier has descended a slope from the last speech sentence which is said after he actually descended the slope. Intu- itively, Vid2Seq is particularly suited for learning from such noisy supervision as it jointly models allnarrations and the corresponding timestamps in the video. We demonstrate the effectiveness of our pretrained model through extensive experiments. We show the im- portance of pretraining on untrimmed narrated videos, the ability of Vid2Seq to use both the visual and speech modali- ties, the importance of the pretraining objectives, the benefit of joint caption generation and localization, as well as the importance of the language model size and the scale of the pretraining dataset. The pretrained Vid2Seq model achieves state-of-the-art performance on various dense video cap- tioning benchmarks [34, 45, 126]. Our model also excels at generating paragraphs of text describing the video: with- out using ground-truth event proposals at inference time, our model outperforms all prior approaches including those that rely on such proposals [49,75,124]. Moreover, Vid2Seq generalizes well to the standard task of video clip cap- tioning [8, 105]. Finally, we introduce a new few-shot dense video captioning setting in which we finetune our pre- trained model on a small fraction of the downstream train- ing dataset and show benefits of Vid2Seq in this setting. In summary, we make the following contributions: (i)We introduce Vid2Seq for dense video captioning. Given multi-modal inputs (transcribed speech and video), Vid2Seq predicts a single sequence of discrete tokens that includes caption tokens interleaved with special time to- kens that represent event timestamps. (ii)We show that transcribed speech and corresponding timestamps in unla- beled narrated videos can be effectively used as a source of weak supervision for dense video captioning. (iii) Fi- nally, our pretrained Vid2Seq model improves the state of the art on three dense video captioning datasets (YouCook2, ViTT, ActivityNet Captions), two video paragraph caption- ing benchmarks (YouCook2, ActivityNet Captions) and two video clip captioning datasets (MSR-VTT, MSVD), and also generalizes well to few-shot settings. Our code implemented in Jax and based on the Scenic library [18] is publicly released at [1].
Yao_DetCLIPv2_Scalable_Open-Vocabulary_Object_Detection_Pre-Training_via_Word-Region_Alignment_CVPR_2023	Abstract This paper presents DetCLIPv2, an efficient and scalable training framework that incorporates large-scale image- text pairs to achieve open-vocabulary object detection (OVD). Unlike previous OVD frameworks that typically rely on a pre-trained vision-language model (e.g., CLIP) or ex- ploit image-text pairs via a pseudo labeling process, Det- CLIPv2 directly learns the fine-grained word-region align- ment from massive image-text pairs in an end-to-end man- ner. To accomplish this, we employ a maximum word-region similarity between region proposals and textual words to guide the contrastive objective. To enable the model to gain localization capability while learning broad concepts, Det- CLIPv2 is trained with a hybrid supervision from detection, grounding and image-text pair data under a unified data formulation. By jointly training with an alternating scheme and adopting low-resolution input for image-text pairs, Det- CLIPv2 exploits image-text pair data efficiently and ef- fectively: DetCLIPv2 utilizes 13 ×more image-text pairs than DetCLIP with a similar training time and improves †Corresponding author: [email protected], [email protected]. With 13M image-text pairs for pre-training, DetCLIPv2 demonstrates superior open-vocabulary detec- tion performance, e.g., DetCLIPv2 with Swin-T backbone achieves 40.4% zero-shot AP on the LVIS benchmark, which outperforms previous works GLIP/GLIPv2/DetCLIP by 14.4/11.4/4.5% AP , respectively, and even beats its fully- supervised counterpart by a large margin.	1. Introduction Traditional object detection frameworks [6,35,36,57] are typically trained to predict a set of predefined categories, which fails to meet the demand of many downstream appli- cation scenarios that require to detect arbitrary categories (denoted as open-vocabulary detection, OVD). For exam- ple, a robust autonomous driving system requires accurate predictions for all classes of objects on the road [26]. Ex- tending traditional object detectors to adapt these scenar- ios needs tremendous human effort for extra instance-level bounding-box annotations, especially for rare classes. To obtain an open-vocabulary detector without the expensive annotation process, the central question we should ask is: where does knowledge about unseen categories come from ? 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. 23497 Recent works [16,44,51] try to achieve open-vocabulary object detection by transferring knowledge from a pre- trained vision-language (VL) model [20, 33, 49]. E.g., ViLD [16] distills the CLIP’s [33] image embeddings of cropped proposals into the proposal features of a detection model. However, these solutions suffer from the domain gap problem: VL models are typically pre-trained with an image-level supervision using a fixed resolution input, which are not capable of recognizing objects with various scales in the detection task, especially for small objects. Another line of work resorts to exploiting massive image-text pairs crawled from the Internet. To utilize the image-text pair data without instance-level annotation, ap- proaches [13, 14, 19, 27, 48, 54] generate pseudo-bounding- box labels following a self-training paradigm [40] or based on a pre-trained VL model [33]. However, their final per- formance is restricted by the quality of pseudo-labels pro- vided by a detector trained with limited human-annotated concepts or a VL model suffering from the aforementioned domain gap problem. Besides, using high-resolution inputs similar to detection data for massive image-text pairs will impose a huge computational burden on training, prevent- ing us from further scaling up image-text pairs. To address the above issues, we present DetCLIPv2, an end-to-end open-vocabulary detection pre-training frame- work that effectively incorporates large-scale image-text pairs. DetCLILPv2 simultaneously learns localization ca- pability and knowledge of broad concepts without relying on a teacher model to provide pseudo labels. Specifically, we perform joint training with heterogeneous data from multiple sources, including detection [38], grounding [22] and image-text pairs [7, 39], under a unified data formula- tion. To enable image-text pairs without instance-level an- notations to facilitate learning of detection, inspired by [49], we employ an optimal matching-based set similarity be- tween visual regions and textual concepts to guide the con- trastive learning. By alternating different types of data for training, we enable a “flywheel effect”: learning from de- tection data provides accurate localization, which helps ex- tract representative regions for contrastive learning, while contrastive learning from image-text pairs helps recognize broader concepts, which further improves the localization of unseen categories. As the training goes on, the detector learns to locate and recognize increasingly rich concepts. Furthermore, to relief the computation burden brought by large-scale image-text pairs, we adopt a low-resolution input for image-text pair data, which significantly improves the training efficiency. This is a reasonable design since the caption of image-text pair data typically describes only the main objects appearing in the image, which alleviates the necessity of high-resolution training. Benefiting from the effective designs, DetCLIPv2 demonstrates superior open-vocabulary detection perfor- Image & Caption Pseudo Box Annotations (b) Pseudo label generation (a) Training with distillation (c) Training with fine-grained word-region alignment (ours) Image Caption Detector Text Encoder... ...Region-level Embed. Word-level Embed.Fine-grained Contrastive LossWord-Region SimilarityImage Detector ... Region Embed.... DetectorVL model DistillVL model ...Cropped Proposals Image Embed.Figure 2. Different OVD training paradigms. (a) Distilling knowledge from a pre-trained VL model [16]. (b) Exploiting image-text pairs via pseudo labeling [27]. (c) Our end-to-end joint training eliminates complex multi-stage training schemes, allow- ing for mutual benefits in learning from different types of data. mance and promising scaling behavior. E.g., compared to the prior work DetCLIP [48], DetCLIPv2 is able to exploit 13 ×more image-text pairs while requiring only a similar training time. Using the vanilla ATSS [53] as the detector, DetCLIPv2 with Swin-T backbone achieves 40.4% zero-shot AP on the LVIS [17] benchmark, sur- passing previous works GLIP [27]/GLIPv2 [52]/DetCLIP [48] by 14.4/11.4/4.5% AP, respectively. DetCLIPv2 also exhibits great generalization when transferring to down- stream tasks, e.g., it achieves SoTA fine-tuning performance on LVIS and ODinW13 [27]. We present a possibility of achieving open-world detection by incorporating large- scale image-text pairs and hope it will enlighten the commu- nity to explore a similar successful trajectory to CLIP [33].
Xu_Q-DETR_An_Efficient_Low-Bit_Quantized_Detection_Transformer_CVPR_2023	Abstract The recent detection transformer (DETR) has ad- vanced object detection, but its application on resource- constrained devices requires massive computation and memory resources. Quantization stands out as a solution by representing the network in low-bit parameters and op- erations. However, there is a significant performance drop when performing low-bit quantized DETR (Q-DETR) with existing quantization methods. We find that the bottle- necks of Q-DETR come from the query information distor- tion through our empirical analyses. This paper addresses this problem based on a distribution rectification distillation (DRD). We formulate our DRD as a bi-level optimization problem, which can be derived by generalizing the informa- tion bottleneck (IB) principle to the learning of Q-DETR. At the inner level, we conduct a distribution alignment for the queries to maximize the self-information entropy. At the upper level, we introduce a new foreground-aware query matching scheme to effectively transfer the teacher informa- tion to distillation-desired features to minimize the condi- tional information entropy. Extensive experimental results show that our method performs much better than prior arts. For example, the 4-bit Q-DETR can theoretically acceler- ate DETR with ResNet-50 backbone by 6.6 ×and achieve 39.4% AP , with only 2.6% performance gaps than its real- valued counterpart on the COCO dataset1.	1. Introduction Inspired by the success of natural language processing (NLP), object detection with transformers (DETR) has been introduced to train an end-to-end detector via a transformer encoder-decoder [4]. Unlike early works [22, 31] that often employ convolutional neural networks (CNNs) and require post-processing procedures, e.g., non-maximum suppres- sion (NMS), and hand-designed sample selection, DETR treats object detection as a direct set prediction problem. †Equal contribution. ∗Corresponding author: [email protected] 1Code: https://github.com/SteveTsui/Q-DETR decoder.5.co_attn.query decoder.0.co_attn.query decoder.2.co_attn.query (b) 4-bit DETR-R50 (a) Real-valued DETR-R50 Figure 1. The histogram of query values q(blue shadow) and cor- responding PDF curves (red curve) of Gaussian distribution [17], w.r.t the cross attention of different decoder layers in (a) real- valued DETR-R50, and (b) 4-bit quantized DETR-R50 (baseline). Gaussian distribution is generated from the statistical mean and variance of the query values. The query in quantized DETR-R50 bears information distortion compared with the real-valued one. Experiments are performed on the VOC dataset [9]. Despite this attractiveness, DETR usually has a tremen- dous number of parameters and float-pointing operations (FLOPs). For instance, there are 39.8M parameters taking up 159MB memory usage and 86G FLOPs in the DETR model with ResNet-50 backbone [12] (DETR-R50). This leads to an unacceptable memory and computation con- sumption during inference, and challenges deployments on devices with limited supplies of resources. Therefore, substantial efforts on network compression have been made towards efficient online inference [7, 33, 43, 44]. Quantization is particularly popular for deploying on AI chips by representing a network in low-bit formats. Yet prior post-training quantization (PTQ) for DETR [26] derives quantized parameters from pre-trained real-valued models, which often restricts the model performance in a sub-optimized state due to the lack of fine-tuning on the training data. In particular, the performance drastically drops when quantized to ultra-low bits (4-bits or less). Al- ternatively, quantization-aware training (QAT) [25, 42] per- forms quantization and fine-tuning on the training 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. 3842 (b) 4-bit DETR-R50(a) Real-valued DETR-R50 Figure 2. Spatial attention weight maps in the last decoder of (a) real-valued DETR-R50, and (b) 4-bit quantized DETR-R50. The green rectangle denotes the ground-truth bounding box. Follow- ing [29], the highlighted area denotes the large attention weights in the selected four heads in compliance with bound predic- tion. Compared to its real-valued counterpart that focuses on the ground-truth bounds, quantized DETR-R50 deviates significantly. simultaneously, leading to trivial performance degradation even with significantly lower bits. Though QAT meth- ods have been proven to be very effective in compressing CNNs [8, 27] for computer vision tasks, an exploration of low-bit DETR remains untouched. In this paper, we first build a low-bit DETR baseline, a straightforward solution based on common QAT tech- niques [2]. Through an empirical study of this baseline, we observe significant performance drops on the VOC [9] dataset. For example, a 4-bit quantized DETR-R50 us- ing LSQ [8] only achieves 76.9% AP 50, leaving a 6.4% performance gaps compared with the real-valued DETR- R50. We find that the incompatibility of existing QAT methods mainly stems from the unique attention mecha- nism in DETR, where the spatial dependencies are first con- structed between the object queries and encoded features. Then the co-attended object queries are fed into box coor- dinates and class labels by a feed-forward network. A sim- ple application of existing QAT methods on DETR leads to query information distortion, and therefore the performance severely degrades. Fig. 1 exhibits an example of informa- tion distortion in query features of 4-bit DETR-R50, where we can see significant distribution variation of the query modules in quantized DETR and real-valued version. The query information distortion causes the inaccurate focus of spatial attention, which can be verified by following [29] to visualize the spatial attention weight maps in 4-bit and real- valued DETR-R50 in Fig. 2. We can see that the quantized DETR-R50 bear’s inaccurate object localization. Therefore, a more generic method for DETR quantization is necessary. To tackle the issue above, we propose an efficient low-bit quantized DETR (Q-DETR) by rectifying the query infor- mation of the quantized DETR as that of the real-valued counterpart. Fig. 3 provides an overview of our Q-DETR, which is mainly accomplished by a distribution rectifica-tion knowledge distillation method (DRD). We find ineffec- tive knowledge transferring from the real-valued teacher to the quantized student primarily because of the information gap and distortion. Therefore, we formulate our DRD as a bi-level optimization framework established on the infor- mation bottleneck principle (IB). Generally, it includes an inner-level optimization to maximize the self-information entropy of student queries and an upper-level optimization to minimize the conditional information entropy between student and teacher queries. At the inner level, we conduct a distribution alignment for the query guided by its Gaussian- alike distribution, as shown in Fig. 1, leading to an explicit state in compliance with its maximum information entropy in the forward propagation. At the upper level, we introduce a new foreground-aware query matching that filters out low- qualified student queries for exact one-to-one query match- ing between student and teacher, providing valuable knowl- edge gradients to push minimum conditional information entropy in the backward propagation. This paper attempts to introduce a generic method for DETR quantization. The significant contributions in this paper are outlined as follows: (1) We develop the first QAT quantization framework for DETR, dubbed Q-DETR. (2) We use a bi-level optimization distillation framework, ab- breviated as DRD. (3) We observe a significant performance increase compared to existing quantized baselines.