Datasets:

bskang

/

CVPR2023_title_abstract_intro

like 0

Tertikas_Generating_Part-Aware_Editable_3D_Shapes_Without_3D_Supervision_CVPR_2023

Abstract Impressive progress in generative models and implicit representations gave rise to methods that can generate 3D shapes of high quality. However, being able to locally con- trol and edit shapes is another essential property that can unlock several content creation applications. Local control can be achieved with part-aware models, but existing meth- ods require 3D supervision and cannot produce textures. In this work, we devise PartNeRF , a novel part-aware gener- ative model for editable 3D shape synthesis that does not require any explicit 3D supervision. Our model generates objects as a set of locally defined NeRFs, augmented with an affine transformation. This enables several editing op- erations such as applying transformations on parts, mixing *Work done during internship at Stanford.parts from different objects etc. To ensure distinct, manip- ulable parts we enforce a hard assignment of rays to parts that makes sure that the color of each ray is only determined by a single NeRF . As a result, altering one part does not af- fect the appearance of the others. Evaluations on various ShapeNet categories demonstrate the ability of our model to generate editable 3D objects of improved fidelity, compared to previous part-based generative approaches that require 3D supervision or models relying on NeRFs.

1. Introduction Generating realistic and editable 3D content is a long- standing problem in computer vision and graphics that has recently gained more attention due to the increased demand for 3D objects in AR/VR, robotics and gaming applications. 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. 4466 However, manual creation of 3D models is a laborious en- deavor that requires technical skills from highly experi- enced artists and product designers. On the other hand, edit- ing 3D shapes, typically involves re-purposing existing 3D models, by manually changing faces and vertices of a mesh and modifying its respective UV-map [95]. To accommo- date this process, several recent works introduced genera- tive models that go beyond generation and allow editing the generated instances [13,18,52,55,62,77,101,116,117,124]. Shape editing involves making local changes on the shape and the appearance of different parts of an object. There- fore, having a basic understanding of the decomposition of the object into parts facilitates controlling what to edit . While Generative Adversarial Networks (GANs) [30] have emerged as a powerful tool for synthesizing photore- alistic images [7, 15, 16, 47–49], scaling them to 3D data is non-trivial as they ignore the physics of image forma- tion process.To address this, 3D-aware GANs incorporate 3D representations such as voxel grids [38, 72, 75] or com- bine them with differentiable renderers [57, 126]. While they faithfully recover the geometry and appearance, they do not allow changing specific parts of the object. Inspired by the rapid evolution of neural implicit ren- dering techniques [68], recent works [8, 32, 77, 96, 114] proposed to combine them with GANs in order to allow for multi-view-consistent generations of high quality. De- spite their impressive performance on novel view synthe- sis, their editing capabilities are limited. To this end, editing operations in the latent space have been explored [21, 42, 62, 115, 124] but these approaches lack intuitive control over the shape. By decomposing shapes into parts, other works facilitate structure-aware shape manipulations [40,70,88,111]. However, they require 3D supervision dur- ing training and can only operate on textureless shapes. To address these limitations, we devise PartNeRF, a novel part-aware generative model, implemented as an auto- decoder [5]. Our model enables part-level control, which fa- cilitates various editing operations on the shape and appear- ance of the generated instance. These operations include rigid and non-rigid transformations on the object parts, part mixing from different objects, removing/adding parts and editing the appearance of specific parts of the object. Our key idea is to represent objects using a set of locally defined Neural Radiance Fields (NeRFs) that are arranged such that the object can be plausibly rendered from a novel view. To enable part-level control, we enforce a hard as- signment between parts and rays that ensures that altering one part does not affect the shape and appearance of the others. Our model does not require 3D supervision; we only assume supervision from images and object masks captured from known cameras. We evaluate PartNeRF on various ShapeNet categories and demonstrate that it generates tex- tured shapes of higher fidelity than both part-based as wellas NeRF-based generative models. Furthermore, we show- case several editing operations, not previously possible. In summary, we make the following contributions : We propose the first part-aware generative model that parametrizes parts as NeRFs and can generate editable 3D shapes. Unlike prior part-based approaches, our model does not require explicit 3D supervision and can gener- ate textured shapes. Compared to NeRF-based generative models, our work is the first that reasons about parts and hence enables operations both on the shape and the tex- ture of the generated object. Code and data is available at https://ktertikas.github.io/part nerf.

Truong_FREDOM_Fairness_Domain_Adaptation_Approach_to_Semantic_Scene_Understanding_CVPR_2023

Abstract Although Domain Adaptation in Semantic Scene Seg- mentation has shown impressive improvement in recent years, the fairness concerns in the domain adaptation have yet to be well defined and addressed. In addition, fairness is one of the most critical aspects when deploying the segmen- tation models into human-related real-world applications, e.g., autonomous driving, as any unfair predictions could influence human safety. In this paper, we propose a novel Fairness Domain Adaptation (FREDOM) approach to se- mantic scene segmentation. In particular, from the proposed formulated fairness objective, a new adaptation framework will be introduced based on the fair treatment of class distri- butions. Moreover, to generally model the context of struc- tural dependency, a new conditional structural constraint is introduced to impose the consistency of predicted segmen- tation. Thanks to the proposed Conditional Structure Net- work, the self-attention mechanism has sufficiently modeled the structural information of segmentation. Through the ab- lation studies, the proposed method has shown the perfor- mance improvement of the segmentation models and pro- moted fairness in the model predictions. The experimental results on the two standard benchmarks, i.e., SYNTHIA → Cityscapes and GTA5 →Cityscapes, have shown that our method achieved State-of-the-Art (SOTA) performance1.

1. Introduction Semantic segmentation has achieved remarkable results in a wide range of practical problems, including scene un- derstanding, autonomous driving, and medical imaging, by using deep learning models, e.g., Convolutional Neural Net- works (CNN) [3, 4, 24], Transformers [45]. Despite the phenomenal achievement, these data-driven approaches still need to improve in treating the prediction of each class. In particular, the segmentation models typically treat unfairly between classes in the dataset according to the class distri- butions. It is known as the fairness problem of semantic 1The implementation of FREDOM is available at https : / / github.com/uark-cviu/FREDOM Figure 1. The class distributions on Cityscapes are defined for Fairness problem and Long-tail problem. Inlong-tail problem, several head classes frequently exist in the dataset, e.g., Pole, Traf- fic Light, or Sign. Still, these classes belong to a minority group in the fairness problem as their appearance on images does not oc- cupy too many pixels. Our FREDOM has promoted the fairness of models illustrated by an increase of mIoU on the minority group. segmentation. The unfair predictions of segmentation mod- els can lead to severe problems, e.g., in autonomous driving, unfair predictions may result in wrong decisions in motion planning control and therefore affect human safety. More- over, the fairness issue of segmentation models is even well observed or exaggerated when the trained models are de- ployed into new domains. Many prior works alleviate the performance drop on new domains by using unsupervised domain adaptation, but these approaches do not guarantee 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. 19988 Figure 2. Illustration of the Presence of Classes between Major (green boxes) and Minor (red boxes) Groups . Classes in the minority group typically occupy fewer pixels than the ones in the majority group (Best view in color and 2×zoom). the fairness property. There needs to be more attention on addressing the fair- ness issue in semantic segmentation under the supervised or domain adaptation settings. Besides, the definition of fair- ness in semantic segmentation needs to be better defined and, therefore, often needs clarification with the long-tail is- sue in segmentation. In particular, the long-tail problem in segmentation is typically caused by the number of existing instances of each class in the dataset [21, 44]. Meanwhile, thefairness problem in segmentation is considered for the number of pixels of each class in the dataset. Although there could be a correlation between fairness and long-tail problems, these two issues are distinct. For example, sev- eral objects constantly exist in the dataset, but their presence often occupies only tiny regions of the given image (con- taining a small number of pixels), e.g., the Pole, which is a head class in Cityscapes, accounts for over 20% of instances while the number of pixels does only less than 0.01% of pixels. Hence, upon the fairness definition, it should belong to the minor group of classes as its presence does not oc- cupy many pixels in the image. Another example is Person, which accounts for over 5%of instances, while the num- ber of pixels does only less than 0.01% of pixels. Traffic Lights or Signs also suffer a similar problem. Fig. 2 illus- trates the appearance of classes in the majority and minority groups. Therefore, although instances of these classes con- stantly exist in the dataset, these are still being mistreated by the segmentation model. Fig. 1 illustrates the class distribu- tions defined based on long-tail and fairness, respectively. Several works reduce the class imbalance effects us- ing weighted (balanced) cross entropy [13, 21, 44], focal loss [1], data augmentation or rare-class sampling tech- niques [1,19]. Still, these need to address the fairness prob- lem directly. Indeed, many prior domain adaptation meth- ods [6, 17, 28, 34, 36–39] have been used to improve the overall performance. However, these methods often ignore unfair effects produced by the model caused by the imbal- anced class distribution. Besides, in some adaptation ap- proaches using entropy minimization [29, 42], the model’s bias caused by the class imbalance between majority and minority groups is even exaggerated [7, 35]. Meanwhile, other approaches using re-weighted or focal loss [1] often assume pixel independence and then penalize the loss con- tribution of each pixel individually and ignore the structuralinformation of images. Then, pixel independence is relaxed by adopting the Markovian assumption [3,48] to model seg- mentation structures based on neighbor pixels. In the scope of our work , we are interested in addressing the fairness problem in semantic segmentation between classes under the unsupervised domain adaptation setting. It should be noted that our interested problem is practical. In real-world applications (e.g., autonomous driving), deep learning mod- els are typically deployed into new domains compared to the training dataset. Then, unsupervised domain adaptation plays a role in bridging the gap between the two domains. Contributions of This Work: This work presents a novel Unsupervised FairnessDom ain Adaptation (FREDOM) approach to semantic segmentation. To the best of our knowledge, this is one of the first works to address the fair- ness problem in semantic segmentation under the domain adaptation setting. Our contributions can be summarized as follows. First, the new fairness objective is formulated for semantic scene segmentation. Then, based on the fair- ness metric, we propose a novel fairness domain adapta- tion approach based on the fair treatment of class distribu- tions. Second, the novel Conditional Structural Constraint is proposed to model the structural consistency of segmen- tation maps. Thanks to our introduced Conditional Struc- ture Network, the spatial relationship and structure infor- mation are well modeled by the self-attention mechanism. Significantly, our structural constraint relaxes the assump- tion of pixel independence held by prior approaches and generalizes the Markovian assumption by considering the structural correlations between all pixels. Finally, our ab- lation studies have shown the effectiveness of different as- pects in our approach to the fairness improvement of seg- mentation models. Through experiments, our FREDOM has promoted the fairness property of segmentation models and achieved state-of-the-art (SOTA) performance on two standard benchmarks of unsupervised domain adaptation, i.e., SYNTHIA →Cityscapes and GTA5 →Cityscapes.

Tan_Sample-Level_Multi-View_Graph_Clustering_CVPR_2023

Abstract Multi-view clustering has hitherto been studied due to their effectiveness in dealing with heterogeneous data. De- spite the empirical success made by recent works, there still exists several severe challenges. Particularly, previous multi-view clustering algorithms seldom consider the topo- logical structure in data, which is essential for clustering data on manifold. Moreover, existing methods cannot fully explore the consistency of local structures between different views as they uncover the clustering structure in a intra- view way instead of a inter-view manner. In this paper, we propose to exploit the implied data manifold by learning the topological structure of data. Besides, considering that the consistency of multiple views is manifested in the gen- erally similar local structure while the inconsistent struc- tures are the minority, we further explore the intersections of multiple views in the sample level such that the cross-view consistency can be better maintained. We model the above concerns in a unified framework and design an efficient al- gorithm to solve the corresponding optimization problem. Experimental results on various multi-view datasets certifi- cate the effectiveness of the proposed method and verify its superiority over other SOTA approaches.

1. Introduction In many real scenarios, data are usually collected from diverse sources in various domains or described by multi- ple feature sets [1, 2]. A case in point is the image dataset, which can be represented by different visual descriptors, such as LBP, GIST, CENTRIST, HOG, SIFT and Color Mo- ment [3]. Moreover, a document can be written in differ- ent languages, and a website can be described by its link- age, page content, etc. These data are generally known as multi-view data [1, 4]. Each view contains partial and com- plementary information, any of which suffices for learning, and they all together agree to a consensus latent represen- tation [5–7]. Multi-view clustering aims to accurately par- *Corresponding author.tition data into distinct clusters according to their compati- ble and complementary information embedded in heteroge- neous features [1, 8]. In general, multi-view clustering methods can be di- vided into four categories according to the mechanisms and principles on which these methods are based. [1] ini- tiated the trend of co-training algorithm by first carrying out a co-training strategy and making the clustering re- sults on all view tend to each other as well as satisfying the broadest agreement across all views. Owing to the ef- ficiency of exploiting similarities among multiple views, multi-kernel learning is widely utilized to boost the perfor- mance of multi-view clustering methods [9–11]. Multi-task multi-view clustering [12,13] inherits the property of multi- task clustering. Each view of the data is treated with at least one related task. Moreover, inter-task knowledge is trans- ferred from one to another so that the relationship between multi-view and multi-task is fully exploited to improve the clustering outcomes. Another kind of graph-based multi- view clustering methods usually attempt to explore an opti- mal consensus graph across all views, and utilize graph-cut algorithm on the optimal graph to obtain final clustering re- sult [3, 5, 14]. A great number of multi-view clustering methods have been proposed and illustrated remarkable empirical suc- cesses up to now [15–17]. However, there are still severe drawbacks urgently need to be overcome. For one thing, existing multi-view clustering algorithms seldom consider the topological structure in data, which is essential for clus- tering data on manifold. Considering that the data sampled from real world typically lie in the nonlinear manifold, data points geometrically far from one to another may keep high consistency when they are linked by a series of consecutive neighbors. For another, they suffer from a coarse-grained problem that ignores the view correlation between samples. Note that redundancy or partial structure mistakes in certain views may lead to a sub-optimal cluster structure. Besides, traditional view-wise fusion strategy would result in a su- perimposition of redundancies, and hence acquire a more imprecise common cluster structure. Regarding the above issues, we propose to exploit 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. 23966 implied data manifold by learning the topological structure of data. Besides, considering that the consistency of multi- ple views is manifested in the generally similar local struc- ture while the inconsistent structures are the minority, we further explore the intersections of multiple views in the sample level such that the cross-view consistency can be better maintained. By leveraging the subtasks of topologi- cal relevance learning and the sample-level graph fusion in our collaborative model, each subtask is alternately boosted towards an optimal solution. Experimental results on var- ious multi-view datasets certificate the effectiveness of the proposed method.

Sun_TRACE_5D_Temporal_Regression_of_Avatars_With_Dynamic_Cameras_in_CVPR_2023

Abstract Although the estimation of 3D human pose and shape (HPS) is rapidly progressing, current methods still can- not reliably estimate moving humans in global coordinates, which is critical for many applications. This is particularly challenging when the camera is also moving, entangling human and camera motion. To address these issues, we adopt a novel 5D representation (space, time, and identity) that enables end-to-end reasoning about people in scenes. Our method, called TRACE, introduces several novel ar- *This work was done when Yu Sun was an intern at JD.com. †Corresponding author.chitectural components. Most importantly, it uses two new “maps” to reason about the 3D trajectory of people over time in camera, and world, coordinates. An additional memory unit enables persistent tracking of people even dur- ing long occlusions. TRACE is the first one-stage method to jointly recover and track 3D humans in global coordinates from dynamic cameras. By training it end-to-end, and us- ing full image information, TRACE achieves state-of-the-art performance on tracking and HPS benchmarks. The code1 and dataset2are released for research purposes. 1https://www.yusun.work/TRACE/TRACE.html 2https://github.com/Arthur151/DynaCam 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. 8856

1. Introduction The estimation of 3D human pose and shape (HPS) has many applications and there has been significant recent progress [6, 7, 21–24, 37, 38, 46, 49, 56, 57, 60]. Most meth- ods, however, reason only about a single frame at a time and estimate humans in camera coordinates . Moreover, such methods do not track people and are unable to recover their global trajectories. The problem is even harder in typical hand-held videos, which are filmed with a dynamic, mov- ing, camera. For many applications of HPS, single-frame estimates in camera coordinates are not sufficient. To cap- ture human movement and then transfer it to a new 3D scene, we must have the movement in a coherent global co- ordinate system. This is a requirement for computer graph- ics, sports, video games, and extended reality (XR). Our key insight is that most methods estimate humans in 3D, whereas the true problem is 5D. That is, a method needs to reason about 3D space, time, and subject identity. With a 5D representation, the problem becomes tractable, enabling aholistic solution that can exploit the full video to infer multiple people in a coherent global coordinate frame. As illustrated in Fig. 1, we develop a unified method to jointly regress the 3D pose, shape, identity, and global trajectory of the subjects in global coordinates from monocular videos captured by dynamic cameras (DC-videos). To achieve this, we deal with two main challenges. First, DC-videos contain both human motion and camera motion and these must be disentangled to recover the human trajec- tory in global coordinates. One idea would be to recover the camera motion relative to the rigid scene using structure- from-motion (SfM) methods (e.g. [31]). In scenes contain- ing many people and human motion, however, such meth- ods can be unreliable. An alternative approach is taken by GLAMR [58], which infers global human trajectories from local 3D human poses, without taking into account the full scene. By ignoring evidence from the full image, GLAMR fails to capture the correct global motion in common scenar- ios, such as biking, skating, boating, running on a treadmill, etc. Moreover, GLAMR is a multi-stage method, with each stage dependent on accurate estimates from the preceding one. Such approaches are more brittle than our holistic, end-to-end, method. The other challenge, as shown in the upper right corner of Fig. 1, is that severe occlusions are common in videos with multiple people. Currently, the most popular tracking strategy is to infer the association between 2D detections using a temporal prior (e.g. Kalman filter) [63]. However, in DC-videos, human motions are often irregular and can easily violate hand-crafted priors. PHALP [40] is one of the few methods to address this for 3D HPS. It uses a clas- sical, multi-stage, detection-and-tracking formulation with heuristic temporal priors. It does not holistically reason about the sequence and is not trained end-to-end.To address these issues, we reason about people using a 5D representation and capture information from the full image and the motion of the scene. This holistic reason- ing enables the reliable recovery of global human trajecto- ries and subject tracking using a single-shot method. This is more reliable than multi-stage methods because the net- work can exploit more information to solve the task and is trained end-to-end. No hand-crafted priors are needed and the network is able to share information among modules. Specifically, we develop TRACE, a unified one-stage method for Temporal Regression of Avatars with dynamic Cameras in 3D Environments. The architecture is inspired by BEV [45], which directly estimates multiple people in depth from a single image using multiple 2D maps. BEV uses a 2D map representing an imaginary, “top down”, view of the scene. This is combined with an image-centric 2D map to reason about people in 3D. Our key insight is that the idea of maps can be extended to represent how people move in 3D . With this idea, TRACE introduces three new modules to holistically model 5D human states, perform- ing multi-person temporal association, and inferring human trajectories in global coordinates; see Fig. 2. First, to construct a holistic 5D representation of the video, we extract temporal image features by fusing single- frame feature maps from the image backbone with a tem- poral feature propagation module. We also compute the optical flow between adjacent frames with a motion back- bone. The optical flow provides short-term motion features that carry information about the motion of the scene and the people. Second, to explicitly track human motions, we introduce a novel 3D motion offset map to establish the as- sociation of the same person across adjacent frames. This map contains a 3D offset vector at each position, which rep- resents the difference between the 3D positions of the same subject from the previous frame to the current frame in cam- era coordinates. We also introduce a memory unit to keep track of subjects under long-term occlusion. Note that the 3D trajectories are built in camera space, and TRACE uses a novel world motion map that transfers the trajectories to global coordinates. At each position, this map contains a 6D vector to represent the difference between the 3D posi- tions of the corresponding subject from the previous frame to the current frame and its 3D orientation in world coordi- nates. Taken together, this novel network architecture goes beyond prior work by taking information from the full video frames to address detection, pose estimation, tracking, and occlusion in a holistic network that is trained end-to-end. To enable training and evaluation of global human tra- jectory estimation from in-the-wild DC-videos, we build a new dataset, DynaCam. Since collecting global human tra- jectories and camera poses with in-the-wild DC-videos is difficult, we simulate a moving camera using publicly avail- able in-the-wild panoramic videos and regular videos cap- 8857 tured by static cameras. In this way, we create more than 500 in-the-wild DC-videos with precise camera pose anno- tations. Then we generate pseudo-ground-truth 3D human annotations via fitting [20] SMPL [32] to detected 2D pose sequences [17,54,63]. With 2D/3D human pose and camera pose annotations, we can obtain the global human trajecto- ries using the PnP algorithm [16]. This dataset is sufficient to train TRACE to deal with dynamic cameras. We evaluate TRACE on two multi-person in-the-wild benchmarks (3DPW [50] and MuPoTS-3D [35]) and our DynaCam dataset. On 3DPW, TRACE achieves the state- of-the-art (SOTA) PA-MPJPE of 37.8mm, less the cur- rent best (42.7 [26]). On MuPoTS-3D, TRACE outper- forms previous 3D-representation-based methods [39, 40] and tracking-by-detection methods [63] on tracking people under long-term occlusion. On DynaCam, TRACE outper- forms GLAMR [58] in estimating the 3D human trajectory in global coordinates from DC-videos. In summary, our main contributions are: (1) We intro- duce a 5D representation and use it to learn holistic tempo- ral cues related to both 3D human motions and the scene. (2) We introduce two novel motion offset representations to explicitly model temporal multi-subject association and global human trajectories from temporal clues in an end-to- end manner. (3) We estimate long-term 3D human motions over time in global coordinates, achieving SOTA results. (4) We collect the DynaCam dataset of DC-videos with pseudo ground truth, which facilitates the training and evaluation of global human trajectory estimation. The code and dataset are publicly available for research purposes.

Tan_Distilling_Neural_Fields_for_Real-Time_Articulated_Shape_Reconstruction_CVPR_2023

Abstract We present a method for reconstructing articulated 3D models from videos in real-time, without test-time optimiza- tion or manual 3D supervision at training time. Prior work often relies on pre-built deformable models (e.g. SMAL/SMPL), or slow per-scene optimization through dif- ferentiable rendering (e.g. dynamic NeRFs). Such methods fail to support arbitrary object categories, or are unsuit- able for real-time applications. To address the challenge of collecting large-scale 3D training data for arbitrary de- formable object categories, our key insight is to use off- the-shelf video-based dynamic NeRFs as 3D supervision to train a fast feed-forward network, turning 3D shape and motion prediction into a supervised distillation task. Our temporal-aware network uses articulated bones and blend skinning to represent arbitrary deformations, and is self- supervised on video datasets without requiring 3D shapes or viewpoints as input. Through distillation, our network learns to 3D-reconstruct unseen articulated objects at in- teractive frame rates. Our method yields higher-fidelity 3D reconstructions than prior real-time methods for animals, with the ability to render realistic images at novel view- points and poses.

1. Introduction We are interested in building high-quality animatable models of articulated 3D objects from videos in real time. One promising application is virtual and augmented real- ity, where the goal is to create high-fidelity 3D experiences from images and videos captured live by users. For rigid scenes, structure from motion (SfM) and neural rendering can be used to build accurate 3D cities and landmarks from Internet image collections [1, 20, 33]. For articulated ob- jects such as friends and pets, many works parameterize the range of motions using category-specific templates such as SMPL [18] for humans and SMAL [4] for quadruped ani- mals. Although these methods can be trained on large-scale video datasets, they rely on parametric body template mod- els built from extensive real-world 3D scans: these body models are not easy to generate for diverse categories in the wild such as clothed humans or pets with distinct morpholo- gies, which are often the focus of user content. Inspired by the breakthrough success of neural radiance fields [21], many works reconstruct arbitrary articulated ob- jects in an analysis-by-synthesis framework [16, 27, 28, 30, 36, 44] by defining time-dependent 3D warping fields and establishing long-range correspondences on top of canon- 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. 4692 ical shape and appearance models. These methods out- put high-quality reconstructions of arbitrary objects without 3D data or pre-defined templates, but the output represen- tations are scene-specific and often require hours to com- pute from scratch on unseen videos - an unacceptable cost for real-time VR/AR tasks. We are thus interested in dy- namic 3D reconstruction algorithms that achieve the best of both worlds: the speed of template-based models and the quality and generalization ability of dynamic NeRFs. To achieve this, our key insight is remarkably simple: we train category-specific feed-forward 3D predictors at scale by self-supervising them with dynamic NeRF “teachers” fit- ted to offline video data. By leveraging scene-fitted dynamic NeRFs for 3D su- pervision at scale, our method learns a feed-forward pre- dictor for appearance, 3D shape, and articulations of non- rigid objects from videos. Our learned 3D models use lin- ear blend skinning to express articulations, allowing it to be animated by manipulating bone transformations. We ad- dress three key challenges in our work: (1) how to super- vise feed-forward models with internal representations of dynamic NeRFs, (2) how to produce temporally consistent predictions of pose, articulation, and appearance, and (3) how to build efficient systems for real-time reconstruction.

Srivastava_How_You_Feelin_Learning_Emotions_and_Mental_States_in_Movie_CVPR_2023

Abstract Movie story analysis requires understanding characters’ emotions and mental states. Towards this goal, we for- mulate emotion understanding as predicting a diverse and multi-label set of emotions at the level of a movie scene and for each character. We propose EmoTx, a multimodal Transformer-based architecture that ingests videos, multi- ple characters, and dialog utterances to make joint pre- dictions. By leveraging annotations from the MovieGraphs dataset [72], we aim to predict classic emotions ( e.g. happy, angry) and other mental states ( e.g. honest, helpful). We conduct experiments on the most frequently occurring 10 and 25 labels, and a mapping that clusters 181 labels to 26. Ablation studies and comparison against adapted state- of-the-art emotion recognition approaches shows the effec- tiveness of EmoTx. Analyzing EmoTx’s self-attention scores reveals that expressive emotions often look at character to- kens while other mental states rely on video and dialog cues.

1. Introduction In the movie The Pursuit of Happyness , we see the pro- tagonist experience a roller-coaster of emotions from the lows of breakup and homelessness to the highs of getting selected for a coveted job. Such heightened emotions are of- ten useful to draw the audience in through relatable events as one empathizes with the character(s). For machines to understand such a movie (broadly, story), we argue that it is paramount to track how characters’ emotions and men- tal states evolve over time. Towards this goal, we lever- age annotations from MovieGraphs [72] and train models to watch the video, read the dialog, and predict the emo- tions and mental states of characters in each movie scene. Emotions are a deeply-studied topic. From ancient Rome and Cicero’s 4-way classification [60], to modern brain re- search [33], emotions have fascinated humanity. Psychol- ogists use of Plutchik’s wheel [53] or the proposal of uni- versality in facial expressions by Ekman [18], structure has been provided to this field through various theories. Affec- tive emotions are also grouped into mental (affective, be- - And he's very bitter . - And he's just gonna walk out the door and never know why she's just lying there on the couch... - That's a chick's movie. - I would say so. - What kind of a person would write to someone they heard on the radio?- Stop it. - Richard Jaeckel had on this shiny helmet 'cause he was the M.P . - No more. Oh, God, I love that movie. CBA ??? ??? ???Figure 1. Multimodal models and multi-label emotions are neces- sary for understanding the story. A: What character emotions can we sense in this scene? Is a single label enough? B: Without the dialog, can we try to guess the emotions of the Sergeant and the Soldier. C: Is it possible to infer the emotions from the characters’ facial expressions (without subtitles and visual background) only? Check the footnote below for the ground-truth emotion labels for these scenes and the supplement for an explanation of the story. havioral, and cognitive) or bodily states [13]. A recent work on recognizing emotions with visual con- text, Emotic [31] identifies 26 label clusters and proposes amulti-label setup wherein an image may exhibit multi- ple emotions ( e.g.peace, engagement ). An alternative to the categorical space, valence, arousal, and dominance are also used as three continuous dimensions [31]. Predicting a rich set of emotions requires analyzing multiple contextual modalities [31, 34, 44]. Popular directions in multimodal emotion recognition are Emotion Recognition in Conver- sations (ERC) that classifies the emotion for every dialog utterance [42,54,83]; or predicting a single valence-activity score for short ∼10s movie clips [4, 45]. We operate at the level of a movie scene : a set of shots telling a sub-story, typically at one location, among a de- fined cast, and in a short time span of 30to60 s. Thus, scenes are considerably longer than single dialogs [54] or Ground-truth emotions and mental states portrayed in movie scenes in Fig. 1: A: excited, curious, confused, annoyed, alarmed; B: shocked, confident; C: happy, excited, amused, shocked, confident, nervous. 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. 2517 movie clips in [4]. We predict emotions and mental states for all characters in the scene and also by accumulating la- bels at the scene level. Estimation on a larger time window naturally lends itself to multi-label classification as charac- ters may portray multiple emotions simultaneously ( e.g.cu- rious andconfused ) or have transitions due to interactions with other characters ( e.g.worried tocalm ). We perform experiments with multiple label sets: Top- 10 or 25 most frequently occurring emotion labels in MovieGraphs [72] or a mapping to the 26 labels in the Emotic space, created by [45]. While emotions can broadly be considered as part of mental states, for this work, we con- sider that expressed emotions are apparent by looking at the character, e.g.surprise, sad, angry ; and mental states are la- tent and only evident through interactions or dialog, e.g.po- lite, determined, confident, helpful1. We posit that classifi- cation in a rich label space of emotions requires looking at multimodal context as evident from masking context in Fig. 1. To this end, we propose EmoTx that jointly models video frames, dialog utterances, and character appearance. We summarize our contributions as follows: (i) Building on rich annotations from MovieGraphs [72], we formulate scene and per-character emotion and mental state classifi- cation as a multi-label problem. (ii) We propose a multi- modal Transformer-based architecture EmoTx that predicts emotions by ingesting all information relevant to the movie scene. EmoTx is also able to capture label co-occurrence and jointly predicts all labels. (iii) We adapt several pre- vious works on emotion recognition for this task and show that our approach outperforms them all. (iv) Through analy- sis of the self-attention mechanism, we show that the model learns to look at relevant modalities at the right time. Self- attention scores also shed light on our model’s treatment of expressive emotions vs. mental states.

Tripathi_Edges_to_Shapes_to_Concepts_Adversarial_Augmentation_for_Robust_Vision_CVPR_2023

Abstract Recent work has shown that deep vision models tend to be overly dependent on low-level or “texture” features, leading to poor generalization. Various data augmentation strategies have been proposed to overcome this so-called texture bias in DNNs. We propose a simple, lightweight adversarial augmentation technique that explicitly incen- tivizes the network to learn holistic shapes for accurate prediction in an object classification setting. Our augmen- tations superpose edgemaps from one image onto another image with shuffled patches, using a randomly determined mixing proportion, with the image label of the edgemap im- age. To classify these augmented images, the model needs to not only detect and focus on edges but distinguish between relevant and spurious edges. We show that our augmenta- tions significantly improve classification accuracy and ro- bustness measures on a range of datasets and neural ar- chitectures. As an example, for ViT-S, We obtain absolute gains on classification accuracy gains up to 6%. We also obtain gains of up to 28% and 8.5% on natural adversarial and out-of-distribution datasets like ImageNet-A (for ViT- B) and ImageNet-R (for ViT-S), respectively. Analysis us- ing a range of probe datasets shows substantially increased shape sensitivity in our trained models, explaining the ob- served improvement in robustness and classification accu- racy.

1. Introduction A growing body of research catalogues and analyzes ap- parent failure modes of deep vision models. For instance, work on texture bias [1,7,12] suggests that image classifiers are overdependent on textural cues and fail against simple (adversarial) texture substitutions. Relatedly, the idea of simplicity bias [25] captures the tendency of deep models to use weakly predictive “simple” features such as color or texture, even in the presence of strongly predictive com- plex features. In psychology & neuroscience, too, evidence *Work done at Google Research India. ImageNet-C mCE (Lower is Better)ImageNet-A Accuracy (Higher is Better) Shape FactorR50 R50R101 R101R152 R152R50(T) R50(T)R101(T) R152(T) R101(T)R152(T)ViT-S ViT-SViT-S(T) ViT-S(T)ViT-S(E) ViT-S(E)R152(E) R152(E)R101(E) R101(E)R50(E) R50(E)Figure 1. Comparison of the models on robustness and shape- bias. The shape factor gives the fraction of dimensions that en- code shape cues [17]. Backbone(T) denotes texture shape debiased (TSD) models [21]. In comparison, EL EAS denoted by Back- bone(E) is more shape biased and shows better performance on ImageNet-C and ImageNet-A datasets. suggests that deep networks focus more on “local” features rather than global features and differ from human behav- ior in related tasks [19]. More broadly speaking, there is a mismatch between the cognitive concepts and associated world knowledge implied by the category labels in image datasets such as Imagenet and the actual information con- tent made available to a model via one-hot vectors encod- ing these labels. In the face of under-determined learning problems, we need to introduce inductive biases to guide 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. 24470 Figure 2. Representative performance comparison of our model, with the ‘Debiased’ models (TSD [21]) on ImageNet-A, R, C, and Sketch datasets. The models trained using EL EAS show improved performance on out-of-distribution robustness datasets. The large performance improvement on the ImageNet-A dataset indicates better robustness to natural adversarial examples. the learning process. To this end, Geirhos et al. [7] pro- posed a data augmentation method wherein the texture of an image was replaced with that of a painting through styl- ization. Follow-on work improved upon this approach by replacing textures from other objects (instead of paintings) and teaching the model to separately label the outer shape and the substituted texture according to their source image categories [21]. Both these approaches discourage overde- pendence on textural features in the learned model; how- ever, they do not explicitly incentivize shape recognition. We propose a lightweight adversarial augmentation technique EL EAS (EdgeLearning for Shape sensitivity) that is designed to increase shape sensitivity in vision mod- els. Specifically, we augment a dataset with superposi- tions of random pairs of images from the dataset, where one image is processed to produce an edge map, and the other is modified by shuffling the location of image patches within the image. The two images are superposed us- ing a randomly sampled relative mixing weight (similar to Mixup [30]), and the new superposed image is assigned the label of the edgemap image. EL EAS is designed to specif- ically incentivize not only edge detection but shape sensi- tivity: 1) classifying the edgemap image requires the model to extract and exploit edges – the only features available in the image, 2) distinguishing the edgemap object cate- gory from the superposed shuffled image requires the model to distinguish the overall edgemap object shape (relevant edges) from the shuffled image edges (irrelevant edges, less likely to be “shape like”). We perform extensive experi- ments over a range of model architectures, image classi- fication datasets, and probe datasets, comparing EL EAS against recent baselines. Figure 1 provides a small visual sample of our findings and results; across various models, a measure of shape sensitivity [17] correlates very strongly with measures of classifier robustness [11, 12] (see Resultsfor more details), validating the shape sensitivity inductive bias. In addition, for a number of model architectures, mod- els trained with EL EAS significantly improve both mea- sures compared to the previous SOTA data augmentation approach [21]. Summing up, we make the following contributions: • We propose an adversarial augmentation technique, ELEAS, designed to incentivize shape sensitivity in vision models. Our augmentation technique is lightweight (needing only an off-the-shelf edge detec- tion method) compared to previous proposals that re- quire expensive GAN-based image synthesis [7, 21]. • In experiments, EL EAS shows increased shape sen- sitivity on a wide range of tasks designed to probe this property. Consequently, we obtain increased ac- curacy in object classification with a 6% improvement on ImageNet-1K classification accuracy for ViT-Small among others. • EL EAS shows high generalizability and out of dis- tribution robustness with 14.2% improvement in ImageNet-C classification performance and 5.89% in- crease in shape-bias for Resnet152.

Tanay_Efficient_View_Synthesis_and_3D-Based_Multi-Frame_Denoising_With_Multiplane_Feature_CVPR_2023

Abstract While current multi-frame restoration methods combine information from multiple input images using 2D align- ment techniques, recent advances in novel view synthesis are paving the way for a new paradigm relying on volu- metric scene representations. In this work, we introduce the first 3D-based multi-frame denoising method that sig- nificantly outperforms its 2D-based counterparts with lower computational requirements. Our method extends the mul- tiplane image (MPI) framework for novel view synthesis by introducing a learnable encoder-renderer pair manip- ulating multiplane representations in feature space. The encoder fuses information across views and operates in a depth-wise manner while the renderer fuses information across depths and operates in a view-wise manner. The two modules are trained end-to-end and learn to separate depths in an unsupervised way, giving rise to Multiplane Feature (MPF) representations. Experiments on the Spaces and Real Forward-Facing datasets as well as on raw burst data validate our approach for view synthesis, multi-frame denoising, and view synthesis under noisy conditions.

1. Introduction Multi-frame denoising is a classical problem of com- puter vision where a noise process affecting a set of images must be inverted. The main challenge is to extract con- sistent information across images effectively and the cur- rent state of the art relies on optical flow-based 2D align- ment [3, 7, 45]. Novel view synthesis, on the other hand, is a classical problem of computer graphics where a scene is viewed from one or more camera positions and the task is to predict novel views from target camera positions. This problem requires to reason about the 3D structure of the scene and is typically solved using some form of volumetric representation [28, 32, 55]. Although the two problems are traditionally considered distinct, some novel view synthe- sis approaches have recently been observed to handle noisy Depth -wise EncoderView -wise RendererNoisy inputsSynthesized outputsMultiplane Feature Representation Noisy inputs IBRNet [48] NAN [31] MPFER (ours) Figure 1. Top: Our Multiplane Features Encoder-Renderer (MPFER) reimagines the MPI pipeline by moving the multiplane representation to feature space. Bottom: MPFER significantly out- performs existing methods in multiple challenging scenarios, in- cluding here, novel view synthesis from 8 highly degraded inputs. inputs well, and to have a denoising effect in synthesized views by discarding inconsistent information across input views [18,26]. This observation opens the door to 3D-based multi-frame denoising, by recasting the problem as a special case of novel view synthesis where the input views are noisy and the target views are the clean input views [26, 31]. Recently, novel view synthesis has been approached as an encoding-rendering process where a scene representa- tion is first encoded from a set of input images and an ar- bitrary number of novel views are then rendered from this scene representation. In the Neural Radiance Field (NeRF) framework for instance, the scene representation is a radi- ance field function encoded by training a neural network on the input views. Novel views are then rendered by querying and integrating this radiance field function over light rays originating from a target camera position [2, 24, 28]. In the Multiplane Image (MPI) framework on the other hand, the scene representation is a stack of semi-transparent colored layers arranged at various depths, encoded by feeding the input views to a neural network trained on a large number of scenes. Novel views are then rendered by warping and overcompositing the semi-transparent layers [8, 41, 55]. 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. 20898 In the present work, we adopt the MPI framework be- cause it is much lighter than the NeRF framework compu- tationally. The encoding stage only requires one inference pass on a network that generalizes to new scenes instead of training one neural network per-scene, and the rendering stage is essentially free instead of requiring a large number of inference passes. However, the standard MPI pipeline struggles to predict multiplane representations that are self- consistent across depths from multiple viewpoints. This problem can lead to depth-discretization artifacts in syn- thesized views [40] and has previously been addressed at the encoding stage using computationally expensive mech- anisms and a large number of depth planes [8, 11, 27, 40]. Here, we propose to enforce cross-depth consistency at the rendering stage by replacing the fixed overcompositing op- erator with a learnable renderer. This change of approach has three important implications. First, the encoder mod- ule can now process depths independently from each other and focus on fusing information across views. This sig- nificantly reduces the computational load of the encoding stage. Second, the scene representation changes from a static MPI to Multiplane Features (MPF) rendered dynam- ically. This significantly increases the expressive power of the scene encoding. Finally, the framework’s overall per- formance is greatly improved, making it suitable for novel scenarios including multi-frame denoising where it outper- forms standard 2D-based approaches at a fraction of their computational cost. Our main contributions are as follow: • We solve the cross-depth consistency problem for multi- plane representations at the rendering stage, by introduc- ing a learnable renderer. • We introduce the Multiplane Feature (MPF) representa- tion, a generalization of the multiplane image with higher representational power. • We re-purpose the multiplane image framework origi- nally developed for novel view synthesis to perform 3D- based multi-frame denoising. • We validate the approach with experiments on 3 tasks and 3 datasets and significantly outperform existing 2D-based and 3D-based methods for multi-frame denoising.

Tang_Master_Meta_Style_Transformer_for_Controllable_Zero-Shot_and_Few-Shot_Artistic_CVPR_2023

Abstract Transformer-based models achieve favorable perfor- mance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer atten- tion operations. Nevertheless, the over-paramerized multi- layer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses con- tent and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as Master specifically for style transfer. On the one hand, in the proposed model, differ- ent Transformer layers share a common group of parame- ters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is read- ily to control the degree of stylization via tuning the num- ber of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learn- able scaling operation on content features before content- style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learn- ing scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one *Equal contribution. †Corresponding author.specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive exper- iments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings.

1. Introduction Artistic style transfer aims at applying style patterns like colors and textures of a reference image to a given content image while preserving the semantic structure of the con- tent. In contrast to the pioneering optimization method [9] and early per-style-per-model methods like [17, 26], arbi- trary style transfer methods [3, 13, 16, 21, 22, 24, 31] enable real time style transfer for any style image in the test time in a zero-shot manner. The flexibility has led to this arbitrary- style-per-model fashion to dominate style transfer research. Recently, to enhance the representation of global infor- mation in arbitrary style transfer, Transformer [40] is intro- duced to this area [4], leveraging the global receptive field and powerful multi-head/layer structure, and achieves su- perior performance. Nevertheless, the over-parameterized multi-layer structure increases model parameters signifi- cantly. As shown in Fig. 1(a), there are 25.94M learn- able parameters for a 3-layer Transformer structure in StyTr2 [4], v.s. 3.50M in AdaIN [13], a simple but effective baseline in arbitrary style transfer. Such a large model for standard Transformer inevitably presents a heavy burden for training. As shown in Fig. 1(b), when there are more than 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. 18329 𝑐!𝑐"𝑐𝑠!𝑐𝑠"𝑐!#𝑐"#𝑐𝑠!𝑐𝑠"𝑠 Style / ContentVanilla Transformer w Residual ConnectionOur Transformer w Learnable Scaling Figure 2. Residual connection in the vanilla Transformer tends to destroy the original similarity relationship on content structure in style transfer task. Our model is designed to address this problem with learnable scaling parameters. The top row shows a simple 2D visualization and the bottom one provides a qualitative example. 4 layers, vanilla Transformer even fails to get convergent in training, which limits the scalability of the Transformer model in style transfer. Moreover, vanilla Transformer relies on residual connec- tions [12] to stylize content features, which suffers from the content-distortion problem. We illustrate this effect with a 2D visualization in Fig. 2(top), where residual connec- tions lead the transformation results of two content feature vectors to move towards the dominated style features and thereby tend to eliminate their original distinction. The vi- sual effect is that the stylized images would be dominated by some strong style features, such as salient edges, with the original self-(dis)similarity of content structures destroyed, as the example shown in Fig. 2(bottom). Focusing on these drawbacks, in this paper, we are ded- icated to devising a novel Transformer architecture specif- ically for artistic style transfer. On the one hand, in the proposed model, different Transformer layers share a com- mon group of parameters and a random number of stacked layers are adopted for each training iteration. Compared with the original version, sharing parameters across differ- ent layers reduces the total number of parameters signifi- cantly and leads to more convenient training convergence. As a byproduct, it is also readily for our model to control the degree of stylization via tuning the number of stacked layers freely in the inference time, as shown in Fig. 1(right). On the other hand, we equip Transformer with learnable scale parameters for content-style interactions instead of residual connections, which alleviates content distortion to a large degree and better preserves content structures while render- ing vivid style patterns simultaneously, as shown in the 2D visualization and the qualitative example in Fig. 2. Furthermore, beyond the typical zero-shot arbitrary style transfer, leveraging a meta learning scheme, our method is adaptable to the few-shot setting. By only fine-tuning the Transformer encoder layer in the few-shot stage, rapidadaptation to the model for a specific style within a lim- ited number of updates is possible, where the stylization with only 1 layer can be further improved, as shown in Fig. 1. Beyond that, we first achieve text-guided few- shot style transfer with this framework, which largely al- leviates the training burden of previous per-text-per-model solution. In this sense, we term the overall pipeline Meta StyleTransform er(Master ). Our contributions are summa- rized as follows: • We propose a novel Transformer architecture specif- ically for artistic style transfer. It shares parameters between different layers, which not only helps training convergence, but also allows convenient control over the stylization effect. • We identity the content distortion problem of resid- ual connections in Transformer and propose learnable scale parameters as an option to alleviate the problem. • We introduce a meta learning framework for adapting original training setting of zero-shot style transfer to the few-shot scenario, with which our Master achieves very good trade-off between flexibility and quality. • Experiments show that our model achieves results bet- ter than those of arbitrary-style-per-model methods. Furthermore, under the few-shot setting, either condi- tioned on image or text, Master can even yield perfor- mance on par with that of per-style-per-model methods with significantly less training cost.

Suhail_Omnimatte3D_Associating_Objects_and_Their_Effects_in_Unconstrained_Monocular_Video_CVPR_2023

Abstract We propose a method to decompose a video into a back- ground and a set of foreground layers, where the background captures stationary elements while the foreground layers capture moving objects along with their associated effects (e.g. shadows and reflections). Our approach is designed for unconstrained monocular videos, with an arbitrary camera and object motion. Prior work that tackles this problem assumes that the video can be mapped onto a fixed 2D can- vas, severely limiting the possible space of camera motion. Instead, our method applies recent progress in monocular camera pose and depth estimation to create a full, RGBD video layer for the background, along with a video layer for each foreground object. To solve the underconstrained decomposition problem, we propose a new loss formulation based on multi-view consistency. We test our method on challenging videos with complex camera motion and show significant qualitative improvement over current approaches.

1. Introduction Decomposing a video into meaningful layers (as show in Fig. 1) is a long-standing and complex problem [42] that hasseen a recent surge in progress with the application of deep neural networks [17, 24, 25, 46]. A challenging variant of layer decomposition is the omnimatte [25] task, which aims to separate an input video into a background and multiple foreground layers, each containing an object of interest along with its correlated effects such as shadows and reflections, thus enabling video editing applications such as object re- moval, background replacement and retiming. The original work on omnimatte [24, 25] introduced a self-supervised ap- proach for decomposing a video by assuming the background can be unwrapped onto a single static 2D background canvas using homography warping. Following work [17,46] relaxed the homography restriction, but maintained the necessity of unwrapping onto a 2D canvas. This 2D modelling, however, limits the applicability of these methods to camera motions that have limited or zero parallax; intuitively, only panning camera motions are al- lowed. If the camera’s center of projection moves a sig- nificant distance over the video, 2D methods are unable to learn an accurate background and are forced to place the background detail in a foreground layer (Figure 2). Since camera parallax is quite common, 2D methods are severely restricted in practice. To handle camera parallax, we exploit another recent line 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. 630 Background RGBForeground RGB Foreground Alpha OmnimatteOurNeural AtlasFigure 2. Existing methods fail on scenes with parallax. Top row: Omnimatte [25] fails due to inaccurate homography regis- tration, reconstructing the majority of the video in the foreground layer. Middle row: Neural-Atlas [17] incorrectly places trees in the foreground layer and struggles to inpaint the person region. Bottom row: Our method obtains a clean background and captures the person and their shadow in the foreground layer. of work on estimating 3D camera position and depth maps from casually-captured video with moving objects [18, 26, 48, 49]. Instead of a 2D layer decomposition, we propose to learn a 3D background model that varies per-frame and includes inpainted RGB and depth in the regions occluded by the foreground. Allowing the background to vary per- frame, rather than assuming a global, static canvas, creates an ill-posed decomposition problem, as variation in each input pixel could be explained by a foreground object, background, or both. To resolve this ambiguity, we enforce that the background varies slowly over time with 3D multi-view constraints using dense depth and camera pose estimates [48]. As shown in Fig. 2, our model can successfully separate the layers obtaining a clean background RGB whilst capturing only the person and their shadow in the foreground layer. Contributions. Our main contribution is a novel video decomposition method capable of accurately separating an input video with complex object and camera motion ( e.g., parallax) into a background and object layers. To address the ill-posed decomposition problem, we design regulariza- tion terms based on multi-view consistency and smoothness priors. The resulting model produces clean decompositions of real-world videos where previous methods fail.

Sun_Regularizing_Second-Order_Influences_for_Continual_Learning_CVPR_2023

Abstract Continual learning aims to learn on non-stationary data streams without catastrophically forgetting previous knowl- edge. Prevalent replay-based methods address this chal- lenge by rehearsing on a small buffer holding the seen data, for which a delicate sample selection strategy is required. However, existing selection schemes typically seek only to maximize the utility of the ongoing selection, overlooking the interference between successive rounds of selection. Motivated by this, we dissect the interaction of sequential selection steps within a framework built on influence func- tions. We manage to identify a new class of second-order influences that will gradually amplify incidental bias in the replay buffer and compromise the selection process. To reg- ularize the second-order effects, a novel selection objective is proposed, which also has clear connections to two widely adopted criteria. Furthermore, we present an efficient im- plementation for optimizing the proposed criterion. Exper- iments on multiple continual learning benchmarks demon- strate the advantage of our approach over state-of-the-art methods. Code is available at https://github.com/ feifeiobama/InfluenceCL .

1. Introduction The ability to continually accumulate knowledge is a hallmark of intelligence, yet prevailing machine learning systems struggle to remember old concepts after acquir- ing new ones. Continual learning has hence emerged to tackle this issue, also known as catastrophic forgetting , and thereby enable the learning on long task sequences which are frequently encountered in real-world applica- tions [15, 17]. Amongst a variety of valid methods, the replay-based approaches [33, 39] achieve evidently strong results by buffering a small coreset of previously seen data for rehearsal in later stages. Due to the rigorous constraints on memory overhead, the replay buffer needs to be carefully maintained such that only the samples deemed most critical *Corresponding author. CoresetNew dataSelectionSelectionInfluenceCoresetNew dataCoresetNew data(a) Greedy selectionDegraded selectionRegularize Oracle selectionNow(b)Figure 1. (a) In continual learning, earlier coreset selection exerts a profound influence on subsequent steps through the data flow. (b) By ignoring this, a greedy selection strategy can degrade over time. Therefore, we propose to model and regularize the influence of each selection on the future. for preserving overall performance are selected during the learning procedure. Despite much effort in sophisticated design of coreset selection criteria [3,7,47,52], existing works are mostly de- veloped for an oversimplified scenario, where each round of selection is considered isolated and primarily focused on refining single-round performance. For example, Borsos et al. [7] greedily minimize the empirical risk on the currently available subset of data. However, as shown in Fig. 1a, the actual selection process within continual learning is rather different in that each prior selection defines the input for subsequent selection and therefore has an inevitable impact on the later decision. Neglecting such interactions between successive selection steps, as in previous practice, may re- sult in a degraded coreset after the prolonged selection pro- cess. To maximize overall performance, an ideal continual learning method should take into account the future influ- ence of each round of selection, which remains unresolved due to the obscure role of intermediate selection 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. 20166 This work models the interplay among multiple selection steps with the classic tool of influence functions [21, 29], which estimates the effect of each training point by perturb- ing sample weights. Similarly, we begin by upweighting two samples from consecutive time steps, and uncover that the latter one’s influence evaluation is biased due to the con- sequent perturbations in test gradient and coreset Hessian. This gives rise to a new class of second-order influences that interfere with the initial influence-based selection strat- egy. To be specific, the selection tends to align across dif- ferent rounds, favoring those samples with a larger gradient projection along the previously upweighted ones. It further suggests that some unintended adverse effects of the early selection steps, such as incidental bias introduced into the replay buffer, will be amplified after rounds of selection and impair the final performance. To address the newly discovered disruptive effects, we propose to regularize each round of selection beforehand based on its expected second-order influence, as illustrated in Fig. 1b. Intuitively, the selection with a large magnitude of second-order influence will substantially interfere with future influence estimates, and therefore should be avoided. However, the magnitude itself cannot be precalculated for direct guidance of the selection, since it is related to some unknown future data. We instead derive a tractable upper bound for the magnitude which results in the form of ℓ2- norm, and integrated it with the vanilla influence functions to serve as the final selection criterion. The proposed regu- larizer can be interpreted with clarity as it equates to a cou- ple of existing criteria ranging from gradient matching to diversity, under varying simplifications. Finally, we present an efficient implementation that tackles the technical chal- lenges in selecting with neural networks. Our contributions are summarized as below: • We investigate the previously-neglected interplay be- tween consecutive selection steps within an influence- based selection framework. A new type of second- order influences is identified, and further analysis states its harmfulness in continual learning. • A novel regularizer is proposed to mitigate the second- order interference. The regularizer is associated with two other popular selection criteria and can be effi- ciently optimized with our implementation. • Comprehensive experiments on three continual learn- ing benchmarks (Split CIFAR-10, Split CIFAR-100 and Split miniImageNet) clearly illustrates that our method achieves new state-of-the-art performance.

Tang_NeuMap_Neural_Coordinate_Mapping_by_Auto-Transdecoder_for_Camera_Localization_CVPR_2023

Abstract This paper presents an end-to-end neural mapping method for camera localization, encoding a whole scene into a grid of latent codes, with which a Transformer- based auto-decoder regresses 3D coordinates of query pix- els. State-of-the-art camera localization methods require each scene to be stored as a 3D point cloud with per-point features, which takes several gigabytes of storage per scene. While compression is possible, the performance drops sig- nificantly at high compression rates. NeuMap achieves ex- tremely high compression rates with minimal performance drop by using 1) learnable latent codes to store scene in- formation and 2) a scene-agnostic Transformer-based auto- decoder to infer coordinates for a query pixel. The scene- agnostic network design also learns robust matching priors by training with large-scale data, and further allows us to just optimize the codes quickly for a new scene while fixing the network weights. Extensive evaluations with five bench- marks show that NeuMap outperforms all the other coor-dinate regression methods significantly and reaches similar performance as the feature matching methods while hav- ing a much smaller scene representation size. For example, NeuMap achieves 39.1% accuracy in Aachen night bench- mark with only 6MB of data, while other compelling meth- ods require 100MB or a few gigabytes and fail completely under high compression settings. The codes are available at https://github.com/Tangshitao/NeuMap.

1. Introduction Visual localization determines camera position and ori- entation based on image observations, an essential task for applications such as VR/AR and self-driving cars. Despite significant progress, accurate visual localization remains a challenge, especially when dealing with large viewpoint and illumination changes. Compact map representation is an- other growing concern, as applications like delivery robots may require extensive maps. Standard visual localization techniques rely on massive databases of keypoints with 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. 929 coordinates and visual features, posing a significant bottle- neck in real-world applications. Visual localization techniques generally establish 2D- 3D correspondences and estimate camera poses using perspective-n-point (PnP) [22] with a sampling method like RANSAC [15]. These methods can be divided into two cat- egories: Feature Matching (FM) [11, 34, 37, 49] and Scene Coordinate Regression (SCR) [3, 5, 7, 40]. FM methods, which are trained on a vast amount of data covering vari- ous viewpoint and illumination differences, use sparse ro- bust features extracted from the query image and matched with those in candidate scene images. This approach ex- ploits learning-based feature extraction and correspondence matching methods [13, 35, 42, 44] to achieve robust local- ization. However, FM methods require large maps, mak- ing them impractical for large-scale scenes. Many meth- ods [11, 27, 49] have been proposed to compress this map representation, but often at the cost of degraded perfor- mance. On the other hand, SCR methods directly regress a dense scene coordinate map using a compact random for- est or neural network, providing accurate results for small- scale indoor scenes. However, their compact models lack generalization capability and are often restricted to lim- ited viewpoint and illumination changes. Approaches such as ESAC [5] handle larger scenes by dividing them into smaller sub-regions, but still struggle with large viewpoint and illumination changes. We design our method to enjoy the benefits of compact scene representation of SCR methods and the robust per- formance of FM methods. Similar to FM methods, we also focus on a sparse set of robustly learned features to deal with large viewpoint and illumination changes. On the other hand, we exploit similar ideas to SCR methods to regress the 3D scene coordinates of these sparse fea- tures in the query images with a compact map represen- tation. Our method, dubbed neural coordinate mapping (NeuMap), first extracts robust features from images and then applies a transformer-based auto-decoder (i.e., auto- transdecoder) to learn: 1) a grid of scene codes encoding the scene information (including 3D scene coordinates and feature information) and 2) the mapping from query im- age feature points to 3D scene coordinates. At test time, given a query image, after extracting image features of its key-points, the auto-transdecoder regresses their 3D coor- dinates via cross attention between image features and la- tent codes. In our method, the robust feature extractor and the auto-transdecoder are scene-agnostic, where only latent codes are scene specific. This design enables the scene- agnostic parameters to learn matching priors across scenes while maintaining a small data size. To handle large scenes, we divide the scene into smaller sub-regions and process them independently while applying a network pruning tech- nique [25] to drop redundant codes.We demonstrate the effectiveness of NeuMap with a di- verse set of five benchmarks, ranging from indoor to out- door and small to large-scale: 7scenes (indoor small), Scan- Net (indoor small), Cambridge Landmarks (outdoor small), Aachen Day &Night (outdoor large), and NA VER LABS (indoor large). In small-scale datasets (i.e., 7scenes and Cambridge Landmarks), NeuMap compresses the scene representation by around 100-1000 times without any per- formance drop compared to DSAC++. In large-scale datasets (i.e., Aachen Day &Night and NA VER LABS), NeuMap significantly outperforms the current state-of-the- art at high compression settings, namely, HLoc [34] with a scene-compression technique [49]. In ScanNet dataset, we demonstrate the quick fine-tuning experiments, where we only optimize the codes for a new scene while fixing the scene-agnostic network weights.

Tang_Parts2Words_Learning_Joint_Embedding_of_Point_Clouds_and_Texts_by_CVPR_2023

Abstract Shape-Text matching is an important task of high-level shape understanding. Current methods mainly represent a 3D shape as multiple 2D rendered views, which obviously can not be understood well due to the structural ambigu- ity caused by self-occlusion in the limited number of views. To resolve this issue, we directly represent 3D shapes as point clouds, and propose to learn joint embedding of point clouds and texts by bidirectional matching between parts from shapes and words from texts. Specifically, we first seg- ment the point clouds into parts, and then leverage optimal transport method to match parts and words in an optimized feature space, where each part is represented by aggregat- ing features of all points within it and each word is ab- stracted by its contextual information. We optimize the fea- ture space in order to enlarge the similarities between the paired training samples, while simultaneously maximizing the margin between the unpaired ones. Experiments demon- strate that our method achieves a significant improvement in accuracy over the SOTAs on multi-modal retrieval tasks under the Text2Shape dataset. Codes are available at here.

1. Introduction Interaction scenarios, such as metaverse, and computer- aided design (CAD), create a larger number of 3D shapes and text descriptions. To enable a more intelligent process of interaction, it is important to bridge the gap between 3D data and linguistic data. Recently, 3D shapes with rich geo- metric details have been available in large-scale 3D deep learning benchmark datasets [5, 34]. Beyond 3D shapes themselves, text descriptions can also provide additional in- formation. However, it is still hard to jointly understand 3D shapes and texts, since representing different modalities in a common semantic space is still very challenging. The existing methods aim at learning a joint embedded *Corresponding authors.space to connect various 3D representations with texts, such as voxel grids [6] and multi-view rendered images [11, 12]. However, due to the low resolution and self-occlusions, it is hard for those methods mentioned above to improve the ability of joint understanding of shapes and texts. On the other hand, previous shape-text matching methods [6, 12, 37] usually take the global features of the entire 3D shape for text matching, making it challenging to capture the local geometries, and thus are not suitable for matching detailed geometric descriptions. Regional-based matching approaches are commonly employed in the image-text matching task [21–23, 27], whereby visual-text alignment is established at the seman- tic level to enhance the performance of retrieval. These models compute the local similarities between regions and words and then aggregate the local information to obtain the global metrics between the heterogeneous pairs. How- ever, these two-stage methods based on the pre-trained seg- mentation networks split the connection between matching embeddings and segmentation prior information. In this paper, we introduce an optimal transport based shape-text matching method to achieve fine-grained align- ment and retrieval of 3D shapes and texts, as shown in Fig- ure 1. To mitigate the influence of low-resolution or self- occlusions, we directly represent the shape as point clouds and learn a part-level segmentation prior. Afterward, we leverage optimal transport to build the regional cross-modal correspondences and achieve more precise retrieval results. Our main contributions are summarized as follows: • We propose a novel end-to-end network framework to learn the joint embedding of point clouds and texts, which enables the bidirectional matching be- tween parts from point clouds and words from texts. • We leverage optimal transport theory to obtain the best matches between parts and words and incorporate Earth Mover’s Distance (EMD) to describe the match- ing score. • To the best of our knowledge, our proposed network 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. 6884 Figure 1. Comparison between the global-based matching method and our proposed method. The proposed end-to-end framework aims to learn the joint embedding of point clouds and text by matching parts to words. It can either retrieve shapes using text or vice versa. Our novelty lies in the way of jointly learning embeddings of point clouds and texts. achieves SOTA results in joint 3D shape/text under- standing tasks in terms of various evaluation metrics.

Uy_SCADE_NeRFs_from_Space_Carving_With_Ambiguity-Aware_Depth_Estimates_CVPR_2023

Abstract Neural radiance fields (NeRFs) have enabled high fidelity 3D reconstruction from multiple 2D input views. However, a well-known drawback of NeRFs is the less-than-ideal per- formance under a small number of views, due to insufficient constraints enforced by volumetric rendering. To address this issue, we introduce SCADE, a novel technique that improves NeRF reconstruction quality on sparse, unconstrained in- put views for in-the-wild indoor scenes. To constrain NeRF reconstruction, we leverage geometric priors in the form of per-view depth estimates produced with state-of-the-art monocular depth estimation models, which can generalize across scenes. A key challenge is that monocular depth es- timation is an ill-posed problem, with inherent ambiguities. To handle this issue, we propose a new method that learns to predict, for each view, a continuous, multimodal distribu- tion of depth estimates using conditional Implicit Maximum Likelihood Estimation (cIMLE). In order to disambiguate exploiting multiple views, we introduce an original space carving loss that guides the NeRF representation to fuse multiple hypothesized depth maps from each view and distill from them a common geometry that is consistent with all views. Experiments show that our approach enables higher fidelity novel view synthesis from sparse views. Our project page can be found at scade-spacecarving-nerfs.github.io.

1. Introduction Neural radiance fields (NeRF) [24] have enabled high fidelity novel view synthesis from dozens of input views. Such number of views are difficult to obtain in practice, however. When given only a small number of sparse views, vanilla NeRF tends to struggle with reconstructing shape accurately, due to inadequate constraints enforced by the volume rendering loss alone. Shape priors can help remedy this problem. Various forms of shape priors have been proposed for NeRFs, such as object category-level priors [13], and handcrafted data-independent priors [25]. The former requires knowledge of category labels and is not applicable to scenes, and the latter is agnos- tic to the specifics of the scene and only encodes low-level regularity (e.g., local smoothness). A form of prior that is both scene-dependent and category-agnostic is per-view monocular depth estimates, which have been explored in prior work [6,32]. Unfortunately, monocular depth estimates are often inaccurate, due to estimation errors and inherent ambiguities, such as albedo vs. shading (cf. check shadow illusion), concavity vs. convexity (cf. hollow face illusion), distance vs. scale (cf. miniature cinematography), etc. As a result, the incorrect or inconsistent priors imposed by such depth estimates may mislead the NeRF into reconstructing incorrect shape and produce artifacts in the generated views. In this paper, we propose a method that embraces the uncertainty and ambiguities present in monocular depth esti- 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. 16518 mates, by modelling a probability distribution over depth es- timates. The ambiguities are retained at the stage of monocu- lar depth estimation, and are only resolved once information from multiple views are fused together. We do so with a prin- cipled loss defined on probability distributions over depth estimates for different views. This loss selects the subset of modes of probability distributions that are consistent across all views and matches them with the modes of the depth distribution along different rays as modelled by the NeRF. It turns out that this operation can be interpreted as a proba- bilistic analogue of classical depth-based space carving [17]. For this reason, we dub our method Space Carving with Ambiguity-aware Depth Estimates , or SCADE for short. Compared to prior approaches of depth supervision [6,32] that only supervise the moments of NeRF’s depth distribu- tion (rather than the modes), our key innovation is that we supervise the modes of NeRF’s depth distribution. The su- pervisory signal provided by the former is weaker than the latter, because the former only constrains the value of an in- tegral aggregated along the ray, whereas the latter constrains the values at different individual points along the ray. Hence, the supervisory signal provided by the former is 2D (because it integrates over a ray), whereas the supervisory signal pro- vided by the our method is 3D (because it is point-wise) and thus can be more fine-grained. An important technical challenge is modelling probability distributions over depth estimates. Classical approaches use simple distributions with closed-form probability densities such as Gaussian or Laplace distributions. Unfortunately these distributions are not very expressive, since they only have a single mode (known as “unimodal”) and have a fixed shape for the tails. Since each interpretation of an ambiguous image should be a distinct mode, these simple unimodal dis- tributions cannot capture the complex ambiguities in depth estimates. Naïve extensions like a mixture of Gaussians are also not ideal because some images are more ambiguous than others, and so the number of modes needed may differ for the depth estimates of different images. Moreover, learning such a mixture requires backpropagating through E-M, which is nontrivial. Any attempt at modifying the probability density to make it capable of handling a variable number of modes can easily run into an intractable partition function [4,10,31], which makes learning difficult because maximum likelihood requires the ability to evaluate the probability density, which is a function of the partition function. To get around this conundrum, we propose representing the probability distribution with a set of samples generated from a neural network. Such a distribution can be learned with a conditional GAN; however, because GANs suffer from mode collapse, they cannot model multiple modes [11] and are therefore unsuited to modelling ambiguity. Instead, we propose leveraging conditional Implicit Maximum Like- lihood Estimation (cIMLE) [19, 28] to learn the distribution, which is designed to avoid mode collapse. We consider the challenging setting of leveraging out-of- Figure 2. Ambiguities from a single image . We show samples from our ambiguity-aware depth estimates that is able to handle different types of ambiguities. (Top) An image of a chair taken from the top-view. Without context of the scene, it is unclear that it is an image of a chair. We show different samples from our ambiguity-aware depth estimates that are able to capture different degrees of convextiy. (Bottom) An image of a cardboard under bad lighting conditions that capture the albedo vs shading ambiguity that is also represented by our different samples. domain depth priors to train NeRFs on real-world indoor scenes with sparse views. Under this setting, the depth priors we use are trained on a different dataset (e.g., Taskonomy) from the scenes our NeRFs are trained on (e.g., ScanNet). This setting is more challenging than usual due to domain gap between the dataset the prior is trained on and the scenes NeRF is asked to reconstruct. We demonstrate that our method outperforms vanilla NeRF and NeRFs with supervi- sion from depth-based priors in novel view synthesis. In summary, our key contributions include: •An method that allows the creation of NeRFs in uncon- strained indoor settings from only a modest number of 2D views by introducing a sophisticated way to exploit ambiguity-aware monocular depth estimation. •A novel way to sample distributions over image depth estimates based on conditional Implicit Maximum Like- lihood Estimation that can represent depth ambiguities and capture a variable number of discrete depth modes. •A new space-carving loss that can be used in the NeRF formulation to optimize for a mode-seeking 3D den- sity that helps select consistent depth modes across the views and thus compensate for the under-constrained photometric information in the few view regime.

Tan_Temporal_Attention_Unit_Towards_Efficient_Spatiotemporal_Predictive_Learning_CVPR_2023

Abstract Spatiotemporal predictive learning aims to generate fu- ture frames by learning from historical frames. In this pa- per, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame fea- tures and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recur- rent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unpar- allelizable architectures. To parallelize the temporal mod- ule, we propose the Temporal Attention Unit (TAU), which decomposes temporal attention into intra-frame statical at- tention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regu- larization to take inter-frame variations into account. Ex- tensive experiments demonstrate that the proposed method enables the derived model to achieve competitive perfor- mance on various spatiotemporal prediction benchmarks.

1. Introduction The last decade has witnessed revolutionary advances in deep learning across various supervised learning tasks such as image classification [34,52,87], object detection [73,75], computational biology [21, 22, 43, 83, 84], and etc. Despite significant breakthroughs in supervised learning, which re- lies on large-scale labeled datasets, the potential of unsuper- vised learning remains largely untapped. Self-supervised learning that designs pretext tasks to produce labels derived from the data itself is recognized as a subset of unsupervised learning. In the context of self-supervised learning, con- trastive self-supervised learning [7,8,28,33,85,86,92,104] predicts the noise contrastive estimation from predefined positive or negative pairs, and masked self-supervised learn- ing[17, 32, 45, 51, 57, 102] predicts the masked patches from the visible patches. Unlike these image-level self- *Equal contribution. †Corresponding author.supervised learning, spatiotemporal predictive learning that predicts future frames from past frames at the video- level [6, 19, 27, 46, 60, 63]. Spatial EncoderSpatial DecoderTemporalModuleConvLSTMConvLSTMConvLSTM3D Conv3D ConvE3D-LSTM ConvConvConvLSTMConvConvTAUPhyCellInvertibleFlowReverseFlowST-LSTM(a) General framework(b) ConvLSTM(c) E3D-LSTM (d) PhyDNet(e) CrevNet(f) Ours Figure 1. Comparison of model architectures among common spa- tiotemporal predictive learning methods. Note that we denote 2D convolutional neural networks as Conv while 3D Conv means 3D convolutional neural networks. Accurate spatiotemporal predictive learning can bene- fit broad practical applications in climate change [74, 77], human motion forecasting [91, 107], traffic flow predic- tion [18, 97], and representation learning [39, 71]. The significance of spatiotemporal predictive learning primarily lies in its potential of exploring both spatial correlation and temporal evolution in the physical world. Moreover, the self-supervised nature of spatiotemporal predictive learn- ing aligns well with human learning styles without a large amount of labeled data. Massive videos can provide a rich source of visual information, enabling spatiotemporal pre- 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. 18770 dictive learning to serve as a generative pre-training strat- egy [35, 60] for feature representation learning towards di- verse downstream visual supervised tasks. Most of the existing methods [2,3,9,11,12,19,24,29,41, 44,65,78,79,82,88,89,93–96,98,100,103] in spatiotempo- ral predictive learning employ hybrid architectures of con- volutional neural networks and recurrent units in which spa- tial correlation and time evolution can be learned, respec- tively. Inspired by the success of long short-term mem- ory (LSTM) [36] in sequential modeling, ConvLSTM [78] is a seminal work on the topic of spatiotemporal predic- tive learning that extends fully connected LSTM to con- volutional LSTM towards accurate precipitation nowcast- ing. PredRNN [95] is an admirable work that proposes Spatiotemporal LSTM (ST-LSTM) units to model spatial appearances and temporal variations in a unified memory pool. This work provides insights on designing typical re- current units for spatiotemporal predictive learning and in- spires a series of subsequent works [4, 93, 96, 98]. E3D- LSTM [94] integrates 3D convolutional neural networks into recurrent units towards good representations with both short-term frame dependencies and long-term high-level re- lations. PhyDNet [29] introduces a two-branch architec- ture that involves physical-based PhyCells and ConvLSTMs for performing partial differential equation constraints in latent space. CrevNet [103] proposes an invertible two- way autoencoder based on flow [14, 15] and a condition- ally reversible architecture for spatiotemporal predictive learning. As shown in Figure 1 (a), we present a gen- eral framework consisting of the spatial encoder/decoder and the middle temporal module, abstracted from these methods. Though these spatiotemporal predictive learning methods have different temporal modules and spatial en- coders/decoders, they basically share a similar framework. Based on the general framework, we argue that the tem- poral module plays an essential role in spatiotemporal pre- dictive learning. While the common choice of the tempo- ral module is the recurrent-based units, we explore a novel parallelizable attention module named Temporal Attention Unit (TAU) to capture time evolution. The proposed tempo- ral attention is decomposed into intra-frame statical atten- tion and inter-frame dynamical attention. Furthermore, we argue that the mean square error loss only focuses on intra- frame differences, and we propose a differential divergence regularization that also cares about the inter-frame varia- tions. Keeping the spatial encoder and decoder as simple as 2D convolutional neural networks, we deliberately imple- ment our proposed TAU modules and surprisingly find the derived model achieves competitive performance as those recurrent-based models. This observation provides a new perspective to improve spatiotemporal predictive learning by parallelizable attention networks instead of common- used recurrent units.We conduct experiments on various datasets with dif- ferent experimental settings: (1) Standard spatiotemporal predictive learning. Quantitative results on various datasets demonstrate our proposed method can achieve competitive performance on standard spatiotemporal predictive learn- ing. (2) Generalization ability. To verify the generaliza- tion ability, we train our model on KITTI and test it on the Caltech Pedestrian dataset with different domains. (3) Pre- dict future frames with flexible lengths. We tackle the long- length frames by feeding the predicted frames as the input and find the performance is consistently well. Through the superior performance in the above three experimental set- tings, we demonstrate that our proposed model can provide a novel manner that learns temporal dependencies without recurrent units.

Tejero_Full_or_Weak_Annotations_An_Adaptive_Strategy_for_Budget-Constrained_Annotation_CVPR_2023

Abstract Annotating new datasets for machine learning tasks is tedious, time-consuming, and costly. For segmentation ap- plications, the burden is particularly high as manual delin- eations of relevant image content are often extremely expen- sive or can only be done by experts with domain-specific knowledge. Thanks to developments in transfer learning and training with weak supervision, segmentation models can now also greatly benefit from annotations of different kinds. However, for any new domain application looking to use weak supervision, the dataset builder still needs to define a strategy to distribute full segmentation and other weak annotations. Doing so is challenging, however, as it is a priori unknown how to distribute an annotation budget for a given new dataset. To this end, we propose a novel ap- proach to determine annotation strategies for segmentation datasets, whereby estimating what proportion of segmen- tation and classification annotations should be collected given a fixed budget. To do so, our method sequentially determines proportions of segmentation and classification annotations to collect for budget-fractions by modeling the expected improvement of the final segmentation model. We show in our experiments that our approach yields annota- tions that perform very close to the optimal for a number of different annotation budgets and datasets.

1. Introduction Semantic segmentation is a fundamental computer vi- sion task with applications in numerous domains such as autonomous driving [11, 43], scene understanding [45], surveillance [50] and medical diagnosis [9, 18]. As the ad- vent of deep learning has significantly advanced the state- of-the-art, many new application areas have come to lightand continue to do so too. This growth has brought and continues to bring exciting domain-specific datasets for seg- mentation tasks [6, 19, 29, 32, 52]. Today, the process of establishing machine learning- based segmentation models for any new application is rel- atively well understood and standard. Only once an image dataset is gathered and curated, can machine learning mod- els be trained and validated. In contrast, building appropri- ate datasets is known to be difficult, time-consuming, and yet paramount. Beyond the fact that collecting images can be tedious, a far more challenging task is producing ground- truth segmentation annotations to subsequently train (semi) supervised machine learning models. This is mainly be- cause producing segmentation annotations often remains a manual task. As reported in [4], generating segmentation annotations for a single PASCAL image [15] takes over 200 seconds on average. This implies over 250 hours of annotation time for a dataset containing a modest 5’000 images. What often further exacerbates the problem for domain-specific datasets is that only the dataset designer, or a small group of individuals, have enough expertise to produce the annotations ( e.g., doctors, experts, etc.), mak- ing crowd-sourcing ill-suited. To overcome this challenge, different paradigms have been suggested over the years. Approaches such as Active Learning [7, 8, 26] aim to iteratively identify subsets of im- ages to annotate so as to yield highly performing models. Transfer learning has also proved to be an important tool in reducing annotation tasks [13, 17, 24, 25, 30, 36]. For in- stance, [37] show that training segmentation models from scratch is often inferior to using pre-training models de- rived from large image classification datasets, even when the target application domain differs from the source do- main. Finally, weakly-supervised methods [2, 40] combine pixel-wise annotations with other weak annotations that are faster to acquire, thereby reducing the annotation burden. 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. 11381 OCTPASCAL VOCSUIMCityscapesFigure 1. Illustration of different semantic segmentation applications; OCT: Pathologies of the eye in OCT images, SUIM: Underwater scene segmentation [19], Cityscape: street level scene segmentation [11], PASCAL VOC: natural object segmentation. In particular, Papandreou et al. [40] showed that combina- tions of strong and weak annotations ( e.g., bounding boxes, keypoints, or image-level tags) delivered competitive results with a reduced annotation effort. In this work, we rely on these observations and focus on the weakly supervised seg- mentation setting. In the frame of designing annotation campaigns, weakly- supervised approaches present opportunities for efficiency as well. Instead of completely spending a budget on a few expensive annotations, weakly-supervised methods allow a proportion of the budget to be allocated to inexpensive, or weak, labels. That is, one could spend the entire annotation budget to manually segment available images, but would ultimately lead to relatively few annotations. Conversely, weak annotations such as image-level labels are roughly 100 times cheaper to gather than their segmentation coun- terparts [4]. Thus, a greater number of weakly-annotated images could be used to train segmentation models at an equal cost. In fact, under a fixed budget, allocating a pro- portion of the budget to inexpensive image-level class labels has been shown to yield superior performance compared to entirely allocating a budget to segmentation labels [4]. Yet, allocating how an annotation budget should be dis- tributed among strong and weak annotations is challenging, and inappropriate allocations may severely impact the qual- ity of the final segmentation model. For example, spend- ing the entire budget on image-level annotations will clearly hurt the performance of a subsequent segmentation model. Instead, a naive solution would be to segment and classify a fixed proportion of each ( e.g., say 80% - 20%). Knowing what proportion to use for a given dataset is unclear, how- ever. Beyond this, there is no reason why the same fixed proportion would be appropriate across different datasets or application domains. That is, it would be highly unlikelythat the datasets shown in Fig. 1 all require the same pro- portion of strong and weak annotations to yield optimal seg- mentation models. Despite its importance, choosing the best proportion of annotation types remains a largely unexplored research question. Weakly-supervised and transfer-learning meth- ods generally assume that the annotation campaign and the model training are independent and that all annotations are simply available at training time. While active learning methods do alternate between annotation and training, they focus on choosing optimal samples to annotate rather than choosing the right type of annotations. Moreover, most ac- tive learning methods ignore constraints imposed by an an- notation budget. More notable, however, is the recent work of Mahmood et. al. [33, 34] which aims to determine what weak and strong annotation strategy is necessary to achieve a target performance level. While noteworthy, this objective differs from that here, whereby given a fixed budget, what strategy is best suited for a given new dataset? To this end, we propose a novel method to find an opti- mal budget allocation strategy in an online manner. Using a collection of unlabeled images and a maximum budget, our approach selects strong and weak annotations, constrained by a given budget, that maximize the performance of the subsequent trained segmentation model. To do this, our method iteratively alternates between partial budget alloca- tions, label acquisition, and model training. At each step, we use the annotations performed so far to train multiple models to estimate how different proportions of weak and strong annotations affect model performance. A Gaussian Process models these results and maps the number of weak and strong annotations to the expected model improvement. Computing the Pareto optima between expected improve- ment and costs, we choose a new sub-budget installment 11382 and its associated allocation so to yield the maximum ex- pected improvement. We show in our experiments that our approach is beneficial for a broad range of datasets, and il- lustrate that our dynamic strategy allows for high perfor- mances, close to optimal fixed strategies that cannot be de- termined beforehand.

Tong_Seeing_Through_the_Glass_Neural_3D_Reconstruction_of_Object_Inside_CVPR_2023

Abstract In this paper, we define a new problem of recovering the 3D geometry of an object confined in a transparent en- closure. We also propose a novel method for solving this challenging problem. Transparent enclosures pose chal- lenges of multiple light reflections and refractions at the interface between different propagation media e.g. air or glass. These multiple reflections and refractions cause seri- ous image distortions which invalidate the single viewpoint assumption. Hence the 3D geometry of such objects can- not be reliably reconstructed using existing methods, such as traditional structure from motion or modern neural re- construction methods. We solve this problem by explicitly modeling the scene as two distinct sub-spaces, inside and outside the transparent enclosure. We use an existing neu- ral reconstruction method (NeuS) that implicitly represents the geometry and appearance of the inner subspace. In or- der to account for complex light interactions, we develop a hybrid rendering strategy that combines volume rendering with ray tracing. We then recover the underlying geome- try and appearance of the model by minimizing the differ- ence between the real and rendered images. We evaluate our method on both synthetic and real data. Experiment results show that our method outperforms the state-of-the- art (SOTA) methods. Codes and data will be available at https://github.com/hirotong/ReNeuS

1. Introduction Digitization of museum collections is an emerging topic of increasing interest, especially with the advancement of virtual reality and Metaverse technology. Many famous mu- seums around the world have been building up their own digital collections1 2for online exhibitions. Among these collections, there is a kind of special but important collec- 1https://collections.louvre.fr/en/ 2https://www.metmuseum.org/ (a) Reference Image (b) COLMAP [26] (c) NeuS [31] (d)Ours Figure 1. We propose a new research problem of recovering shape of objects inside a transparent container as shown in (a) and solve it by a physically-based neural reconstruction solution. Other meth- ods (b) and (c) without consideration of the optical effects fail to reconstruct, (d) our proposed ReNeuS can retrieve high-quality mesh of the insect object. tions such as insects [7, 23], human tissues [30], aquatic creatures [6] and other fragile specimens that need to be preserved inside some hard but transparent materials ( e.g. resin, glass) as shown in Fig. 1a. In order to digitize such collections, we abstract a distinct research problem which is seeing through the glassy outside and recovering the geom- etry of the object inside. To be specific, the task is to recon- struct the 3D geometry of an object of interest from multiple 2D views of the scene. The object of interest is usually con- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12555 tained in a transparent enclosure, typically a cuboid or some regular-shaped container. A major challenge of this task is the severe image dis- tortion caused by repeated light reflection and refraction at the interface of air and transparent blocks. Since it inval- idates the single viewpoint assumption [9], most existing multi-view 3D reconstruction methods tend to fail. One of the most relevant research problems is 3D re- construction through refractive interfaces, known as Refrac- tive Structure-from-Motion (RSfM) [1, 4]. In RSfM, one or more parallel interfaces between different propagation media, with different refractive indices, separate the opti- cal system from the scene to be captured. Light passing through the optical system refracts only once at each inter- face causing image distortion. Unlike RSfM methods, we deal with intersecting interfaces where multiple reflections and refractions take place. Another related research topic is reconstruction of transparent objects [13, 17, 19]. In this problem, lights only get refracted twice while entering and exiting the target. In addition to the multiple refractions considered by these methods, our method also tackles the problem of multiple reflections within the transparent en- closure. Recently, neural implicit representation methods [21,22, 28, 29, 31] achieve state-of-the-art performance in the task of novel view synthesis and 3D reconstruction showing promising ability to encode the appearance and geometry. However, these methods do not consider reflections. Al- though NeRFReN [12] attempts to incorporate the reflec- tion effect, it cannot handle more complicated scenarios with multiple refraction and reflection. In order to handle such complicated cases and make the problem tractable, we make the following two reasonable simplifications: i.Known geometry and pose of the transparent block. It is not the focus of this paper to estimate the pose of the interface plane since it is either known [1,4] or cal- culated as a separate research problem [14] as is typi- cally assumed in RSfM. ii.Homogeneous background and ambient lighting. Since the appearance of transparent objects is highly affected by the background pattern, several transparent object reconstruction methods [13, 19] obtain ray cor- respondence with a coded background pattern. To con- centrate on recovering the shape of the target object, we further assume monochromatic background with only homogeneous ambient lighting conditions similar to a photography studio. To handle both Reflection and Refraction in 3D recon- struction, we propose ReNeuS , a novel extension of a neu- ral implicit representation method NeuS [31]. In lieu ofdealing with the whole scene together, a divide-and-conquer strategy is employed by explicitly segmenting the scene into internal space, which is the transparent container containing the object, and an external space which is the surrounding air space. For the internal space, we use NeuS [31] to en- code only the geometric and appearance information. For the external space, we assume a homogeneous background with fixed ambient lighting as described in simplification ii. In particular, we use a novel hybrid rendering strategy that combines ray tracing with volume rendering techniques to process the light interactions across the two sub-spaces. The main contributions of this paper are as follows: • We introduce a new research problem of 3D recon- struction of an object confined in a transparent enclo- sure, a rectangular box in this paper. • We develop a new 3D reconstruction method, called ReNeuS, that can handle multiple light reflections and refractions at intersecting interfaces. ReNeuS employs a novel hybrid rendering strategy to combine the light intensities from the segmented inner and outer sub- spaces. • We evaluate our method on both synthetic as well as a newly captured dataset that contains 10real insect specimens enclosed in a transparent glass container.

Sun_Stimulus_Verification_Is_a_Universal_and_Effective_Sampler_in_Multi-Modal_CVPR_2023

Abstract To comprehensively cover the uncertainty of the future, the common practice of multi-modal human trajectory pre- diction is to first generate a set/distribution of candidate fu- ture trajectories and then sample required numbers of tra- jectories from them as final predictions. Even though a large number of previous researches develop various strong models to predict candidate trajectories, how to effectively sample the final ones has not received much attention yet. In this paper, we propose stimulus verification, serving as a universal and effective sampling process to improve the multi-modal prediction capability, where stimulus refers to the factor in the observation that may affect the future movements such as social interaction and scene context. Stimulus verification introduces a probabilistic model, de- noted as stimulus verifier, to verify the coherence between a predicted future trajectory and its corresponding stimulus. By highlighting prediction samples with better stimulus- coherence, stimulus verification ensures sampled trajecto- ries plausible from the stimulus’ point of view and there- fore aids in better multi-modal prediction performance. We implement stimulus verification on five representative pre- diction frameworks and conduct exhaustive experiments on three widely-used benchmarks. Superior results demon- strate the effectiveness of our approach.

1. Introduction Human trajectory prediction [3, 8, 9, 17, 35, 38, 39, 43] is a vital task towards autonomous driving systems and social robots, bridging the perception and decision mod- ules [4, 26, 27, 42]. Due to the fact that there is no single correct future, one of the most important aspects of trajec- tory prediction lies in multi-modal prediction, studying to *contributed equally §Cewu Lu is the corresponding author, the member of Qing Yuan Re- search Institute and MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China and Shanghai Qi Zhi institute. ℳ 𝛩 (a) (b) Possible Prediction Impossible Prediction Observed Trajectory Sampled Trajectory Distribution SetFigure 1. (a) Multi-modal trajectory prediction pipeline, including two major components: base prediction model Mand sampler Θ. (b) Examples of stimulus(context)-inconsistent trajectories sam- pled from a stimulus(context)-aware base prediction model, So- phie [33]. Predictions in red leads to impossible regions. cover multiple distinct future possibilities with finite pre- dictions. Specifically, for an observation, a set/distribution of candidate trajectories is first generated by a base predic- tion model, after which final ones of required number are sampled from the candidates, seeing Fig. 1-a. To guarantee the performance of multi-modal trajectory prediction, countless efforts have been put into the design of prediction models in prior works using either gener- ative or classification frameworks. For generative ones, some popular architectures such as GAN [7], V AE [14] and Normalizing Flow [29] have been applied and im- proved in plenty of studies on multi-modal trajectory pre- diction [8, 23, 33, 34, 40]. In the meantime, impressive re- sults are also achieved with classification models such as MultiPath [3] and PCCSNet [39]. Yet it is often neglected that an effective sampling strategy, which selects the ap- propriate trajectories from the collection of candidates, also has a great impact on the prediction performance. Typically, random or Monte-Carlo sampling is adopted as the default one. Only a handful of works [2, 22, 46] have studied 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. 22014 sampling strategy, while they all seek to solve this prob- lem by searching better latent vectors for generative frame- works. Although they have achieved remarkable outcomes, the applications of such approaches are strongly limited as they are confined to generative frameworks only, ignoring the other mainstream frameworks, i.e.classification ones. Besides, there are some factors in the observation that may influence the future movements. These factors, such as social interaction and scene context, are referred as stim- uli[32] and their significance of guiding accurate predic- tions has been proven in numerous researches [15, 25, 33, 34]. Although base frameworks can learn from stimu- lus information to generate candidate predictions, if such information is missing in the sampling process, out-of- distribution predictions are still highly probable to be sam- pled and harm the overall accuracy, as illustrated in Fig. 1-b. To raise a universal sampler taking stimulus into con- sideration, in this paper, we propose an explicit sampling process called stimulus verification. It first verifies the co- herence between a candidate trajectory and its correspond- ing stimulus, and then samples highly stimulus-coherent ones as final results. Since the verification is an external process beyond the base prediction model, it can be uni- versally applied to any multi-modal prediction frameworks. And meanwhile, stimulus information is successfully intro- duced, ensuring that the selected samples comply with the constrains of the involved stimuli. Stimulus verification is built on a conditionally parame- terized probabilistic model optimized by Maximum Likeli- hood Estimation, namely stimulus verifier. By feeding the candidate trajectory along with its stimulus into the veri- fier, corresponding output likelihood reveals the trajectory- stimulus coherence, where a higher likelihood indicates bet- ter coherence. To this end, we can map the likelihood to a coherence score, which further serves as the basis to sample more stimulus-coherent predictions from a large group of trajectory candidates. Our method is simple to implement for any base prediction framework in a plug-and-play manner. We implement our approach on five representative frame- works [8, 29, 33, 39, 43] and conduct exhaustive experi- ments on three widely-used trajectory prediction bench- marks, i.e.ETH [28]/UCY [16] Dataset, Grand Central Sta- tion Dataset [47] and NBA SportVU Dataset [20], to vali- date our approach. Superior results confirm the effective- ness of stimulus verification.

Song_OPE-SR_Orthogonal_Position_Encoding_for_Designing_a_Parameter-Free_Upsampling_Module_CVPR_2023

Abstract Arbitrary-scale image super-resolution (SR) is often tackled using the implicit neural representation (INR) ap- proach, which relies on a position encoding scheme to im- prove its representation ability. In this paper, we introduce orthogonal position encoding (OPE), an extension of po- sition encoding, and an OPE-Upscale module to replace the INR-based upsampling module for arbitrary-scale im- age super-resolution. Our OPE-Upscale module takes 2D coordinates and latent code as inputs, just like INR, but does not require any training parameters. This parameter- free feature allows the OPE-Upscale module to directly perform linear combination operations, resulting in con- tinuous image reconstruction and achieving arbitrary-scale image reconstruction. As a concise SR framework, our method is computationally efficient and consumes less mem- ory than state-of-the-art methods, as confirmed by exten- sive experiments and evaluations. In addition, our method achieves comparable results with state-of-the-art methods in arbitrary-scale image super-resolution. Lastly, we show that OPE corresponds to a set of orthogonal basis, validat- ing our design principle.1

1. Introduction Photographs are composed of discrete pixels of vary- ing precision due to the limitations of sampling frequency, which breaks the continuous visual world into discrete parts. The single image super-resolution (SISR) task aims to restore the original continuous world in the image as much as possible. In an arbitrary-scale SR task, one of- ten reconstructs the continuous representation of a low- resolution image and then adjusts the resolution of the target image as needed. The recent rise of implicit neural repre- sentation (INR) in 3D vision has enabled the representation *Corresponding author. 1Project page: https://github.com/gaochao-s/ope-srof complex 3D objects and scenes in a continuous man- ner [14,19,41,42,44,45,47,49,57,58], which also opens up possibilities for continuous image and arbitrary-scale image super-resolution [5, 18, 32, 72]. Existing methods for arbitrary-scale SR typically use a post-upsampling framework [70]. In this approach, low- resolution (LR) images first pass through a deep CNN network (encoder) without improving the resolution, and then pass through an INR-based upsampling module (de- coder) with a specified target resolution to reconstruct high- resolution (HR) images. The decoder establishes a mapping from feature maps (the output of encoder) to target image pixels using a pre-assigned grid partitioning and achieves arbitrary-scale with the density of the grid in Cartesian co- ordinate system. However, the INR approach has a defect of learning low-frequency information, also known as spectral bias [50]. To address this issue, sinusoidal positional en- coding is introduced to embed input coordinates to higher dimensions and enable the network to learn high-frequency details. This inspired recent works on arbitrary-scale SR to further improve the representation ability [32, 72]. Despite its effectiveness in arbitrary-scale SR, the INR- based upsampling module increases the complexity of the entire SR framework as two different networks are jointly trained. Additionally, as a black-box model, it represents a continuous image with a strong dependency on both the feature map and the decoder (e.g., MLP). However, its rep- resentation ability decreases after flipping the feature map, a phenomenon known as flipping consistency decline. As shown in Fig. 1, flipping the feature map horizontally be- fore the upsampling module of LIIF results in a blurred tar- get image that does not have the expected flip transforma- tion. This decline could be due to limitations of the MLP in learning the symmetry feature of the image. MLP is a universal function approximator [17], which tries to fit a mapping function from feature map to the con- tinuous image, therefore, it is reasonable to assume that such process could be solved by an analytical solution. In this paper, we re-examine position encoding from the per- 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. 10009 spective of orthogonal basis and propose orthogonal posi- tion encoding (OPE) for continuous image representation. The linear combination of 1D latent code and OPE can directly reconstruct continuous image patch without using implicit neural function [5]. To prove OPE’s rationality, we analyse it both from functional analysis and 2D-Fourier transform. We further embed it into a parameter-free up- sampling module, called OPE-Upscale Module, to replace INR-based upsampling module in deep SR framework, then currently deep SR framework can be greatly simplified. Unlike the state-of-the-art method by Lee et al. [32], which enhances MLP with position encoding, we explore the possibility of extending position encoding without MLP. By providing a more concise SR framework, our method achieves high computing efficiency and consumes less memory than the state-of-the-art, while also achieving com- parable image performance in arbitrary-scale SR tasks. Our contributions are as follows: • We propose a novel position encoding, called orthogo- nal position encoding (OPE), which takes the form of a 2D-Fourier series and corresponds to a set of orthog- onal basis. Building on OPE, we introduce the OPE- Upscale Module, a parameter-free upsampling module for arbitrary-scale image super-resolution. • Our method significantly reduces the consumption of computing resources, resulting in high computing effi- ciency for arbitrary-scale SR tasks. • The OPE-Upscale Module is interpretable, parameter- free and does not require training, resulting in a con- cise SR framework that elegantly solves the flipping consistency problem. • Extensive experiments demonstrate that our method achieves comparable results with the state-of-the-art. Furthermore, our method enables super-resolution up to a large scale of ×30.

Sosea_MarginMatch_Improving_Semi-Supervised_Learning_with_Pseudo-Margins_CVPR_2023

Abstract We introduce MarginMatch, a new SSL approach combin- ing consistency regularization and pseudo-labeling, with its main novelty arising from the use of unlabeled data train- ing dynamics to measure pseudo-label quality. Instead of using only the model’s confidence on an unlabeled example at an arbitrary iteration to decide if the example should be masked or not, MarginMatch also analyzes the behavior of the model on the pseudo-labeled examples as the training progresses, to ensure low quality predictions are masked out. MarginMatch brings substantial improvements on four vi- sion benchmarks in low data regimes and on two large-scale datasets, emphasizing the importance of enforcing high- quality pseudo-labels. Notably, we obtain an improvement in error rate over the state-of-the-art of 3.25% on CIFAR-100 with only 25labels per class and of 3.78% on STL-10 using as few as 4labels per class. We make our code available at https://github.com/tsosea2/MarginMatch .

1. Introduction Deep learning models have seen tremendous success in many vision tasks [14, 22, 27, 42, 43]. This success can be attributed to their scalability, being able to produce better re- sults when they are trained on large datasets in a supervised fashion [15, 27, 34, 35, 43, 47]. Unfortunately, large labeled datasets annotated for various tasks and domains are diffi- cult to acquire and demand considerable annotation effort or domain expertise. Semi-supervised learning (SSL) is a powerful approach that mitigates the requirement for large labeled datasets by effectively making use of information from unlabeled data, and thus, has been studied extensively in vision [4, 5, 23, 25, 30, 36, 38, 39, 44–46]. Recent SSL approaches integrate two important com- ponents: consistency regularization [46, 49] and pseudo- labeling [25]. Consistency regularization works on the as- sumption that a model should output similar predictions when fed perturbed versions of the same image, whereas pseudo-labeling uses the model’s predictions of unlabeled examples as labels to train against. For example, Sohn et al. [41] introduced FixMatch that combines consistency reg-ularization on weak and strong augmentations with pseudo- labeling. FixMatch relies heavily on a high-confidence threshold to compute the unsupervised loss, disregarding any pseudo-labels whose confidence falls below this threshold. While training using only high-confidence pseudo-labels has shown to consistently reduce the confirmation bias [1], this rigid threshold allows access only to a small amount of un- labeled data for training, and thus, ignores a considerable amount of unlabeled examples for which the model’s predic- tions do not exceed the confidence threshold. More recently, Zhang et al. [49] introduced FlexMatch that relaxes the rigid confidence threshold in FixMatch to account for the model’s learning status of each class in that it adaptively scales down the threshold for a class to encourage the model to learn from more examples from that class. The flexible thresh- olds in FlexMatch allow the model to have access to a much larger and diverse set of unlabeled data to learn from, but lowering the thresholds can lead to the introduction of wrong pseudo-labels, which are extremely harmful for generaliza- tion. Interestingly, even when the high-confidence threshold is used in FixMatch can result in wrong pseudo-labels. See Figure 1 for incorrect pseudo-labels detected in the training set after we apply FixMatch and FlexMatch on ImageNet. We posit that a drawback of FixMatch and FlexMatch and in general of any pseudo-labeling approach is that they use the confidence of the model only at the current iteration to enforce quality of pseudo-labels and completely ignore model’s predictions at prior iterations. In this paper, we propose MarginMatch, a new SSL ap- proach that monitors the behavior of the model on the unla- beled examples as the training progresses, from the begin- ning of tr

Tao_GALIP_Generative_Adversarial_CLIPs_for_Text-to-Image_Synthesis_CVPR_2023

Abstract Synthesizing high-fidelity complex images from text is challenging. Based on large pretraining, the autoregres- sive and diffusion models can synthesize photo-realistic im- ages. Although these large models have shown notable progress, there remain three flaws. 1) These models re- quire tremendous training data and parameters to achieve good performance. 2) The multi-step generation design slows the image synthesis process heavily. 3) The synthe- sized visual features are challenging to control and require delicately designed prompts. To enable high-quality, effi- cient, fast, and controllable text-to-image synthesis, we pro- pose Generative Adversarial CLIPs, namely GALIP . GALIP leverages the powerful pretrained CLIP model both in the discriminator and generator. Specifically, we propose a CLIP-based discriminator. The complex scene understand- ing ability of CLIP enables the discriminator to accurately assess the image quality. Furthermore, we propose a CLIP- empowered generator that induces the visual concepts from CLIP through bridge features and prompts. The CLIP- integrated generator and discriminator boost training ef- ficiency, and as a result, our model only requires about 3% training data and 6%learnable parameters, achieving com- parable results to large pretrained autoregressive and diffu- sion models. Moreover, our model achieves ∼120×faster synthesis speed and inherits the smooth latent space from GAN. The extensive experimental results demonstrate the excellent performance of our GALIP . Code is available at https://github.com/tobran/GALIP .

1. Introduction Over the last few years, we have witnessed the great suc- cess of generative models for various applications [4, 47]. Among them, text-to-image synthesis [3, 5, 16, 19–22, 26, 29, 30, 34, 43, 48, 50–53, 60] is one of the most appealing *Corresponding Author Figure 1. (a) Existing text-to-image GANs conduct adversarial training from scratch. (b) Our proposed GALIP conducts adver- sarial training based on the integrated CLIP model. applications. It generates high-fidelity images according to given language guidance. Owing to the convenience of lan- guage for users, text-to-image synthesis has attracted many researchers and has become an active research area. Based on a large scale of data collections, model size, and pretraining, recently proposed large pretrained autore- gressive and diffusion models, e.g., DALL-E [34] and LDM [36], show the impressive generative ability to syn- thesize complex scenes and outperform the previous text-to- image GANs significantly. Although these large pretrained generative models have achieved significant advances, they still suffer from three flaws. First, these models require tremendous training data and parameters for pretraining. The large data and model size brings an extremely high computing budget and hardware requirements, making it inaccessible to many researchers and users. Second, the generation of large models is much slower than GANs. The token-by-token generation and progressive denoising require hundreds of inference steps and make the generated results lag the language inputs seriously. Third, there is no intuitive smooth latent space as GANs, which maps mean- ingful visual attributes to the latent vector. The multi-step generation design breaks the synthesis process and scatters the meaningful latent space. It makes the synthesis process require delicately designed prompts to 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. 14214 To address the above limitations, we rethink Generative Adversarial Networks (GAN). GANs are much faster than autoregressive and diffusion models and have smooth latent space, which enables more controllable synthesis. How- ever, GAN models are known for potentially unstable train- ing and less diversity in the generation [7]. It makes current text-to-image GANs suffer from unsatisfied synthesis qual- ity under complex scenes. In this work, we introduce the pretrained CLIP [31] into text-to-image GANs. The large pretraining of CLIP brings two advantages. First, it enhances the complex scene un- derstanding ability. The pretraining dataset has many com- plex images under different scenes. Armed with the Vision Transformer (ViT) [9], the image encoder can extract in- formative and meaningful visual features from complex im- ages to align the corresponding text descriptions after ade- quate pretraining. Second, the large pretraining dataset also enables excellent domain generalization ability. It contains various kinds of images, e.g., photos, drawings, cartoons, and sketches, collected from a variety of publicly available sources. The various images make the CLIP model can map different kinds of images to the shared concepts and en- able impressive domain generalization and zero-shot trans- fer ability. These two advantages of CLIP, complex scene understanding and domain generalization ability, motivate us to build a more powerful text-to-image model. We propose a novel text-to-image generation framework named Generative Adversarial CLIPs (GALIP). As shown in Figure 1, the GALIP integrates the CLIP model [31] in both the discriminator and generator. To be specific, we pro- pose the CLIP-based discriminator and CLIP-empowered generator. The CLIP-based discriminator inherits the com- plex scene understanding ability of CLIP [31]. It is com- posed of a frozen ViT-based CLIP image encoder (CLIP- ViT) and a learnable mate-discriminator (Mate-D). The Mate-D is mated to the CLIP-ViT for adversarial training. To retain the knowledge of complex scene understanding in the CLIP-ViT, we freeze its weights and collect the pre- dicted CLIP image features from different layers. Then, the Mate-D further extracts informative visual features from collected CLIP features to distinguish the synthesized and real images. Based on the complex scene understanding ability of CLIP-ViT and the continuous analysis of Mate- D, the CLIP-based discriminator can assess the quality of generated complex images more accurately. Furthermore, we propose the CLIP-empowered genera- tor, which exerts the domain generalization ability of CLIP [31]. It is hard for the generator to synthesize complex im- ages directly. Some works employ sketch [11] and lay- out [21, 23] as bridge domains to alleviate the difficulty. However, such a design requires additional labeled data. Different from these works, the excellent domain general- ization of CLIP [31] motivates us that there may be an im-plicit bridge domain, which is easier to synthesize but can be mapped to the same visual concepts through the CLIP- ViT. Thus, we design the CLIP-empowered generator. It is composed of a frozen CLIP-ViT and a learnable mate- generator (Mate-G). The Mate-G first predicts the implicit bridge features from text and noise. Then the bridge feature will be mapped to the visual concepts through CLIP-ViT. Furthermore, we add some text-conditioned prompts to the CLIP-ViT for task adaptation. The predicted visual con- cepts close the gap between text features and target images which enhances the complex image synthesis ability. Overall, our contributions can be summarized as follows: • We propose an efficient, fast, and more controllable model for text-to-image synthesis that can synthesize high-quality complex images. • We propose the CLIP-based discriminator, which as- sesses the quality of complex images more accurately. • We propose the CLIP-empowered generator, which syn- thesizes images based on text features and predicted CLIP visual features. • Extensive experiments demonstrate that the proposed GALIP can achieve comparable performance with large pertaining models based on significantly smaller compu- tational costs.

Sun_Ultrahigh_Resolution_ImageVideo_Matting_With_Spatio-Temporal_Sparsity_CVPR_2023

Abstract Commodity ultrahigh definition (UHD) displays are be- coming more affordable which demand imaging in ultrahigh resolution (UHR). This paper proposes SparseMat , a com- putationally efficient approach for UHR image/video mat- ting. Note that it is infeasible to directly process UHR im- ages at full resolution in one shot using existing matting algorithms without running out of memory on consumer- level computational platforms, e.g., Nvidia 1080Ti with 11G memory, while patch-based approaches can introduce un- sightly artifacts due to patch partitioning. Instead, our method resorts to spatial and temporal sparsity for ad- dressing general UHR matting. When processing videos, huge computation redundancy can be reduced by exploit- ing spatial and temporal sparsity. In this paper, we show how to effectively detect spatio-temporal sparsity, which serves as a gate to activate input pixels for the matting model. Under the guidance of such sparsity, our method with sparse high-resolution module (SHM) can avoid patch- based inference while memory efficient for full-resolution matte refinement. Extensive experiments demonstrate that SparseMat can effectively and efficiently generate high- quality alpha matte for UHR images and videos at the orig- inal high resolution in a single pass. Project page is in https://github.com/nowsyn/SparseMat.git.

1. Introduction Ultrahigh resolution (UHR) matting is an important problem [36, 49], and with increasing demand due to the fast advent and accessibility of commodity ultrahigh def- inition displays in real-world applications, such as gam- ing, TV/movie post-production, and image/video editing, UHR matting becomes ever relevant. However, modern consumer-level GPU and mobile devices still have limited hardware resources. Despite good technical contributions, guided filters and patch-based techniques (Figure 1) are not applicable, when unsightly blurry and seams artifacts are unacceptable in UHR imaging. Matting is a primary technique for image/video editing and plays an important role in many applications. The goal of matting is to extract a detailed alpha matte of the fore-ground object from a given image/video. The matted fore- ground can be composited on other background images. As known, matting is an ill-posed problem defined as Equa- tion 1 with the given image I, foreground F, background B and alpha α∈[0,1]to be extracted: I=αF+ (1−α)B. (1) Most of the existing state-of-the-art matting meth- ods [28, 33, 48] take the whole image as input in a forward pass, and thus the resolution they can handle is bounded by available memory. Given limited memory, to process UHR images, a straightforward approach is to first process the downsampled input image which will inevitably lead to blurry artifacts. Thus, super-resolution methods, such as guided filter (GF) [21] or deep guided filter (DGF) [47], have been proposed to recover missing details. However, guided filter or deep guided filter easily produces fuzzy ar- tifacts when processing complex hairy structures as shown in Figure 1-(a). Patch-based inference [23, 35] is another plausible strategy. However, small patch can cause artifacts due to insufficient global context and inconsistent local con- text as shown in Figure 1-(b). On the other hand, using large patch (e.g., 2048) with large overlap produces defective al- pha matte with missing details or blurry artifacts in long hair region due to the lack of long-range dependency in UHR images as shown Figure 1-(b), not to mention that heavy computation and memory overhead are introduced with in- creasing patch size making some methods cannot even run, as shown in Figure 1-(c). In conclusion, super resolving based on (deep) guided filter or patch-based inference are not ideal choices for handling UHR matting. In this paper we propose a general image/video matting framework SparseMat to address the problem of UHR mat- ting, which is both memory and computation efficient while generating high-quality alpha mattes. The core idea of our method is to skip a large amount of redundant computation on many pixels during processing UHR images or videos. In general, our method takes the low-resolution prior as input to generate spatio-temporal sparsity, which serves as the gate to activate pixels consumed by a sparse convolution module. Both temporal and spatial information contribute to the sparsity estimation. Specifically, we compute color difference between adjacent frames to obtain the temporal 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. 14112 Image LPN + SHM (Ours) LPNLPN + GF (a)Patch=512 Patch=2048Image Full-resolution(b)20040060080010001200 25651210242048GMACsPatchsize0246810 25651210242048Memory(GB)Patchsize (c) RVMBGMv2Figure 1. (a)Blurry artifacts when UHR images are not matted in the original full resolution: guided filter (GF), deep guided filter (RVM [36]), small patch replacement method (BGMv2 [35]). The low-resolution alpha matte obtained by the self-trained low-resolution prior network (LPN) is for reference here. Our sparse high-resolution module (SHM) produces high-quality alpha matte for UHR matting. (b)Seam artifacts of patch-based inference from FBA [13] under different patch sizes with a fixed ratio (1/8) of overlap. Full-resolution result is ours. (c)Memory consumption and computation (GMACs) with different patch sizes. When the patch size exceeds 2K, some methods cannot even run on Nvidia 1080Ti with 11G memory. sparsity. For the spatial sparsity, we can derive from any lightweight low-resolution matting model such as [28, 36]. The two sparsity maps are combined together to determi- nate the pixels which need high-resolution processing by the sparse module. Unlike super-resolving strategy per- formed on the whole image, our method with sparse high- resolution module (SHM) can safely skip expensive com- putations in large solid pixel regions, only paying attention to irregular, sparse and (oftentimes) thin border regions sur- rounding the object or transitional regions within the ob- ject. In contrast to restrictive views of patch-based strategy to local regions, our method takes a more global perspec- tive of the foreground object thus avoiding potential arti- facts due to inadequate context consideration. This design allows SparseMat to process an ultrahigh resolution image or video frame in only one shot without suffering any in- formation loss caused by down-sampling or patch partition, which thus produces high-quality alpha matte for ultrahigh resolution images or videos. Our contributions are summarized below: 1. This is the first work for general UHR image/video matting which enables full-resolution inference in one shot without running out of memory, thus eliminating the need of patch partitioning and patch artifacts. 2. We show how to obtain accurate spatio-temporal spar- sity for general UHR matting, which has never been adequately discussed in previous works. In addition, this is the first work that proposes to apply sparse con- volution network to skip unnecessary computations in dealing with UHR matting. 3. We conduct extensive experiments in multiple popu- lar image/video matting datasets, including Adobe Im- age Matting Dataset [48], VideoMatte240K [42] and our self-collected UHR matting dataset, and providepromising qualitative results, which demonstrate the superiority of our SparseMat in dealing with general image/video matting.

Tian_Multi-Object_Manipulation_via_Object-Centric_Neural_Scattering_Functions_CVPR_2023

Abstract Learned visual dynamics models have proven effective for robotic manipulation tasks. Yet, it remains unclear how best to represent scenes involving multi-object interactions. Current methods decompose a scene into discrete objects, but they struggle with precise modeling and manipulation amid challenging lighting conditions as they only encode appearance tied with specific illuminations. In this work, we propose using object-centric neural scattering functions (OSFs) as object representations in a model-predictive con- trol framework. OSFs model per-object light transport, en- abling compositional scene re-rendering under object re- arrangement and varying lighting conditions. By combin- ing this approach with inverse parameter estimation and graph-based neural dynamics models, we demonstrate im- proved model-predictive control performance and general- ization in compositional multi-object environments, even in previously unseen scenarios and harsh lighting conditions.

1. Introduction Predictive models are the core components of many robotic systems for solving inverse problems such as plan- ning and control. Physics-based models built on first prin- ciples have shown impressive performance in domains such as drone navigation [3] and robot locomotion [28]. How- ever, such methods usually rely on complete a priori knowl- edge of the environment, limiting their use in complicated manipulation problems where full-state estimation is com- plex and often impossible. Therefore, a growing number of approaches alternatively propose to learn dynamics models directly from raw visual observations [2, 12, 14, 17, 21, 48]. Although using raw sensor measurements as inputs to predictive models is an attractive paradigm as they are readily available, visual data can be challenging to work with directly due to its high dimensionality. Prior meth- ods proposed to learn dynamics models over latent vec- indicates equal contribution. Yancheng is affiliated with Fudan Uni- versity; this work was done while he was a summer intern at Stanford. Typical visual manipulation settingSimulatedRealOur settings Figure 1. While typically studied visual manipulation settings are carefully controlled environments, we consider scenarios with varying and even harsh lighting, in addition to novel object con- figurations, that are more similar to real-world scenarios. tors, demonstrating promising results in a range of robotics tasks [17, 18, 44, 52]. However, with multi-object interac- tions, the underlying physical world is 3D and composi- tional. Encoding everything into a single latent vector fails to consider the relational structure within the environment, limiting its generalization outside the training distribution. Another promising strategy is to build more structured visual representations of the environment, including the use of particles [31, 33, 36], keypoints [32, 37, 38], and object meshes [22]. Among the structured representations, Driess et al. [10] leveraged compositional neural implicit represen- tations in combination with graph neural networks (GNNs) for the dynamic modeling of multi-object interactions. The inductive bias introduced by GNNs captures the environ- ment’s underlying structure, enabling generalization to sce- narios containing more objects than during training, and the neural implicit representations allow precise estimation and modeling of object geometry and interactions. However, Driess et al. [10] only considered objects of uniform color 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. 9021 in well-lit scenarios. It is unclear how the method works for objects with more complicated geometries and textures. The lack of explicit modeling of light transport also limits its use in scenarios of varying lighting conditions, especially those vastly different from the training distributions. In this paper, we propose to combine object-centric neural scattering functions (OSFs) [57] and graph neural networks for the dynamics modeling and manipulation of multi-object scenes. OSFs explicitly model light transport and learn to approximate the cumulative radiance transfer, which allows relighting and inverse estimation of scenes involving multiple objects and the change of lights, such as those shown in Figure 1. Combined with gradient-free evolutionary algorithms like covariance matrix adaption (CMA), the learned neural implicit scattering functions sup- port inverse parameter estimation, including object poses and light directions, from visual observations. Based on the estimated scene parameters, a graph-based neural dynamics model considers the interactions between objects and pre- dicts the evolution of the underlying system. The predictive model can then be used within a model-predictive control (MPC) framework for downstream manipulation tasks. Experiments demonstrate that our method performs more accurate reconstruction in harsh lighting conditions compared to prior methods, producing higher-fidelity long horizon prediction compared to video prediction models. When combined with inverse parameter estimation, our en- tire control pipeline improves on simulated object manipu- lation tasks in settings with varying lighting and previously unseen object configurations, compared to performing MPC directly in image space. We make three contributions. First, the use of neural scattering functions supports inverse parameter estimation in scenarios with challenging and previously unseen light- ing conditions. Second, our method models the composi- tionality of the underlying scene and can make long-term future predictions about the system’s evolution to support downstream planning tasks. Third, we conduct and show successful manipulation of simulated multi-object scenes involving extreme lighting directions.

Tang_Label_Information_Bottleneck_for_Label_Enhancement_CVPR_2023

Abstract In this work, we focus on the challenging problem of La- bel Enhancement (LE), which aims to exactly recover label distributions from logical labels, and present a novel Label Information Bottleneck (LIB) method for LE. For the recov- ery process of label distributions, the label irrelevant infor- mation contained in the dataset may lead to unsatisfactory recovery performance. To address this limitation, we make efforts to excavate the essential label relevant information to improve the recovery performance. Our method formu- lates the LE problem as the following two joint processes: 1) learning the representation with the essential label rel- evant information, 2) recovering label distributions based on the learned representation. The label relevant informa- tion can be excavated based on the “bottleneck” formed by the learned representation. Significantly, both the label rel- evant information about the label assignments and the label relevant information about the label gaps can be explored in our method. Evaluation experiments conducted on several benchmark label distribution learning datasets verify the ef- fectiveness and competitiveness of LIB. Our source codes are available at https://github.com/qinghai- zheng/LIBLE

1. Introduction Learning with label ambiguity is important in computer vision and machine learning. Different from the traditional Multi-Label Learning (MLL), which employs multiple log- ical labels to annotate one instance to address the label am- biguity issue [20], Label Distribution Learning (LDL) con- siders the relative importance of different labels and draws much attention in recent years [6, 8, 14, 18, 26]. By distin- guishing the description degrees of all labels, LDL anno- tates one instance with a label distribution. Therefore, LDL is a more general learning paradigm, MLL can be regarded as a special case of LDL [8, 10, 12]. *Corresponding author, E-mail: [email protected], many LDL methods are proposed and achieve great success in practice [3, 9, 14, 18]. Instances with exact label distributions are vital for the training process of LDL methods. Nevertheless, annotating instances with label dis- tributions is time-consuming [24,28]. We take the label dis- tribution annotation process of SJAFFE dataset for example here. SJAFFE dataset is the facial expression dataset, which contains 213 grayscale images collected from 10 Japanese female models, each facial expression image is rated by 60 persons on 6 basic emotions, including happiness, surprise, sadness, fear, anger, and disgust, with a five-level scale from 1 - 5, the higher value indicates the higher emotion intensity. Consequently, the average score of each emotion is served as the emotion label distribution [14,28]. Clearly, the above annotation process is costly and it is unpractical to annotate data with label distributions manually, especially when the number of data is large. Fortunately, most existing datasets in the field of computer vision and machine learning are an- notated by single-label or multi-labels [7, 29], therefore, a highly recommended promising solution is Label Enhance- ment (LE), which attempts to recover the desired label dis- tributions exactly from existing logical labels [24, 28, 32]. Driven by the urgent requirement of obtaining label dis- tributions and the convenience of LE, some LE methods are proposed in recent years [5, 7, 11, 13, 15, 17, 21, 24, 28, 29]. Given a dataset X={x1,x2,···,xn} ∈Rq×n, in which qandndenote the number of dimensions and the number of instances, the potential label set is {y1, y2,···, yc}. The available logical labels and the desired distribution labels ofXare separately indicated by L={l1,l2,···,ln}and D={d1,d2,···,dn}, where lianddiare: li= (ly1 i, ly2 i,···, lyc i)T,di= (dy1 i, dy2 i,···, dyc i)T.(1) To be specific, LE aims to recover Dbased on the informa- tion provided by XandL. For most existing LE methods, their objectives can be concisely summarized as follows: min θ∥fθ(X)−L∥2 F+γreg(fθ(X)), (2) in which D=fθ(X),fθ(·)indicates the mapping from XtoD,reg(·)denotes the regularization function, and γ 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. 7497 Logical label Description Degree Figure 1. Illustration of label relevant information. Excavating the essential label relevant information directly is challenging and we adopt a indirect way here. We jointly investigate the information about the assignments of labels to the instance and the informa- tion about the label gaps between logical labels and label distribu- tions. Given the i-th instance xi, the label gap of the yjlabel is δyj i=lyj i−dyj i. The information contained in lyj iandδyj ican be amalgamated to form the essential label relevant information. In other words, we employ lyj ito explore the label relevant infor- mation about the label assignments and δyj ito excavate the label relevant information about the label gaps. To a certain degree, the combination of δyj iandlyj iis equivalent to dyj i. As depicted here, lyj iindicates that y1andy3are related labels and δyj iprovides the importance of y1andy3. is the trade-off parameter. Most existing LE methods vary inreg(·). For example, GLLE [28] calculates the distance- based similarity matrix of data and employs the smoothness assumption [33] to construct reg(·); LESC [24] considers the global sample correlations and introduces the low-rank constraint as the regularization; PNLR [13] leverages reg(·) to maintain positive and negative label relations during the recovery process. Although a remarkable progress can be made by aforementioned methods, they ignore the label ir- relevant information contained in X, which prevents the further improvement of recovery results. For example, in the LE task of recovering facial age label distributions, the label irrelevant information, such as specularities informa- tion, cast shadows information, and occlusions information, may result in the incorrect mapping process of fθ(·)and the unsuitable regularization of reg(·), eventually leads to the unsatisfactory recovery performance. To overcome the aforementioned limitation, we present a Label Information Bottleneck (LIB) method for LE. Con- cretely, the core idea of LIB is to learn the latent representa- tionH, which preserves the maximum label relevant infor- mation, from X, and jointly recovers the label distributions based on the latent representation. For the LE problem, the label relevant information is the information that describes the description degrees of labels. It is tough to explore the label relevant information directly. As shown in Fig. 1, wedecompose the label relevant information into two compo- nents, namely the assignments of labels to the instance and the label gaps between label distributions and logical labels. Inspired by Information Bottleneck (IB) [25], LIB utilizes the existing logical labels to explore the information about the assignments of labels to the instance. Unlike simply em- ploying the original IB on the LE task, our method further considers the information about the label gaps between la- bel distributions and logical labels. It is noteworthy that the above two components of the label relevant information are jointly explored in our method, and that is why we term the proposed method Label Information Bottleneck (LIB). The main contributions can be summarized as follows: •We decompose the label relevant information into the information about the assignments of labels to instance and the information about the label gaps between logi- cal labels, both of which can be jointly explored during the learning process of our method. •We introduce a novel LE method, termed LIB, which excavates the label relevant information to exactly re- cover the label distributions. Based on the original IB, which explores the label assignments information for LE, LIB further explores the label gaps information. •We verify the effectiveness of LIB by performing ex- tensive experiments on several datasets. Experimental results show that the proposed method can achieve the competitive performance, compared to state-of-the-art LE methods.

Tu_Toward_Accurate_Post-Training_Quantization_for_Image_Super_Resolution_CVPR_2023

Abstract Model quantization is a crucial step for deploying super resolution (SR) networks on mobile devices. However, exist- ing works focus on quantization-aware training, which re- quires complete dataset and expensive computational over- head. In this paper, we study post-training quantization (PTQ) for image super resolution using only a few unla- beled calibration images. As the SR model aims to maintain the texture and color information of input images, the distri- bution of activations are long-tailed, asymmetric and highly dynamic compared with classification models. To this end, we introduce the density-based dual clipping to cut off the outliers based on analyzing the asymmetric bounds of acti- vations. Moreover, we present a novel pixel aware calibra- tion method with the supervision of the full-precision model to accommodate the highly dynamic range of different sam- ples. Extensive experiments demonstrate that the proposed method significantly outperforms existing PTQ algorithms on various models and datasets. For instance, we get a 2.091 dB increase on Urban100 benchmark when quantiz- ing EDSR4 to 4-bit with 100 unlabeled images. Our code is available at both PyTorch and MindSpore.

1. Introduction Image super resolution (SR) is a classical image pro- cessing task in computer vision, which reconstructs high- resolution (HR) images from the corresponding low- resolution (LR) images. SR has been widely applied in the real-world scenarios, such as medical imaging [12, 35], surveillance [1, 49], satellite imagery [31, 36] and smart- phone display [8, 19]. With the rapid development of deep learning in recent years, SR models with deep neural net- work (DNN) structure have continued to achieve state-of- the-art performance on various datasets. However, these SR models require significant storage and computational re- sources, which makes their deployment on mobile devices extremely difficult. To improve the inference efficiency, various techniques have been proposed to compress the models, such as network pruning [16, 50], model quantiza-Table 1. Computational overhead of different quantization meth- ods on EDSR model. The FP denotes full-precision training, the Gt denotes the ground-truth, and the Bs denotes batch size. Method Type Data Gt Bs Iters Run time EDSR [28] FP 800 316 15,000 240 PAMS [25] QAT 800 316 1,500 24 FQSR [40] QAT 800 316 15,000 120 CADyQ [14] QAT 800 3 8 30,000 240 DAQ [15] QAT 800 3 4 300,000 1200 DDTB [52] QAT 800 316 3,000 48 Ours PTQ 100 5 2 500 1 tion [13, 38], compact architecture design [8, 9] and knowl- edge distillation [29, 41, 45, 46]. Among these approaches, model quantization is much benefit to existing artificial in- telligent (AI) accelerators [3, 42], which generally focus on low-precision arithmetic, resulting in lower latency, smaller memory footprint and less energy consumption. Although the previous SR quantization methods make great effort on improving the performance with given bit- width, their main drawback is that they require quantiza- tion aware training (QAT) with complete datasets and ex- pensive computational overhead. As shown in Table 1, the full-precision EDSR model needs to train 15,000 iters with the batch size of 16, takes 8 days on NVIDIA Titan X GPUs [28]. To recover the performance drop of the quan- tized models, most methods also need to train with the same iterative steps on the complete training dataset, in which one training step in QAT actually takes more GPU memory and longer running time than those of the regular floating-point. On the contrast, post-training quantization (PTQ) only requires a few unlabeled calibration images without train- ing, which enables fast deployment on various devices within minutes. Nevertheless, different from the image clas- sification, super resolution requires accurate prediction for each pixel of the output images, which is much sensitive to low-bit compression for feature maps. Figure 2 shows the original floating-point activations of different layers and samples, we observe three properties of their distributions that are much unfriendly to quantization: (1) Long-tailed : 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. 5856 𝑢 𝑢 𝑙𝐻𝑙<𝐻𝑢 𝑙=𝑙+∆𝐻𝑙>𝐻𝑢 𝑢=𝑢−∆ 𝑙 𝑢 ∆ሼ𝑙∆ሼ∆ሼ…… 𝑙𝑢∆ሼFull-precision model Step 1: Density -based Dual ClippingFull-precision model with ሼ𝑙𝑎,𝑢𝑎,𝑙𝑤,𝑢𝑤}𝐾 Quantized model with ሼ𝑙𝑎∗,𝑢𝑎,∗,𝑙𝑤∗,𝑢𝑎∗}𝐾 Step 2: Pixel -aware CalibrationFull-precision neuron Quantized neuronpixel transfer loss pixel transfer lossonly optimize ሼ𝑙𝑎,𝑢𝑎}𝐾 L1 loss L1 lossonly optimize ሼ𝑙𝑤,𝑢w}𝐾Figure 1. The overview of the proposed post-training quantization framework for image super resolution. the distribution shows to be dense in the middle yet sparse in the tails, which means most of values lie in a small range, while only a few outliers have larger amplitude; (2) Asym- metric : the density on the two tails of the distribution is asymmetric, the skewness differ for different layers; (3) Highly-dynamic : the activation range varies, or even by twice, for different input samples. Therefore, the existing PTQ methods which are designed for image classification can not be transferred to the SR task directly. In this paper, we propose a coarse-to-fine method to get the accurate quantized SR model with post-training quan- tization. We first introduce the density-based dual clipping (DBDC) to cut off most of the outliers for narrowing the distribution to a valid range. Different from previous meth- ods [25, 40], the amplitudes of lower and upper clip are not same and the clipping position is depend on the den- sity of two tails. The clipping scheme is employed itera- tively to eliminate the long-tail distribution. The asymmet- ric quantizer with adjustable lower and upper clip values is adopted to solve the asymmetric distribution in SR models. And then we further propose a novel pixel-aware calibration (PaC) to help the quantized network fit the highly dynamic activations for different samples. The PaC leverages fea- ture maps of the full-precision model to supervise those of the quantized model. To stabilize the finetune process, we only update the quantization parameters instead of the orig- inal weights. The whole quantization process of our method can be finished within minutes with a few unlabeled images. The contributions of this paper are summarized as follow: (1) We present a detailed analysis to demonstrate the challenge of post-training quantization on image super reso- lution, indicating that the performance degradation of quan- tized SR model suffers from the long-tailed, asymmetric and highly-dynamic distribution of feature maps. (2) We introduce a coarse-to-fine quantization method to accommodate above problems. With the density-based dual clipping and the pixel-aware calibration, the proposed method is able to conduct accurate quantization with only afew unlabeled calibration images. To the best of our knowl- edge, we are the first to optimize the post-training quantiza- tion for image super resolution task. (3) Extensive experiments on various benchmark models and datasets demonstrate that our method significantly out- performs the existing PTQ methods, and is able to achieve comparable performance with the QAT in some setting. Further, our method can speed up the convergence and bring up the performance when combined with QAT methods.

Song_Advancing_Visual_Grounding_With_Scene_Knowledge_Benchmark_and_Method_CVPR_2023

Abstract Visual grounding (VG) aims to establish fine-grained alignment between vision and language. Ideally, it can be a testbed for vision-and-language models to evaluate their understanding of the images and texts and their reason- ing abilities over their joint space. However, most existing VG datasets are constructed using simple description texts, which do not require sufficient reasoning over the images and texts. This has been demonstrated in a recent study [27], where a simple LSTM-based text encoder without pretrain- ing can achieve state-of-the-art performance on mainstream VG datasets. Therefore, in this paper, we propose a novel benchmark of Scene Knowledge-guided Visual Grounding (SK-VG), where the image content and referring expressions are not sufficient to ground the target objects, forcing the models to have a reasoning ability on the long-form scene knowledge. To perform this task, we propose two approaches to accept the triple-type input, where the former embeds knowledge into the image features before the image-query interaction; the latter leverages linguistic structure to assist in computing the image-text matching. We conduct exten- sive experiments to analyze the above methods and show that the proposed approaches achieve promising results but still leave room for improvement, including performance and interpretability. The dataset and code are available at https://github.com/zhjohnchan/SK-VG .

1. Introduction Visual grounding (VG), aiming to locate an object re- ferred to by a description phrase/text in an image, has emerged as a prominent attractive research direction. It can be applied to various tasks (e.g., visual question answering [4,13,38,51] and vision-and-language navigation [1,11,35]) and also be treated as a proxy to evaluate machines for *Equal contribution †Corresponding author ThemanJakeiscelebratingChristmaswithhisfriends.Jakeputshislefthandontheglassesandholdsawineglassinhisrighthand.HisfriendAlanissittingonhisrightinasuit.Carol,Jake'sanotherfriend,standsonJake'sleftwithawineglassinhislefthand,grinningwithwhiteteeth.Q1:TheblackglassesQ2:Jake’swineglass (Traditional)(SK-VG(Ours)) ImageQuerySceneKnowledge Figure 1. An example from the proposed SK-VG dataset for scene knowledge-guided visual grounding. The task requires a model to reason over the (image, scene knowledge, query) triple to locate the target object referred to by the query. open-ended scene recognition. Typically, VG requires mod- els to reason over vision and language and build connec- tions through single-modal understanding and cross-modal matching. Yet, current VG benchmarks (e.g., RefCOCO [47], RefCOCO+ [47], RefCOCOg [29], ReferItGame [17], and CLEVR-Ref+ [23]) can not serve as a good test bed to eval- uate the reasoning ability since they only focus on simple vision-language alignment. In addition to the simple nature of constructed referring expressions, this can be reflected in the recent state-of-the-art study [27], where they showed that VG models are less affected by language modeling through extensive empirical analyses . In this paper, we believe that the intrinsic difficulty of VG lies in the difference between perceptual representations of images and cognitive representations of texts. Specifically, vi- sual features are obtained through perceptual learning, which 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. 15039 ImageCategoryModel(V)BoundingBox(many) (i)GroundingCategories(e.g.,PASCAL VOC 2007,MS-COCO)ImagePhraseModel(V+L)BoundingBox(many) (ii)GroundingPhrases(e.g.,Flickr30k Entities)ImageLinguistic ExpressionModel(V+L)BoundingBox(exactone) (iii)GroundingExpressions(e.g.,RefCOCO,CLEVR-Ref+,Cops-Ref)ImageLinguistic Expression+SceneKnowledgeModel(V+L)BoundingBox(exactone) (iv)GroundingExpressionsGivenSceneKnowledge(Ours)Figure 2. Illustrations of four categories of grounding tasks, including categories, phrases, linguistic expressions, and linguistic expres- sion+scene knowledge. The height of the input green and blue rectangles denotes its relative information. only maps visual appearances in images to semantic con- cepts. However, open-ended queries might require VG mod- els to understand the whole scene knowledge before perform- ing reasoning to locate the target object. As shown in Figure 1, the perceptual features can encode the information about “a wine glass ”, but it would struggle to locate “ Jake’s wine glass ” without the scene knowledge about “ who is Jake? ”. This is a challenging task owing to two facts: (i) From the dataset perspective, there are no relevant benchmarks for the VG researchers to evaluate their models; (ii) From the model/algorithm perspective, it is not easy to design models to perform reasoning among images, scene knowledge, and open-ended querying texts. Therefore, we propose to break this limitation of cur- rent VG research and construct a new benchmark requir- ingVGto perform reasoning over Scene Knowledge (i.e., text-based stories). The benchmark named SK-VG contains ∼40,000 referring expressions and 8,000 scene stories from 4,000 images, where each image contains 2 scene stories with 5 referring expressions for each story. Moreover, to evaluate the difficulty levels of queries, we curate the test set by splitting the samples into easy/medium/hard cate- gories to provide a detailed evaluation of the vision-language models. Under this new setting, we develop a one-stage ap- proach (i.e., Knowledge-embedded Vision-Language Inter- action (KeViLI)) and a two-stage approach (i.e., Linguistic- enhanced Vision-Language Matching (LeViLM)). In KeViLI, the scene knowledge is firstly embedded into the image fea- tures, and then the interaction between the image and the query is performed; In LeViLM, the image features and the text features are first extracted, and then the match- ing between the (image) regions and the (text) entities are computed, assisted by the structured linguistic information. Through extensive experiments, we show that the proposed approaches can achieve the best performance but still leave room for improvement, especially in the hard split. It chal- lenges the models from three perspectives: First, it is anopen-ended grounding task; Second, the scene stories are long narratives consisting of multiple sentences; Third, it might require the multi-hop reasoning ability of the models. In summary, the contributions of this paper are three-fold: •We introduce a challenging task that requires VG mod- els to reason over (image, scene knowledge, query) triples and build a new dataset named SK-VG on top of real images through manual annotations. •We propose two approaches to enhance the reasoning in SK-VG, i.e., one one-stage approach KeViLI and one two-stage approach LeViLM. •Extensive experiments demonstrate the effectiveness of the proposed approaches. Further analyses and discus- sions could be a good starting point for future study in the vision-and-language field.

Tu_Learning_From_Noisy_Labels_With_Decoupled_Meta_Label_Purifier_CVPR_2023

Abstract Training deep neural networks (DNN) with noisy labels is challenging since DNN can easily memorize inaccurate labels, leading to poor generalization ability. Recently, the meta-learning based label correction strategy is widely adopted to tackle this problem via identifying and correcting potential noisy labels with the help of a small set of clean validation data. Although training with purified labels can effectively improve performance, solving the meta-learning problem inevitably involves a nested loop of bi-level opti- mization between model weights and hyper-parameters (i.e., label distribution). As compromise, previous methods re- sort to a coupled learning process with alternating update. In this paper, we empirically find such simultaneous opti- mization over both model weights and label distribution can not achieve an optimal routine, consequently limiting the representation ability of backbone and accuracy of cor- rected labels. From this observation, a novel multi-stage label purifier named DMLP is proposed. DMLP decou- ples the label correction process into label-free representa- tion learning and a simple meta label purifier, In this way, DMLP can focus on extracting discriminative feature and label correction in two distinctive stages. DMLP is a plug- and-play label purifier, the purified labels can be directly reused in naive end-to-end network retraining or other ro- bust learning methods, where state-of-the-art results are obtained on several synthetic and real-world noisy datasets, especially under high noise levels. Code is available at https://github.com/yuanpengtu/DMLP .

1. Introduction Deep learning has achieved significant progress on vari- ous recognition tasks. The key to its success is the availabil- * Equal contribution. †Corresponding author. Figure 1. (a) Traditional coupled alternating update to solve meta label purification problem, and (b) the proposed DMLP method that decouples the label purification process into representation learning and a simple non-nested meta label purifier. ity of large-scale datasets with reliable annotations. Collect- ing such datasets, however, is time-consuming and expensive. Easy ways to obtain labeled data, such as web crawling [31], inevitably yield samples with noisy labels, which is not appropriate to be directly utilized to train DNN since these complex models are vulnerable to memorize noisy labels [2]. Towards this problem, numerous Learning with Noisy La- bel (LNL) approaches were proposed. Classical LNL meth- ods focus on identifying the noisy samples and reducing their effect on parameter updates by abandoning [12] or assigning smaller importance. However, when it comes to extremely noisy and complex scenarios, such scheme struggles since there is no sufficient clean data to train a discriminative clas- sifier. Therefore, label correction approaches are proposed to augment clean training samples by revising noisy labels to underlying correct ones. Among them, meta-learning based approaches [9, 16, 25] achieve state-of-the-art performance via resorting to a small clean validation set and taking noisy labels as hyper-parameters, which provides sound guidance 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. 19934 toward underlying label distribution of clean samples. How- ever, such meta purification inevitably involves a nested bi-level optimization problem on both model weight and hyper-parameters (shown as Fig. 1 (a)), which is computa- tionally infeasible. As a compromise, the alternating update between model weights and hyper-parameters is adopted to optimize the objective [9, 16, 25], leading to a coupled solution for representation learning and label purification. Empirical observation. Intuitively, alternate optimiza- tion over a large search space (model weight and hyper- parameters) may lead to sub-optimal solutions. To investi- gate how such approximation affects results in robust learn- ing, we conduct empirical analysis on CIFAR-10 [14] with recent label purification methods MLC [41] and MSLC [9], which consist of a deep model and a meta label correction network, and make observation as Fig. 2. •Coupled optimization hinders quality of corrected la- bels. We first compare the Coupled meta corrector MLC with its extremely Decoupled variant where the model weights are first optimized for 70epochs with noisy labels and get fixed, then labels are purified with the guidance of validation set. We adopt the accuracy of corrected label to measure the performance of purification. From Fig. 2 (a), we can clearly observe that compared with Decoupled counterpart, joint optimization yields inferior correction per- formance, and these miscorrection will reversely affect the representation learning in coupled optimization. •Coupled optimization hinders representation ability. We investigate the representation quality by evaluating the linear prob accuracy [6] of extracted feature in Fig. 2 (b). We find the representation quality of Coupled training is much worse at the beginning, which leads to slow and un- stable representation learning in the later stage. To further investigate the effect on representation learning, we also resort to a well pretrained backbone with self-supervised learning [5] as initialization, recent research [40] shows pre- trained representation is substantially helpful for LNL frame- work. However, we find this conclusion does not strictly hold for coupled meta label correctors. As shown in Fig. 2 (c), by comparing the classification accuracy from classi- fier of MLC/MSLC, we observe the pretrained model only brings marginal improvement if model weights is still cou- pled with hyper-parameters. In contrast, when the weight of backbone is fixed and decoupled from the label purification and classifier, the improvement becomes more significant. Decoupled Meta Purification. From the observation above, we find the decoupling between model weights and hyperparameters of meta correctors is essential to label ac- curacy and final results. Therefore, in this paper, we aim at detaching the meta label purification from representation learning and designing a simple meta label purifier which is more friendly to optimization of label distribution prob- lem than existing complex meta networks [9, 41]. Hencewe propose a general multi-stage label correction strategy, named Decoupled Meta Label Purifier (DMLP). The core of DMLP is a meta-learning based label purifier, however, to avoid solving the bi-level optimization with a coupled solution, DMLP decouples this process into self-supervised representation learning and a linear meta-learner to fit un- derlying correct label distribution (illustrated as Fig. 1 (b)), thus simplifies the label purification stage as a single-level optimization problem. The simple meta-learner is carefully designed with two mutually reinforcing correcting processes, named intrinsic primary correction (IPC) and extrinsic aux- iliary correction (EAC) respectively. IPC plays the role of purifying labels in a global sense at a steady pace, while EAC targets at accelerating the purification process via looking ahead (i.e., training with) the updated labels from IPC. The two processes can enhance the ability of each other and form a positive loop of label correction. Our DMLP framework is flexible for application, the purified labels can either be directly applied for naive end-to-end network retraining, or exploited to boost the performance of existing LNL frame- works. Extensive experiments conducted on mainstream benchmarks, including synthetic (noisy versions of CIFAR) and real-world (Clothing1M) datasets, demonstrate the supe- riority of DMLP. In a nutshell, the key contributions of this paper include: •We analyze the necessity of decoupled optimization for label correction in robust learning, based on which we propose DMLP, a flexible and novel multi-stage label purifier that solves bi-level meta-learning problem with a decoupled manner, which consists of representation learning and non- nested meta label purification; •In DMLP, a novel non-nested meta label purifier equipped with two correctors, IPC and EAC is proposed. IPC is a global and steady corrector, while EAC accelerates the correction process via training with the updated labels from IPC. The two processes form a positive training loop to learn more accurate label distribution; •Deep models trained with purified labels from DMLP achieve state-of-the-art results on several synthetic and real- world noisy datasets across various types and levels of label noise, especially under high noise levels. Extensive ablation studies are provided to verify the effectiveness.

Stathopoulos_Learning_Articulated_Shape_With_Keypoint_Pseudo-Labels_From_Web_Images_CVPR_2023

Abstract This paper shows that it is possible to learn models for monocular 3D reconstruction of articulated objects ( e.g. horses, cows, sheep), using as few as 50-150 images la- beled with 2D keypoints. Our proposed approach involves training category-specific keypoint estimators, generating 2D keypoint pseudo-labels on unlabeled web images, and using both the labeled and self-labeled sets to train 3D re- construction models. It is based on two key insights: (1) 2D keypoint estimation networks trained on as few as 50- 150 images of a given object category generalize well and generate reliable pseudo-labels; (2) a data selection mech- anism can automatically create a “curated” subset of the unlabeled web images that can be used for training – we evaluate four data selection methods. Coupling these two insights enables us to train models that effectively utilize web images, resulting in improved 3D reconstruction per- formance for several articulated object categories beyond the fully-supervised baseline. Our approach can quickly bootstrap a model and requires only a few images labeled with 2D keypoints. This requirement can be easily satisfied for any new object category. To showcase the practicality of our approach for predicting the 3D shape of arbitrary object categories, we annotate 2D keypoints on 250 giraffe and bear images from COCO in just 2.5 hours per category.

1. Introduction Predicting the 3D shape of an articulated object from a single image is a challenging task due to its under- constrained nature. Various successful approaches [14, 19] have been developed for inferring the 3D shape of humans. These approaches rely on strong supervision from 3D joint locations acquired using motion capture systems. Similar breakthroughs for other categories of articulated objects, such as animals, remain elusive. This is primarily due to the scarcity of appropriate training data. Some works (such as CMR [15]) learn to predict 3D shapes using only 2D labels for supervision. However, for most object categories even train(a) Train KeypointEstimator Labeled set 𝒮inference(b) Generate Pseudo-labels (PLs)𝒰w/ PLs(c) Data Selection𝒰′(d) Train 3D Shape Predictor Unlabeled set 𝒰 𝒰w/ PLs 𝒰′𝒮∪{}trainℎ$ℎ$𝑓%Figure 1. Overview of the proposed framework . It includes: (a) training a category-specific keypoint estimator with a limited labeled set S, (b) generating keypoints pseudo-labels on web im- ages, (c) automatic curation of web images to create a subset U′, and (d) training a model for 3D shape prediction with images from SandU′. 2D labels are limited or non-existent. We ask: how can we learn models that predict the 3D shape of articulated ob- jects in-the-wild when limited or no annotated images are available for a given object category? In this paper we propose an approach that requires as few as 50-150 images labeled with 2D keypoints. This labeled set can be easily and quickly created for any object category. Our proposed approach is illustrated in Figure 1 and sum- marized as follows: (a) train a category-specific keypoint estimation network using a small set Sof images labeled with 2D keypoints; (b) generate 2D keypoint pseudo-labels on a large unlabeled set Uconsisting of automatically ac- quired web images; (c) automatically curate Uby creating a subset of images and pseudo-labels U′according to a se- lection criterion; (d) train a model for 3D shape prediction with data from both SandU′. A key insight is that current 2D keypoint estimators [30, 34, 38] are accurate enough to create robust 2D key- point detections on unlabeled data, even when trained with a limited number of images. Another insight is that images fromUincrease the variability of several factors, such as This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13092 UUφ,θS US Uφ,ψ /bardbl−/bardbl2 hφ hφ fθ fθ(c) Cross-Model Consistency(d) Cross-Model & Cross-Modality Consistency(a) CF-CM(b) CF-CM2 gψ/bardbl()−/bardbl2Figure 2. Given a small set Sof images labeled with 2D keypoints, we train a 2D keypoint estimation netowrk hϕand generate keypoint pseudo-labels on web images (set U). We select a subset of Uto train a 3D shape predictor fθ. Two methods for data selection can be seen here: (a) CF-CM : an auxiliary 2D keypoint estimator gψgenerates predictions on Uand images with the smallest discrepancy between the keypoint estimates of hϕandgψare selected (criterion (c)); (b) CF-CM2:fθis trained with samples from Sand generates predictions onU. Images with the smallest discrepancy between the keypoint estimates of hϕandfθare selected (criterion (d)) to retrain fθ. camera viewpoints, articulations and image backgrounds, that are important for training generalizable models for 3D shape prediction. However, the automatically acquired web images contain a high proportion of low-quality images with wrong or heavy truncated objects. Naively using all web images and pseudo-labels during training leads to de- graded performance as can be seen in our experiments in Section 4.2. While successful pseudo-label (PL) selection techniques [4, 5, 23, 29, 40] exist for various tasks, they do not address the challenges in our setting. These works investigate PL selection when the unlabeled images come from curated datasets ( e.g. CIFAR-10 [20], Cityscapes [6]), while in our setting they come from the web. In addition, they eventually use all unlabeled images during training while in our case most of the web images should be dis- carded. To effectively utilize images from the web we inves- tigate four criteria to automatically create a “curated” sub- set that includes images with high-quality pseudo-labels. These contain a confidence-based criterion as well as three consistency-based ones (see Figure 2 for two examples). Through extensive experiments on five different articu- lated object categories (horse, cow, sheep, giraffe, bear) and three public datasets, we demonstrate that training with the proposed data selection approaches leads in considerably better 3D reconstructions compared to the fully-supervised baseline. Using all pseudo-labels leads to degraded perfor- mance. We analyze the performance of the data selection methods used and conclude that consistency-based selec- tion criteria are more effective in our setting. Finally, we conduct experiments with varying number of images in the labeled set S. We show that even with only 50 annotated instances and images from web, we can train models that lead to better 3D reconstructions than the fully-supervised models trained with more labels.

Tang_Visual_Recognition_by_Request_CVPR_2023

Abstract Humans have the ability of recognizing visual semantics in an unlimited granularity, but existing visual recognition algorithms cannot achieve this goal. In this paper, we estab- lish a new paradigm named visual recognition by request (ViRReq1) to bridge the gap. The key lies in decompos- ing visual recognition into atomic tasks named requests and leveraging a knowledge base, a hierarchical and text-based dictionary, to assist task definition. ViRReq allows for (i) learning complicated whole-part hierarchies from highly incomplete annotations and (ii) inserting new concepts with minimal efforts. We also establish a solid baseline by in- tegrating language-driven recognition into recent seman- tic and instance segmentation methods, and demonstrate its flexible recognition ability on CPP and ADE20K, two datasets with hierarchical whole-part annotations.

1. Introduction Visual recognition is one of the fundamental problems in computer vision. In the past decade, visual recognition algorithms have been largely advanced with the availability of large-scale datasets and deep neural networks [ 8,13,18]. Typical examples include the ability of recognizing 10,000s of object classes [ 6], segmenting objects into parts or even parts of parts [ 43], using natural language to refer to open- world semantic concepts [ 31],etc. Despite the increasing recognition accuracy in standard benchmarks, we raise a critical yet mostly uncovered issue, namely, the unlimited granularity of visual recognition. As shown in Figure 1, given an image, humans have the ability of recognizing arbitrarily fine-grained contents from it, but existing visual recognition algorithms cannot achieve the same goal. Superficially, this is caused by limited anno- tation budgets so that few training data is available for fine- 1We recommend the readers to pronounce ViRReq as /’virik/ . eye, mouth, …~10% annotated head, torso, …~40% annotated person, bag, …~80% annotated original imageregions & instancesparts & parts of parts Person #1: head, torso, arms, legs, eyes, mouth, …uncertain boundary between head/armand torsoPerson #2:head, torso, arms, legs, eyes, mouth, …Person #3:head, torso, arms, legs, eyes, mouth, …Person ★:instance was not annotatedOther objects: bag, shoes, necklace, …overlaying #1#2#3uncertainboundaryFigure 1. An illustration of unlimited granularity in visual recog- nition. Top: an example image from ADE20K [ 43] with instance- level and part-level annotations. Middle : more and more incom- plete annotations of instances, parts, and parts of parts. Bottom : as granularity goes finer, higher uncertainty occurs in recogniz- ing the boundary (left) and semantic class (right) of objects and/or parts – here, green ,blue andredtexts indicate labeled ,unlabeled (but defined) , and unlabeled (and undefined) objects, respectively. grained and/or long-tailed concepts, but we point out that a more essential reason lies in the conflict between granular- ity and recognition certainty – as shown in Figure 1,when annotation granularity goes finer, annotation certainty will inevitably be lower . This motivates us that the granular- ity of visual recognition shall be variable across instances and scenarios. For this purpose, we propose to interpret semantic annotations into smaller units ( i.e., requests) and assume that recognition is performed only when it is asked, so that one can freely adjust the granularity according to object size, importance, clearness, etc. In this paper, we establish a new paradigm named visual recognition by request (ViRReq ). We only consider the segmentation task in this paper because it best fits the need of unlimited granularity in the spatial domain. Compared 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

Song_Learning_With_Fantasy_Semantic-Aware_Virtual_Contrastive_Constraint_for_Few-Shot_Class-Incremental_CVPR_2023

Abstract Few-shot class-incremental learning (FSCIL) aims at learning to classify new classes continually from limited samples without forgetting the old classes. The mainstream framework tackling FSCIL is first to adopt the cross-entropy (CE) loss for training at the base session, then freeze the feature extractor to adapt to new classes. However, in this work, we find that the CE loss is not ideal for the base ses- sion training as it suffers poor class separation in terms of representations, which further degrades generalization to novel classes. One tempting method to mitigate this prob- lem is to apply an additional na ¨ıve supervised contrastive learning (SCL) in the base session. Unfortunately, we find that although SCL can create a slightly better representa- tion separation among different base classes, it still strug- gles to separate base classes and new classes. Inspired by the observations made, we propose Semantic-Aware Vir- tual Contrastive model (SAVC), a novel method that facili- tates separation between new classes and base classes by introducing virtual classes to SCL. These virtual classes, which are generated via pre-defined transformations, not only act as placeholders for unseen classes in the repre- sentation space, but also provide diverse semantic infor- mation. By learning to recognize and contrast in the fan- tasy space fostered by virtual classes, our SAVC signifi- cantly boosts base class separation and novel class gen- eralization, achieving new state-of-the-art performance on the three widely-used FSCIL benchmark datasets. Code is available at: https://github.com/zysong0113/SAVC.

1. Introduction Like humans, the modern artificial intelligence models are expected to be equipped with the ability of learning continually. For instance, a face recognition system needs †: authors contributed equally.∗: corresponding author. Base Session Few -Shot Incremental SessionCE CE+SCL SAVC(a) Comparison of the embedding spaces trained with different methods. AP1 P2P1A N2N1 N1 P2+ +-Different Classes N3 N4Fantasy ClassesSame Classes- - Al Al-Local Crop Global Crop Negative Pairs … Positive Pairs…Fantasy Samples Update … Positive Selection Negative Selection Original Feature Queue-N2Semantic-Aware Fantasy N3 N4 (b) Illustration of our Semantic-Aware Virtual Contrastive model (SA VC). Figure 1. The motivation of the proposed approach. Under the incremental-frozen framework in FSCIL, our SA VC learns to rec- ognize and contrast in the fantasy space, which leads to better base class separation and novel class generalization than CE and SCL. to continually learn to recognize new faces without forget- ting those has already kept in mind. In this case, Class- Incremental Learning (CIL) [19, 25, 30, 37, 50] draws atten- tion extensively, that sufficient and disjoint data are com- ing session-by-session. However, it is unrealistic to always rely on accessing of multitudinous data. Taking the case of face recognition system again, whether it can quickly learn a new concept with just a few images, is an important cri- terion for evaluating its performance. As a result, design- ing effective and efficient algorithms to resolve Few-Shot Class-Incremental Learning (FSCIL) problem has drawn in- creasing attention [6, 8, 12, 40, 44, 51, 55]. Unlike CIL, only a few data are provided for FSCIL at incremental session 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. 24183 while sufficient samples are only visible at the base session. The core challenge of CIL is to strike a balance of stabil- ity and plasticity, i.e., suppressing forgetting on old knowl- edge while adapting smoothly to new knowledge. Previ- ous works on FSCIL have demonstrated that an incremen- tally trained model is prone to overfit on limited new data and suffer catastrophic forgetting [8, 12, 44, 54]. Thus the mainstream FSCIL framework [40, 51, 58] is to first use the cross-entropy (CE) loss for training in the base session, then freeze the backbone to adapt to new classes. Given such baseline framework, the main problem that needs to be taken into consideration is: What makes a good base- session-model that can generalize well to new classes with limited data? Intuitively, if all representations of base classes concen- trate around their own clustering centers ( i.e., prototypes), and all prototypes are far away from each other, then novel classes should be easily incorporated into the current space without overlapping. Therefore, we conjecture that base class separation facilitates novel class generalization. How- ever, the current widely adopted CE loss in FSCIL cannot well separate the class margins, which causes poor gen- eralization to new classes. This motivates our attempt at methods with better clustering effects, such as supervised contrastive learning (SCL) [21]. We conduct experiments to validate our conjecture by training models optimized by only CE loss, and both CE and SCL. We empirically find that CE leads to lower base class separation and classifica- tion accuracy than SCL. Although SCL indeed slightly im- proves the base class separation with higher accuracy, we find it implicitly decreases the inter-class distance. There- fore, the two methods both leave inadequate room for future updates, which leads to the inevitable overlapping between novel classes and old classes in the representation space (as shown in the left and middle columns of Fig. 1a). The unsatisfactory performance of SCL implies that the limited semantic information base session provided may need our fantasy remedy. Although we have no access to future classes in the base session, we can create many vir- tual classes with various semantics filling the unallocated representation space, which not only holds enough space for future classes, but also enables the learning of diverse semantics for better inference and generalization. To this end, we propose a Semantic-Aware Virtual Con- trastive (SA VC) model to improve the FSCIL by boosting base class separation (see Fig. 1b). Our fantasy space is con- structed with pre-defined transformations [11, 15, 35, 52], which can be regarded as the semantic extension of the orig- inal space in finer grains. The type of transformations rep- resents the richness of our imagination, i.e., how compact we want the embedding space to be and how fine the se- mantic grains we want to acquire. Learning to contrast and recognize in the fantasy space not only brings a better classseparation effect, but also provides richer semantic infor- mation to improve the accuracy. In contrast to the CE base- line and SCL, our SA VC can simultaneously decrease intra- class distance and increase inter-class distance, which leads to better base class separation and novel class generalization (as given in the right of Fig. 1a). Extensive experiments on various benchmark datasets further verify the superiority of our method. Our main contributions are summarized in three folds: 1) We empirically discover that it is crucial to boost class separation degree in base session for the incremental-frozen framework in FSCIL, which helps fast generalization for novel classes with only a few samples. 2) We propose a novel Semantic-Aware Virtual Con- trastive (SA VC) framework, enhancing the FSCIL learner by learning to recognize and contrast in the fantasy space, which preserves enough place for novel generalization and enables a multi-semantic aggregated inference effect. 3) Extensive experiments on three FSCIL bench- marks, i.e., CIFAR100, miniImageNet and CUB200, demonstrate that our SA VC outperforms all approaches by a large margin and achieves the state-of-the-art performance.

Streicher_BASiS_Batch_Aligned_Spectral_Embedding_Space_CVPR_2023

Abstract Graph is a highly generic and diverse representation, suitable for almost any data processing problem. Spec- tral graph theory has been shown to provide powerful algo- rithms, backed by solid linear algebra theory. It thus can be extremely instrumental to design deep network build- ing blocks with spectral graph characteristics. For in- stance, such a network allows the design of optimal graphs for certain tasks or obtaining a canonical orthogonal low- dimensional embedding of the data. Recent attempts to solve this problem were based on minimizing Rayleigh- quotient type losses. We propose a different approach of directly learning the graph’s eigensapce. A severe prob- lem of the direct approach, applied in batch-learning, is the inconsistent mapping of features to eigenspace coordinates in different batches. We analyze the degrees of freedom of learning this task using batches and propose a stable align- ment mechanism that can work both with batch changes and with graph-metric changes. We show that our learnt spec- tral embedding is better in terms of NMI, ACC, Grassman distnace, orthogonality and classification accuracy, com- pared to SOTA. In addition, the learning is more stable.

1. Introduction Representing information by using graphs and analyz- ing their spectral properties has been shown to be an effec- tive classical solution in a wide range of problems including clustering [8,21, 32], classification [13], segmentation [26], dimensionality reduction [5, 10, 23] and more. In this set- ting, data is represented by nodes of a graph, which are em- bedded into the eigenspace of the graph-Laplacian, a canon- ical linear operator measuring local smoothness. Incorporating analytic data structures and methods within a deep learning framework has many advantages. It yields better transparency and understanding of the net- work, allows the use of classical ideas, which were thor- oughly investigated and can lead to the design of new ar- chitectures, grounded in solid theory. Spectral graph algo-rithms, however, are hard to incorporate directly in neural- networks since they require eigenvalue computations which cannot be integrated in back-propagation training algo- rithms. Another major drawback of spectral graph tools is their low scalability. It is not feasible to hold a large graph containing millions of nodes and to compute its graph- Laplacian eigenvectors. Moreover, updating the graph with additional nodes is combersome and one usually resorts to graph-interpolation techniques, referred to as Out Of Sam- ple Extension (OOSE) methods. An approach to solve the above problems using deep neural networks (DNNs), firstly suggested in [24] and re- cently also in [9], is to train a network that approximates the eigenspace by minimizing Rayleigh quotient type losses. The core idea is that the Rayleigh quotient of a sum of nvec- tors is minimized by the neigenvectors with the correspond- ingnsmallest eigenvalues. As a result, given the features of a data instance (node) as input, these networks generate the respective coordinate in the spectral embedding space. This space should be equivalent in some sense to the ana- lytically calculated graph-Laplacian eigenvector space. A common way to measure the equivalence of these spaces is using the Grassman distance. Unfortunately, applying this indirect approach does not guarantee convergence to the de- sired eigenspace and therefore the captured might not be faithful. An alternative approach, suggested in [18] for computing the diffusion map embedding, is a direct supervised method. The idea is to compute the embedding analytically, use it as ground-truth and train the network to map features to eigenspace coordinates in a supervised manner. In order to compute the ground truth embedding, the authors used the entire training set. This operation is very demanding com- putationally in terms of both memory and time and is not scalable when the training set is very large. Our proposed method is to learn directly the eigenspace in batches. We treat each batch as sampling of the full graph and learn the eigenvector values in a supervised manner. A major problem of this kind of scheme is the i

Speth_Non-Contrastive_Unsupervised_Learning_of_Physiological_Signals_From_Video_CVPR_2023

Abstract Subtle periodic signals such as blood volume pulse and respiration can be extracted from RGB video, enabling non- contact health monitoring at low cost. Advancements in remote pulse estimation – or remote photoplethysmogra- phy (rPPG) – are currently driven by deep learning solu- tions. However, modern approaches are trained and evalu- ated on benchmark datasets with ground truth from contact- PPG sensors. We present the first non-contrastive unsuper- vised learning framework for signal regression to mitigate the need for labelled video data. With minimal assumptions of periodicity and finite bandwidth, our approach discov- ers the blood volume pulse directly from unlabelled videos. We find that encouraging sparse power spectra within nor- mal physiological bandlimits and variance over batches of power spectra is sufficient for learning visual features of periodic signals. We perform the first experiments utilizing unlabelled video data not specifically created for rPPG to train robust pulse rate estimators. Given the limited induc- tive biases and impressive empirical results, the approach is theoretically capable of discovering other periodic signals from video, enabling multiple physiological measurements without the need for ground truth signals.

1. Introduction Camera-based vitals estimation is a rapidly growing field enabling non-contact health monitoring in a variety of set- tings [23]. Although many of the signals avoid detection from the human eye, video data in the visible and infrared ranges contain subtle intensity changes caused by physio- logical oscillations such as blood volume and respiration. Significant remote photoplethysmography (rPPG) research for estimating the cardiac pulse has leveraged supervised deep learning for robust signal extraction [7, 18, 30, 38, 51, 52]. While the number of successful approaches has rapidly increased, the size of benchmark video datasets with simul- taneous vitals recordings has remained relatively stagnant. Robust deep learning-based systems for deployment re- quire training on larger volumes of video data with di-verse skin tones, lighting, camera sensors, and movement. However, collecting simultaneous video and physiologi- cal ground truth with contact-PPG or electrocardiograms (ECG) is challenging for several reasons. First, many hours of high quality videos is an unwieldy volume of data. Sec- ond, recording a diverse subject population in conditions representative of real-world activities is difficult to conduct in the lab setting. Finally, synchronizing contact measure- ments with video is technically challenging, and even con- tact measurements used for ground truth contain noise. Fortunately, recent works find that contrastive unsuper- vised learning for rPPG is a promising solution to the data scarcity problem [13, 42, 45, 50]. We extend this research line into non-contrastive unsupervised learning to discover periodic signals in video data. With end-to-end unsuper- vised learning, collecting more representative training data to learn powerful visual features is much simpler, since only video is required without associated medical information. In this work, we show that non-contrastive unsupervised learning is especially simple when regressing rPPG signals. We find weak assumptions of periodicity are sufficient for learning the minuscule visual features corresponding to the blood volume pulse from unlabelled face videos. The loss functions can be computed in the frequency domain over batches without the need for pairwise or triplet compar- isons. Figure 1 compares the proposed approach with su- pervised and contrastive unsupervised learning approaches. This work creates opportunities for scaling deep learning models for camera-based vitals and estimating periodic or quasi-periodic signals from unlabelled data beyond rPPG. Ournovel contributions are: 1. A general framework for physiological signal estima- tion via non-contrastive unsupervised learning (SiNC) by leveraging periodic signal priors. 2. The first non-contrastive unsupervised learning method for camera-based vitals measurement. 3. The first experiments and results of training with non- rPPG-specific video data without ground truth vitals. Source code to replicate this work is available at https: //github.com/CVRL/SiNC-rPPG . 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. 14464 𝑥"!𝑥""𝑥𝜙" ~ Φ 𝜙! ~ Φ Invariant ViewEquivariant View𝑦"𝑦!Waveform PredictionSpectral Prediction𝐹𝐹𝑇(+)𝐹𝐹𝑇(+)𝐹"𝐹!𝑥""#𝑥"!#𝜏" ~ Τ𝜏! ~ ΤSupervisedlearning framework𝜙 ~ Φ 𝜏 ~ Τ𝐹𝐹𝑇(+)𝑥"𝑥𝑦𝑥"#𝐿𝐹=𝐿$+𝐿% +𝐿& 𝐹𝑓(+)𝑓(+)𝑓(+)𝑥"𝑥𝜙 ~ Φ 𝑦𝑥"#𝜏 ~ Τ𝑦'(𝑓(+) t 𝐿(𝐹",𝐹!)𝐿(𝑦,𝑦'() bandwidth (𝐿$) sparsity (𝐿%)<=>batch-wise variance (𝐿&)Unsupervised contrastive learningframeworkUnsupervised non-contrastive learningframework(proposed)Figure 1. Overview of the SiNC framework for rPPG compared with traditional supervised and unsupervised learning. Supervised and contrastive losses use distance metrics to the ground truth or other samples. Our framework applies the loss directly to the prediction by shaping the frequency spectrum, and encouraging variance over a batch of inputs. Power outside of the bandlimits is penalized to learn invariances to irrelevant frequencies. Power within the bandlimits is encouraged to be sparsely distributed near the peak frequency.

Tian_Manipulating_Transfer_Learning_for_Property_Inference_CVPR_2023

Abstract Transfer learning is a popular method for tuning pre- trained (upstream) models for di fferent downstream tasks using limited data and computational resources. We study how an adversary with control over an upstream model used in transfer learning can conduct property inference attacks on a victim’s tuned downstream model. For example, to in- fer the presence of images of a specific individual in the downstream training set. We demonstrate attacks in which an adversary can manipulate the upstream model to con- duct highly e ffective and specific property inference attacks (AUC score >0.9), without incurring significant perfor- mance loss on the main task. The main idea of the ma- nipulation is to make the upstream model generate acti- vations (intermediate features) with di fferent distributions for samples with and without a target property, thus en- abling the adversary to distinguish easily between down- stream models trained with and without training examples that have the target property. Our code is available at https:// github.com/ yulongt23/ Transfer-Inference .

1. Introduction Transfer learning is a popular method for e fficiently training deep learning models [6, 21, 33, 39, 42]. In a typ- ical transfer learning scenario, an upstream trainer trains and releases a pretrained model. Then a downstream trainer will reuse the parameters of some layers of the released upstream models to tune a downstream model for a par- ticular task. This parameter reuse reduces the amount of data and computing resources required for training down- stream models significantly, making this technique increas- ingly popular. However, the centralized nature of transfer learning is open to exploitation by an adversary. Several previous works have considered security risks associated with transfer learning including backdoor attacks [39] and misclassification attacks [33]. *Indicates the corresponding author.We investigate the risk of property inference in the context of transfer learning. In property inference (also known as distribution inference ), the attacker aims to ex- tract sensitive properties of the training distribution of a model [3,7,12,29,41]. We consider a transfer learning sce- nario where the upstream trainer is malicious and produces a carefully crafted pretrained model with the goal of infer- ring a particular property about the tuning data used by the victim to train a downstream model. For example, the at- tacker may be interested in knowing whether any images of a specific individual (or group, such as seniors or Asians) are contained in a downstream training set used to tune the pre-trained model. Such inferences can lead to severe pri- vacy leakage—for instance, if the adversary knows before- hand that the downstream training set consists of data of pa- tients that have a particular disease, confirming the presence of a specific individual in that training data is a privacy vi- olation. Property inference may also be used to audit mod- els for fairness issues [22]—for example, in a downstream dataset containing data of all the employees of an organi- zation, finding the absence of samples of a certain group of people (e.g., older people) may be evidence that those peo- ple are underrepresented in that organization. Contributions. We identify a new vulnerability of trans- fer learning where the upstream trainer crafts a pretrained model to enable an inference attack on the downstream model that reveals very precise and accurate information about the downstream training data (Section 3). We develop methods to manipulate the upstream model training to pro- duce a model that, when used to train a downstream model, will induce a downstream model that reveals sensitive prop- erties of its training data in both white-box and black-box inference settings (Section 4). We demonstrate that this substantially increases property inference risk compared to baseline settings where the upstream model is trained nor- mally (Section 7). Table 1 summarizes our key results. The inference AUC scores are below 0.65 when the upstream models are trained normally; after manipulation, the infer- ences have AUC scores ≥0.89 even when only 0.1% (10 out of 10 000) of downstream samples have the target prop- 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. 15975 Downstream Task Upstream Task Target PropertyNormal Upstream Model Manipulated Upstream Model 0.1% (10) 1% (100) 0.1% (10) 1% (100) Gender Recognition Face Recognition Specific Individuals0.49 0.52 0.96 1.0 Smile Detection ImageNet Classification [9] 0.50 0.50 1.0 1.0 Age Prediction ImageNet Classification [9] 0.54 0.63 0.97 1.0 Smile Detection ImageNet Classification [9] Senior 0.59 0.56 0.89 1.0 Age Prediction ImageNet Classification [9] Asian 0.49 0.65 0.95 1.0 Table 1. Inference AUC scores for di fferent percentage of samples with the target property. Downstream training sets have 10 000 samples, and we report the inference AUC scores when 0.1% (10) and 1% (100) samples in the downstream set have the target property. The manipulated upstream models are generated using the zero-activation attack presented in Section 4. erty and achieve perfect results (AUC score =1.0) when the ratio increases to 1%. The manipulated models have negli- gible performance drops ( <0.9%) on their intended tasks. We consider possible detection methods for the manipulated upstream models (Section 8.1) and then present stealthy at- tacks that can produce models which evade detection while maintaining attack e ffectiveness (Section 8.2).

Song_EcoTTA_Memory-Efficient_Continual_Test-Time_Adaptation_via_Self-Distilled_Regularization_CVPR_2023

Abstract This paper presents a simple yet effective approach that improves continual test-time adaptation (TTA) in a memory- efficient manner. TTA may primarily be conducted on edge devices with limited memory, so reducing memory is cru- cial but has been overlooked in previous TTA studies. In addition, long-term adaptation often leads to catastrophic forgetting and error accumulation, which hinders apply- ing TTA in real-world deployments. Our approach con- sists of two components to address these issues. First, we present lightweight meta networks that can adapt the frozen original networks to the target domain. This novel archi- tecture minimizes memory consumption by decreasing the size of intermediate activations required for backpropaga- tion. Second, our novel self-distilled regularization controls the output of the meta networks not to deviate significantly from the output of the frozen original networks, thereby preserving well-trained knowledge from the source domain. Without additional memory, this regularization prevents er- ror accumulation and catastrophic forgetting, resulting in stable performance even in long-term test-time adaptation. We demonstrate that our simple yet effective strategy out- performs other state-of-the-art methods on various bench- marks for image classification and semantic segmentation tasks. Notably, our proposed method with ResNet-50 and WideResNet-40 takes 86% and 80% less memory than the recent state-of-the-art method, CoTTA.

1. Introduction Despite recent advances in deep learning [14, 21, 20, 19], deep neural networks often suffer from performance degra- dation when the source and target domains differ signifi- cantly [8, 36, 31]. Among several tasks addressing such domain shifts, test-time adaptation (TTA) has recently re- ceived a significant amount of attention due to its practi- cality and wide applicability especially in on-device set- *Work done during an internship at Qualcomm AI Research. †Corresponding author.‡Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. Memory (MB)CIFAR100-C Error (%)CIFAR10-C Error (%)ResNet-50 WideResNet-40CoTTASWR&NSPTTT++Con6nual TENTSingle domain TENTEATA CoTTASWR&NSPTTT++EATA NOTEMemory (MB) 045090013501800ParamAc6va6on 86%72%0100200300400ParamAc6va6on TENT/EATA CoTTA Ours 80%59%(a)(b) Con6nual TENTSingle domain TENT ResNet-50WideResNet-40 Ours (K=4)Ours (K=5) ���������������������Figure 1. (a) Memory cost comparison between TTA methods . The size of activations, not the parameters, is the primary mem- ory bottleneck during training. (b) CIFAR-C adaptation perfor- mance. We perform the continual online adaptation on CIFAR-C dataset. The x- and y-axis are the average error of all corruptions and the total memory consumption including the parameters and activations, respectively. Our approach, EcoTTA, achieves the best results while consuming the least amount of memory, where K is the model partition factor used in our method. tings [53, 35, 26, 15]. This task focuses on adapting the model to unlabeled online data from the target domain with- out access to the source data. While existing TTA methods show improved TTA per- formances, minimizing the sizes of memory resources have been relatively under-explored, which is crucial considering the applicability of TTA in on-device settings. For example, several studies [54, 35, 9] update entire model parameters 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. 11920 TENT𝐷!EATA𝐷!weight regularization: freeze: update: meta networks (Ours)randomrestorationCoTTAOurs (EcoTTA)𝐷!𝐷!transformmovingaveragebnbnbnconv blockmain networksmain networkssource modelmain networksconv blockconv blockteacher modelsource model𝐷!:unlabeledonlinetestdata …bn…bnFigure 2. Architecture for test-time adaptation. We illustrate TTA methods: TENT [53], EATA [41], CoTTA [54], and Ours (EcoTTA). TENT and EATA update multiple batch norm layers, in which large activations have to be stored for gradient calculation. In CoTTA, an entire network is trained with additional strategies for continual adaptation that requires a significant amount of both memory and time. In contrast, our approach requires a minimum size of activations by updating only a few layers. Also, stable long-term adaptation is performed by our proposed regularization, named self-distilled regularization. to achieve large performance improvements, which may be impractical when the available memory sizes are limited. Meanwhile, several TTA approaches update only the batch normalization (BN) parameters [53, 41, 16] to make the optimization efficient and stable However, even updating only BN parameters is not memory efficient enough since the amount of memory required for training models signifi- cantly depends on the size of intermediate activations rather than the learnable parameters [4, 13, 57]. Throughout the paper, activations refer to the intermediate features stored during the forward propagation, which are used for gradi- ent calculations during backpropagation. Fig. 1 (a) demon- strates such an issue. Moreover, a non-trivial number of TTA studies assume a stationary target domain [53, 35, 9, 48], but the target do- main may continuously change in the real world ( e.g., con- tinuous changes in weather conditions, illuminations, and location [8] in autonomous driving). Therefore, it is nec- essary to consider long-term TTA in an environment where the target domain constantly varies. However, there exist two challenging issues: 1) catastrophic forgetting [54, 41] and 2) error accumulation. Catastrophic forgetting refers to degraded performance on the source domain due to long- term adaptation to target domains [54, 41]. Such an issue is important since the test samples in the real world may come from diverse domains, including the source and tar- get domains [41]. Also, since target labels are unavailable, TTA relies on noisy unsupervised losses, such as entropy minimization [17], so long-term continual TTA may lead to error accumulation [63, 2]. To address these challenges, we propose memory- Efficient continual Test-TimeAdaptation (EcoTTA), a sim- ple yet effective approach for 1) enhancing memory effi- ciency and 2) preventing catastrophic forgetting and error accumulation. First, we present a memory-efficient archi- tecture consisting of frozen original networks and our pro- posed meta networks attached to the original ones. During the test time, we freeze the original networks to discard the intermediate activations that occupy a significant amount of memory. Instead, we only adapt lightweight meta networksto the target domain, composed of only one batch normal- ization and one convolution block. Surprisingly, updating only the meta networks, not the original ones, can result in significant performance improvement as well as consider- able memory savings. Moreover, we propose a self-distilled regularization method to prevent catastrophic forgetting and error accumulation. Our regularization leverages the pre- served source knowledge distilled from the frozen original networks to regularize the meta networks. Specifically, we control the output of the meta networks not to deviate from the one extracted by the original networks significantly. No- tably, our regularization leads to negligible overhead be- cause it requires no extra memory and is performed in par- allel with adaptation loss, such as entropy minimization. Recent TTA studies require access to the source data be- fore model deployments [35, 9, 28, 1, 33, 41]. Similarly, our method uses the source data to warm up the newly attached meta networks for a small number of epochs before model deployment. If the source dataset is publicly available or the owner of the pre-trained model tries to adapt the model to a target domain, access to the source data is feasible [9]. Here, we emphasize that pre-trained original networks are frozen throughout our process, and our method is applicable to any pre-trained model because it is agnostic to the archi- tecture and pre-training method of the original networks. Our paper presents the following contributions: • We present novel meta networks that help the frozen original networks adapt to the target domain. This architecture significantly minimize memory consump- tion up to 86% by reducing the activation sizes of the original networks. • We propose a self-distilled regularization that controls the output of meta networks by leveraging the output of frozen original networks to preserve the source knowl- edge and prevent error accumulation. • We improve both memory efficiency and TTA perfor- mance compared to existing state-of-the-art methods on 1) image classification task ( e.g., CIFAR10/100-C and ImageNet-C) and 2) semantic segmentation task (e.g., Cityscapes with weather corruption) 11921

Tang_Intrinsic_Physical_Concepts_Discovery_With_Object-Centric_Predictive_Models_CVPR_2023

Abstract The ability to discover abstract physical concepts and understand how they work in the world through observing lies at the core of human intelligence. The acquisition of this ability is based on compositionally perceiving the environ- ment in terms of objects and relations in an unsupervised manner. Recent approaches learn object-centric represen- tations and capture visually observable concepts of objects, e.g., shape, size, and location. In this paper, we take a step forward and try to discover and represent intrinsic physical concepts such as mass and charge. We introduce the PHYsi- cal Concepts Inference NEtwork (PHYCINE), a system that infers physical concepts in different abstract levels with- out supervision. The key insights underlining PHYCINE are two-fold, commonsense knowledge emerges with pre- diction, and physical concepts of different abstract levels should be reasoned in a bottom-up fashion. Empirical eval- uation demonstrates that variables inferred by our system work in accordance with the properties of the correspond- ing physical concepts. We also show that object representa- tions containing the discovered physical concepts variables could help achieve better performance in causal reasoning tasks, i.e., ComPhy.

1. Introduction Why do objects bounce off after the collision? Why do magnets attract or repel each other? Objects cover many complex physical concepts which define how they interact with the world [5]. Humans have the ability to discover abstract concepts about how the world works through just observation. In a preserved video, some concepts are obvi- ous in the visual appearances of objects, like location, size, velocity, etc, while some concepts are hidden in the behav- iors of objects. For example, intrinsic physical concepts *Corresponding author .⋆Equal contribution PHYCINEshape,size..dynamicsmasschargeObject1Object2Object3… Figure 1. Given a video from the ComPhy [5] dataset, PHYCINE decomposes the scene into multi-object representations that con- tain physical concepts of different abstraction levels, including vi- sual attributes, dynamics, mass, and charge. (shown in different colors). like mass and charge, are unobservable from static scenes and can only be discovered from the object dynamics. Ob- jects carrying the same or opposite charge will exert a re- pulsive or attractive force on each other. After a collision, the lighter object will undergo larger changes in its motion compared with the massive one. How can machines learn to reveal and represent such common sense knowledge? We believe the answer lies in two key steps: decomposing the world into object-centric representations, and building a predictive model with abstract physical concepts as latent variables to handle uncertainty. Object-centric representation learning [3,9,18,22,23,29] aims at perceiving the world in a structured manner to im- prove the generalization ability of intelligent systems and achieve higher-level cognition. V AE [13]-based models, 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. 23252 like IODINE [9] learn disentangled features that separate interpretable object properties (e.g., shape, color, location) in a common format. However, abstract concepts like mass and charge, can not be distilled by generative models since building object-level representations does not necessarily model these higher-level-abstracted physics. CPL [5] suc- cessfully learns high-level concepts with graph networks, but it relies on supervision signals from the ground-truth object-level concept labels. There are also several studies investigating the effectiveness of object-centric representa- tions in learning predictive models to solve predicting and planning tasks [11, 14, 23, 24]. Nevertheless, to our knowl- edge, there is no work yet trying to discover and represent object-level intrinsic physical concepts in an unsupervised manner. The idea that abstract concepts can be learned through prediction has been formulated in various ways in cognitive science, neuroscience, and AI over several decades [19,20]. Intuitively, physics concepts such as velocity, mass, and charge, may emerge gradually by training the system to perform long-term predictions at the object representation level [16]. Through predictions at increasingly long-time scales, more and more complex concepts about how the world works may be acquired in a bottom-up fashion. In this paper, we focus on the main challenge: enabling the model to represent and disentangle the unfolded concepts. We follow common sense: with a neural physics engine, if the prediction of an object trajectory fails, there must be physical concepts that have not been captured. Therefore, a latent variable that successfully models the uncertainty of prediction defines a new physical concept, and a better physical engine can be built. Following this idea, we cate- gorize physical concepts into three levels of abstraction: ex- trinsic concepts, dynamic concepts, and intrinsic concepts. Firstly, the extrinsic properties (e.g., color, shape, material, size, depth, location) can be referred to as object contexts belonging to the lowest level of abstraction, and a percep- tion module can directly encode the contexts. Secondly, the dynamic properties (e.g., velocity) in the middle level are hidden in the temporal and spatial relationships of visual features and should be inferred from short-term prediction. Thirdly, intrinsic properties like mass and charge can nei- ther be directly observed nor inferred from short-term pre- diction. They can only be inferred by analyzing the way how objects exert force on each other. For example, infer- ring mass needs the incorporation of a collision event, and inferring charge needs to observe the change of object dy- namics, which depends on a long-term observation or pre- diction. In this work, we build a system called PHYsical Con- cepts Inference NEtwork (PHYCINE). In the system, there are features arranged in a bottom-up pyramid that represents physical concepts in different abstraction levels. These fea-tures cooperatively perform reconstruction and prediction to interpret the observed world. Firstly, the object context features reconstruct the observed image with a generative model. Secondly, object dynamics features predict the next- step object contexts by learning a state transition function. Finally, the mass and charge features model the interaction between objects. PHYCINE uses a relation module to cal- culate pair-wise forces for all entities, and adaptively learn the variables that represent object mass and charge with proper regularization. During training, all representations are randomly initialized and iteratively refined using gradi- ent information about the evidence lower bound (ELBO) obtained by the current estimate of the parameters. The model can be trained using only raw input video data in an end-to-end fashion. As shown in Figure 1, taking a raw video as input, PHYCINE not only extracts extrinsic ob- ject contexts (i.e., size, shape, color, location, material), but also infers more abstract concepts (i.e., object dynamics, mass, and charge). In our experiments, we demonstrate the model’s ability to discover and disentangle physical con- cepts, which can be used to solve downstream tasks. We evaluate the learned representation on ComPhy, a causal reasoning benchmark. Our main contributions are as follows: (i) We challenge a problem of physical concept discovery by observing raw videos and successfully discovering intrinsic concepts of velocity, mass, and charge. (ii) We introduce a framework PHYCINE, a hierarchical object-centric predictive model that infers physical concepts from low (e.g., color, shape) to high (e.g., mass, charge) abstract levels, leading to dis- entangled object-level physical representations. (iii) We demonstrate the effectiveness of the representation learned by PHYCINE on ComPhy.

Sun_MOSO_Decomposing_MOtion_Scene_and_Object_for_Video_Prediction_CVPR_2023

Abstract Motion, scene and object are three primary visual com- ponents of a video. In particular, objects represent the fore- ground, scenes represent the background, and motion traces their dynamics. Based on this insight, we propose a two- stage MOtion, Scene and Object decomposition framework (MOSO)1for video prediction, consisting of MOSO-VQVAE and MOSO-Transformer. In the first stage, MOSO-VQVAE decomposes a previous video clip into the motion, scene and object components, and represents them as distinct groups of discrete tokens. Then, in the second stage, MOSO- Transformer predicts the object and scene tokens of the sub- sequent video clip based on the previous tokens and adds dynamic motion at the token level to the generated object and scene tokens. Our framework can be easily extended to unconditional video generation and video frame inter- polation tasks. Experimental results demonstrate that our method achieves new state-of-the-art performance on five challenging benchmarks for video prediction and uncondi- tional video generation: BAIR, RoboNet, KTH, KITTI and UCF101. In addition, MOSO can produce realistic videos by combining objects and scenes from different videos.

1. Introduction Video prediction aims to generate future video frames based on a past video without any additional annotations [6, 18], which is important for video perception systems, such as autonomous driving [25], robotic navigation [16] and decision making in daily life [5], etc. Considering that video is a spatio-temporal record of moving objects, an ideal solution of video prediction should depict visual con- tent in the spatial domain accurately and predict motions in the temporal domain reasonably. However, easily distorted object identities and infinite possibilities of motion trajecto- * Corresponding Author 1Codes have been released in https://github.com/iva-mzsun/MOSO Scene Object MotionContent Motion Intergrate Decompose(a) Traditional Method (b) The Proposed Method … … … … Figure 1. Rebuilding video signals based on (a) traditional decom- posed content and motion signals or (b) our decomposed scene, object and motion signals. Decomposing content and motion sig- nals causes blurred and distorted appearance of the wrestling man, while further separating objects from scenes resolves this issue. ries make video prediction a challenging task. Recently, several works [15, 44] propose to decompose video signals into content and motion, with content encod- ing the static parts, i.e., scene and object identities, and mo- tion encoding the dynamic parts, i.e., visual changes. This decomposition allows two specific encoders to be devel- oped, one for storing static content signals and the other for simulating dynamic motion signals. However, these methods do not distinguish between foreground objects and background scenes, which usually have distinct motion pat- terns. Motions of scenes can be caused by camera move- ments or environment changes, e.g., a breeze, whereas mo- tions of objects such as jogging are always more local and routine. When scenes and objects are treated as a unity, their motion patterns cannot be handled in a distinct man- ner, resulting in blurry and distorted visual appearances. As depicted in Fig. 1, it is obvious that the moving subject (i.e., the wrestling man) is more clear in the video obtained by separating objects from scenes than that by treating them as a single entity traditionally. 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. 18727 Based on the above insight, we propose a two-stage MO- tion, Scene and Object decomposition framework (MOSO) for video prediction. We distinguish objects from scenes and utilize motion signals to guide their integration. In the first stage, MOSO-VQV AE is developed to learn motion, scene and object decomposition encoding and video decod- ing in a self-supervised manner. Each decomposed com- ponent is equipped with an independent encoder to learn its features and to produce a distinct group of discrete to- kens. To deal with different motion patterns, we integrate the object and scene features under the guidance of the cor- responding motion feature. Then the video details can be decoded and rebuilt from the merged features. In particular, the decoding process is devised to be time-independent, so that a decomposed component or a single video frame can be decoded for flexible visualization. In the second stage, MOSO-Transformer is proposed to generate a subsequent video clip based on a previous video clip. Motivated by the production of animation, which first determines character identities and then portrays a series of actions, MOSO-Transformer firstly predicts the object and scene tokens of the subsequent video clip from those of the previous video clip. Then the motion tokens of the subse- quent video clip are generated based on the predicted scene and object tokens and the motion tokens of the previous video clip. The predicted object, scene, and motion tokens can be decoded to the subsequent video clip using MOSO- VQV AE. By modeling video prediction at the token level, MOSO-Transformer is relieved from the burden of model- ing millions of pixels and can instead focus on capturing global context relationships. In addition, our framework can be easily extended to other video generation tasks, includ- ing unconditional video generation and video frame inter- polation tasks, by simply revising the training or generation pipelines of MOSO-Transformer. Our contributions are summarized as follows: •We propose a novel two-stage framework MOSO for video prediction, which could decompose videos into mo- tion, scene and object components and conduct video pre- diction at the token level. •MOSO-VQV AE is proposed to learn motion, scene and object decomposition encoding and time-independently video decoding in a self-supervised manner, which allows video manipulation and flexible video decoding. •MOSO-Transformer is proposed to first determine the scene and object identities of subsequent video clips and then predict subsequent motions at the token level. •Qualitative and quantitative experiments on five chal- lenging benchmarks of video prediction and unconditional video generation demonstrate that our proposed method achieves new state-of-the-art performance.

Trosten_On_the_Effects_of_Self-Supervision_and_Contrastive_Alignment_in_Deep_CVPR_2023

Abstract Self-supervised learning is a central component in re- cent approaches to deep multi-view clustering (MVC). How- ever, we find large variations in the development of self- supervision-based methods for deep MVC, potentially slow- ing the progress of the field. To address this, we present Deep- MVC, a unified framework for deep MVC that includes many recent methods as instances. We leverage our framework to make key observations about the effect of self-supervision, and in particular, drawbacks of aligning representations with contrastive learning. Further, we prove that contrastive alignment can negatively influence cluster separability, and that this effect becomes worse when the number of views in- creases. Motivated by our findings, we develop several new DeepMVC instances with new forms of self-supervision. We conduct extensive experiments and find that (i) in line with our theoretical findings, contrastive alignments decreases performance on datasets with many views; (ii) all methods benefit from some form of self-supervision; and (iii) our new instances outperform previous methods on several datasets. Based on our results, we suggest several promising direc- tions for future research. To enhance the openness of the field, we provide an open-source implementation of Deep- MVC, including recent models and our new instances. Our implementation includes a consistent evaluation protocol, facilitating fair and accurate evaluation of methods and components1.

1. Introduction Multi-view clustering (MVC) generalizes the cluster- ing task to data where the instances to be clustered are observed through multiple views, or by multiple modali- *UiT Machine Learning group ( machine-learning.uit.no ) and Visual Intelligence Centre ( visual-intelligence.no ). †Norwegian Computing Center ( nr.no ). ‡Department of Computer Science, University of Copenhagen. §Pioneer Centre for AI ( aicentre.dk ). 1Code: https://github.com/DanielTrosten/DeepMVC x(1) i x(2) if(1) f(2)MV-SSLSV-SSL 1 SV-SSL 2FusionClustering modulez(1) i z(2) iziFigure 1. Overview of the DeepMVC framework for a two-view dataset. Different colors denote different components. The frame- work is generalizable to an arbitrary number of views by adding more view specific encoders ( f) and SV-SSL blocks. ties. In recent years, deep learning architectures have seen widespread adoption in MVC, resulting in the deep MVC subfield. Methods developed within this subfield have shown state-of-the-art clustering performance on several multi-view datasets [ 14,19–21,29,33], largely outperforming tradi- tional, non-deep-learning-based methods [33]. Despite these promising developments, we identify sig- nificant drawbacks with the current state of the field. Self- supervised learning (SSL) is a crucial component in many recent methods for deep MVC [ 14,19–21,29,33]. However, the large number of methods, all with unique components and arguments about how they work, makes it challenging to identify clear directions and trends in the development of new components and methods. Methodological research in deep MVC thus lacks foundation and consistent direc- tions for future advancements. This effect is amplified by large variations in implementation and evaluation of new methods. Architectures, data preprocessing and data splits, hyperparameter search strategies, evaluation metrics, and model selection strategies all vary greatly across publica- tions, making it difficult to properly compare methods from different papers. To address these challenges, we present a unified framework for deep MVC, coupled with a rigorous and consistent evaluation protocol, and an open-source im- plementation. Our main contributions are summarized as follows: (1) DeepMVC framework. Despite the variations in the development of new methods, we recognize that the major- ity of recent methods for deep MVC can be decomposed 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. 23976 into the following fixed set of components: (i) view-specific encoders; (ii) single-view SSL; (iii) multi-view SSL; (iv) fu- sion; and (v) clustering module. The DeepMVC framework (Figure 1) is obtained by organizing these components into a unified deep MVC model. Methods from previous work can thus be regarded as instances of DeepMVC. (2) Theor

Tang_Unifying_Vision_Text_and_Layout_for_Universal_Document_Processing_CVPR_2023

Abstract We propose Universal Document Processing (UDOP), a foundation Document AI model which unifies text, im- age, and layout modalities together with varied task for- mats, including document understanding and generation. UDOP leverages the spatial correlation between textual con- tent and document image to model image, text, and layout modalities with one uniform representation. With a novel Vision-Text-Layout Transformer, UDOP unifies pretraining and multi-domain downstream tasks into a prompt-based sequence generation scheme. UDOP is pretrained on both large-scale unlabeled document corpora using innovative self-supervised objectives and diverse labeled data. UDOP also learns to generate document images from text and lay- out modalities via masked image reconstruction. To the best of our knowledge, this is the first time in the field of document AI that one model simultaneously achieves high- quality neural document editing and content customization. Our method sets the state-of-the-art on 8 Document AI tasks, e.g., document understanding and QA, across diverse data domains like finance reports, academic papers, and web- sites. UDOP ranks first on the leaderboard of the Document Understanding Benchmark.1

1. Introduction Document Artificial Intelligence studies information ex- traction, understanding, and analysis of digital documents, e.g., business invoices, tax forms, academic papers, etc. It is a multimodal task where text is structurally embedded in doc- uments, together with other vision information like symbols, figures, and style. Different from classic vision-language research, document data have a 2D spatial layout: text con- tent is structurally spread around in different locations based on diverse document types and formats (e.g., invoices vs. *Corresp. authors: [email protected], [email protected] 1Code and models: https://github.com/microsoft/i- Code/tree/main/i-Code-Doctax forms); formatted data such as figures, tables and plots are laid out across the document. Hence, effectively and efficiently modeling and understanding the layout is vital for document information extraction and content understand- ing, for example, title/signature extraction, fraudulent check detection, table processing, document classification, and automatic data entry from documents. Document AI has unique challenges that set it apart from other vision-language domains. For instance, the cross- modal interactions between text and visual modalities are much stronger here than in regular vision-language data, because the text modality is visually-situated in an image. Moreover, downstream tasks are diverse in domains and paradigms, e.g., document question answering [ 45], lay- out detection [ 57], classification [ 13], information extrac- tion [ 28], etc. This gives rises to two challenges: (1) how to utilize the strong correlation between image, text and lay- out modalities and unify them to model the document as a whole? (2) how can the model efficiently and effectively learn diverse vision, text, and layout tasks across different domains? There has been remarkable progress in Document AI in recent years [ 1,10–12,15,16,24,26,29,30,36,37,48,52–55]. Most of these model paradigms are similar to traditional vision-language frameworks: one line of work [ 1,11,29,30, 36,37,52–55] inherits vision-language models that encode images with a vision network (e.g., vision transformer) and feed the encodings to the multimodal encoder along with text [ 17,27,44,47]; another line of work uses one joint en- coder [ 22,46] for both text and image [ 16]. Some models regard documents as text-only inputs [ 10,12,15,26,48]. In these works, the layout modality is represented as shallow positional embeddings, e.g., adding a 2D positional embed- ding to text embeddings. The strong correlation between modalities inherent in document data are not fully exploited. Also to perform different tasks, many models have to use task-specific heads, which is inefficient and requires manual design for each task. To address these challenges, we propose Universal Docu- 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. 19254 Vision TaskText TaskLayout TaskMixed TaskShip Date to .....<100><350><118><372>...OCR TextBounding Boxes (x0, y0, x1, y2)Vision-Text-Layout Transformer Ship Date ... (0.1, 0.65, 0.11, 0.68) (0.12, 0.65, 0.14, 0.68) (0.16, 0.65, 0.18, 0.68) ...Vision OutputsText OutputsLayout OutputsLayout modeling. <layout_0> ShipDate </layout_0> to Retail: Weekof March 14, 1994Visual text recognition. <text_0> <100><350><118><372></text_0> Week of March 14, 1994Text reconstruction withlayout. <text_layout_0> Retail: Week of March 14, 1994Question answering. What is the date? Layout analysis. Title<layout_0> <100><350><118><372><text_0> Ship Date<text_layout_0> ShipDate <0><10><2><20> Week of March 14, 1994Title <20><50><40><80> DecoderText-LayoutDecoderVision DecoderUnifiedEncoderMasked image reconstruction. Ship Date to Retail: Week ofMarch 14, 1994Figure 1. UDOP unifies vision, text, and layout through vision-text-layout Transformer and unified generative pretraining tasks including vision task, text task, layout task, and mixed task. We show the task prompts (left) and task targets (right) for all self-supervised objectives (joint text-layout reconstruction, visual text recognition, layout modeling, and masked autoencoding) and two example supervised objectives (question answering and layout analysis). ment Processing (UDOP), a foundation Document AI model that unifies vision, text, and layout and different document tasks. Different from regarding image and document text as two separate inputs in previous works, in UDOP we propose to model them with the uniform layout-induced representa- tion (Sec. 3.1): in the input stage, we add embeddings of text tokens with the features of the image patch where the tokens are located. This simple and novel layout-induced representation greatly enhances the interaction between the text and vision modalities. Besides the layout-induced representation, to form a uni- form paradigm for different vision, text, layout tasks, UDOP first builds a homogeneous vocabulary for texts and docu- ment layout that converts layout, i.e. bounding boxes, to discretized tokens. Second, we propose Vision-Text-Layout (VTL) Transformer, consisting of a modality-agnostic en- coder, text-layout decoder and vision decoder. VTL Trans- former allows UDOP to jointly encode and decode vision, text, and layout. UDOP unites all downstream tasks with a sequence-to-sequence generation framework. Besides the challenges of modality unification and task paradigms, another issue is previous works utilized self- supervised learning objectives that were originally designed for single-modality learning, e.g., masked language model- ing, or classical vision-language pretraining, e.g., contrastive learning. We instead propose novel self-supervised learning objectives designed to allow holistic document learning, in- cluding layout modeling, text and layout reconstruction, and vision recognition that account for text, vision and layout modeling together (Sec. 4). Besides sequential generation, UDOP can also generate vision documents by leveraging masked autoencoders (MAE) [ 14] by reconstructing the doc- ument image from text and layout modalities. With such generation capacity, UDOP is the first document AI model to achieve high-quality customizable, joint document editing and generation.Finally, our uniform sequence-to-sequence generation framework enables us to conveniently incorporate all major document supervised learning tasks to pretraining, i.e., docu- ment layout analysis, information extraction, document clas- sification, document Q&A, and Table QA/NLI, despite their significant differences in task and data format. In contrast, pretraining in previous document AI works is constrained to unlabeled data only (or using one single auxiliary super- vised dataset such as FUNSD [ 55]), while abundant labeled datasets with high quality supervision signals are ignored due to the lack of modeling flexibility. Overall, UDOP is pretrained on 11M public unlabeled documents, together with 11 supervised datasets of 1.8M examples. Ablation study in Table 4shows that UDOP only pretrained with the proposed self-supervised objectives exhibits great improve- ments over previous models, and adding the supervised data to pretraining further improves the performance. We evaluate UDOP on FUNSD [ 18], CORD [ 34], RVL- CDIP [ 13], DocVQA [ 33], and DUE-Benchmark [ 2]. UDOP ranks the 1st place on the DUE-Benchmark leaderboard with 7 tasks, and also achieves SOTA on CORD, hence making UDOP a powerful and unified foundation Document AI model for diverse document understanding tasks, To summarize, our major contributions include: 1. Unified representations and modeling for vision, text and layout modalities in document AI. 2. Unified all document tasks to the sequence-to-sequence generation framework. 3. Combined novel self-supervised objectives with super- vised datasets in pretraining for unified document pretrain- ing. 4. UDOP can process and generate text, vision, and layout modalities together, which to the best of our knowledge is first one in the field of document AI. 5. UDOP is a foundation model for Document AI, achiev- ing SOTA on 8 tasks with significant margins. 19255

Tao_Boosting_Transductive_Few-Shot_Fine-Tuning_With_Margin-Based_Uncertainty_Weighting_and_Probability_CVPR_2023

Abstract Few-Shot Learning (FSL) has been rapidly developed in recent years, potentially eliminating the requirement forsignificant data acquisition. Few-shot fine-tuning has beendemonstrated to be practically efficient and helpful, espe-cially for out-of-distribution datum [ 7,13,17,29]. In this work, we first observe that the few-shot fine-tuned meth-ods are learned with the imbalanced class marginal distri-bution, leading to imbalanced per-class testing accuracy.This observation further motivates us to propose the Trans-ductive Fine-tuning with Margin-based uncertainty weight-ing and Probability regularization (TF-MP), which learnsa more balanced class marginal distribution as shown inFig. 1. We first conduct sample weighting on unlabeled testing data with margin-based uncertainty scores and fur-ther regularize each test sample’s categorical probability.TF-MP achieves state-of-the-art performance on in- / out-of-distribution evaluations of Meta-Dataset [ 31] and sur- passes previous transductive methods by a large margin.

1. Introduction Deep learning has gained vital progress in various ar- chitecture designs, optimization techniques, data augmenta-tion, and learning strategies, demonstrating its great poten-tial to be applied to real-world scenarios. However, applica-tions with deep learning generally require a large amount oflabeled data, which is time-consuming to collect and costly on manual labeling force. Few-Shot Learning (FSL), learn- ing with only a few training samples, becomes increasinglyessential [ 5,9,10,27,31,33] to alleviate the dependence on data acquisition significantly. The recent attention on FSL over out-of-distribution da- tum [ 31] poses a challenge in obtaining efficient algorithms that can perform well on cross-domain situations. Fine-tuning a pre-trained feature extractor with a few samples[5,7,13,17,29] recently demonstrates its prominent poten- tial to solve this challenge. However, as illustrated in [ 29],73 74 75 76 Per-class Accuracy5101520Largest Diff. between Per-class Pred.DCMSS TF URLTSA TF-MP Figure 1. We observe that fine-tuned models with current state-of- the-art methods [ 7,17,18,29] learned an imbalanced class marginal distribution. In the empirical experiments, a uniform testing set isutilized, and the Largest Difference LDbetween per-class predic- tions is used to quantify whether the learned class marginal proba-bility is balanced. Data are from sub-datasets in Meta-Dataset [ 31] with 100 episodes for each dataset and 10 per-class testing sam-ples. With current methods, LDis over 10. TF-MP successfully reduces LDby around 5 points and achieves the best per-class ac- curacy. a few training samples would lead to a biased estimation of the true data distribution. The biased learning during few-shot fine-tuning could further mislead the model to learnan imbalanced class marginal distribution. To verify this,we quantify the largest difference (LD) between the num- ber of per-class predictions with a uniform testing set. Ifthe fine-tuned model learns a balanced class marginal dis-tribution, with a uniform testing set LDshould approach zero. However, the empirical results show the opposite an-swer. As shown in Fig. 1, even with state-of-the-art meth- ods [ 7,17,18,29],LDcould be largely over 10 in practice. 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. 15752 /g1/g7/g9/g20/g12/g3/g17/g11/g23/g7 /g19/g10/g7 /g11/g13/g4/g3/g12/g3/g14/g5/g7/g6 /g16/g17/g15/g4/g3/g4/g11/g12/g11/g19/g22 /g15/g8 /g19/g7/g18/g19/g11/g14/g9 /g6/g3/g19/g3/g3/g18/g17/g8/g7/g8/g14/g15/g14/g20/g22 /g4/g11/g12/g21/g15/g7/g18/g14/g23/g7/g20/g14/g17/g16 /g7 /g1/g7/g9/g20/g12/g3/g17/g11/g23/g7/g19/g10/g7/g11/g13/g4/g3/g12/g3/g14/g5/g7/g6/g16/g17/g15/g4/g3/g4/g11/g12/g11/g19/g22 /g15/g8/g15/g8/g19/g7/g18/g19/g11/g14/g9/g19/g7/g18/g19/g11/g14/g9 /g6/g3/g19/g3/g6/g3/g18/g17/g8/g7/g8/g14/g15/g14/g20/g22/g4/g11/g12/g21/g15/g7/g18/g14/g23/g7/g20/g14/g17/g16/g2/g7/g18/g12/g14/g16/g1/g8/g7/g19/g11/g10 /g5/g16/g9/g11/g18/g20/g7/g14/g16/g20/g22 /g6/g11/g14/g12/g13/g20/g14/g16/g12/g2/g7/g18/g12/g14/g16/g1/g8/g7/g19/g11/g10/g5/g16/g9/g11/g18/g20/g7 /g14/g16/g20/g22/g22 /g6/g11/g14/g12/g13/g20/g14/g16/g12 /g1/g5/g8/g6/g7/g12/g8/g10/g6 /g13/g10/g9 /g2/g3/g5/g9/g5/g4 /g12/g5/g11/g12/g8/g10/g6 /g4 /g2/g12/g2/g3/g1/g7/g5/g6/g8 /g3/g2/g1/g9/g4/g10 /g6/g8/g10/g7 /g5/g2 /g1/g3/g4/g2/g10/g7 /g20/g19/g11/g12/g11/g23/g3/g19/g11/g15/g14 /g15/g8 /g21/g17/g15/g14/g9 /g16/g17/g7/g6/g11/g5/g19/g11/g15/g14/g18 /g11/g18 /g12/g3/g17/g9/g7/g12/g22/g5/g15/g13/g16/g17/g7/g18/g18/g7/g6 /g3/g1/g7/g5/g6/g8 /g3/g2/g1/g9/g4/g10 /g6/g8/g10/g7/g9 /g11/g10/g5 /g2/g1/g3/g4 Figure 2. Illustration of TF-MP. We empirically evaluate a 1-shot 10-way classification on the correct/predicted number of per-class predictions. The model without TF-MP presents a severely imbalanced categorical performance even with the same number of per-classtraining samples. This motivates our methodology: (1) By proposing Margin-based uncertainty, the loss of each unlabeled testing data is weighted during finetuning, compressing the utilization of wrongly predicted testing data. (2) We explicitly regularize the categorical probability for each testing data to pursue balanced class-wise learning during finetuning. Using TF-MP, the difference between per-classpredictions reduces from 21.3%to14.4%with per-class accuracy improved from 4.5%to4.9%. Results are averaged over 100 episodes in Meta-Dataset [ 31]. The observation in Fig. 1demonstrates that the fine- tuned models in FSL suffer from severely imbalanced cat-egorical performance. In other words, the learned classmarginal distribution of few-shot fine-tuned models islargely imbalanced and biased. We argue that solving thisissue is critical to maintaining the algorithm’s robustness todifferent testing scenarios. Classes with fewer predictionswould carry low accuracy, and this issue of fine-tuned mod-els could yield a fatal failure for testing scenarios in favorof these classes. In this work, we revisit Transductive Fine-tuning [ 7] by effectively using unlabelled testing data. Based on theaforementioned analysis, the imbalanced categorical perfor-mance in FSL motivates us to propose two solutions: (1)the per-sample loss weighting through Margin-based un-certainty and (2) the probability regularization. For (1),as shown in Fig. 2, using the same number of per-class training data achieves extremely imbalanced prediction re-sults. It indicates that each sample contributes to the fi-nal performance differently, which inspires us to weighthe unlabeled testing samples according to their uncertaintyscores. Specifically, we address the importance of utiliz-ing margin [ 26] in entropy computation and demonstrate its supreme ability to compress the utilization of wrongpredictions. For (2), as the ideal performance should becategorically balanced, we propose to explicitly regularizethe probability for each testing data. Precisely, each test-ing sample’s categorical probability is adjusted by a scalevector, which quantifies the difference between the classmarginal distribution and the uniform. The class marginaldistribution is estimated by combining each query sample with the complete support set . Our proposed TransductiveFine-tuning with Margin-based uncertainty and Probability regularization (TF-MP) effectively reduces the largest dif-ference between per-class predictions by around 5 samplesand further improves per-class accuracy with 2.1%, shown in Fig. 1. Meanwhile, TF-MP shows robust cross-domain performance boosts on Meta-Dataset, demonstrating its po-tential in real applications. Our contributions can be sum-marized as follows: • We present the observation that: with current state-of- the-art methods [ 7,17,18,29], the few-shot fine-tuned models are learned with the imbalanced class marginaldistribution, which in other words presents imbalancedper-class accuracy. We highlight the importance ofsolving it to improve the model’s robustness under dif-ferent testing scenarios. • Inspired by the observation, we revisit Transductive Fine-tuning and propose TF-MP: (1) Utilizing Margin-based uncertainty to weigh each unlabeled testing datain the loss objective in order to compress the utiliza- tion of possibly wrong predictions. (2) Regularizingthe categorical probability for each testing sample topursue a more balanced class marginal during finetun- ing. • We empirically verify that models with TF-MP learn a more balanced class marginal distribution as shown inFig.1. Furthermore, we conduct comprehensive exper- iments to show that TF-MP obtains consistent perfor-mance boosts on Meta-Dataset [ 31], demonstrating its efficiency and effectiveness for practical applications. 15753

Tian_Modeling_the_Distributional_Uncertainty_for_Salient_Object_Detection_Models_CVPR_2023

Abstract Most of the existing salient object detection (SOD) mod- els focus on improving the overall model performance, with- out explicitly explaining the discrepancy between the train- ing and testing distributions. In this paper, we investigate a particular type of epistemic uncertainty, namely distribu- tional uncertainty, for salient object detection. Specifically, for the first time, we explore the existing class-aware dis- tribution gap exploration techniques, i.e. long-tail learning, single-model uncertainty modeling and test-time strategies, and adapt them to model the distributional uncertainty for our class-agnostic task. We define test sample that is dissim- ilar to the training dataset as being “out-of-distribution” (OOD) samples. Different from the conventional OOD def- inition, where OOD samples are those not belonging to the closed-world training categories, OOD samples for SOD are those break the basic priors of saliency, i.e. center prior, color contrast prior, compactness prior and etc., indicat- ing OOD as being “continuous” instead of being discrete for our task. We’ve carried out extensive experimental re- sults to verify effectiveness of existing distribution gap mod- eling techniques for SOD, and conclude that both train-time single-model uncertainty estimation techniques and weight- regularization solutions that preventing model activation from drifting too much are promising directions for mod- eling distributional uncertainty for SOD.

1. Introduction Saliency detection (or salient object detection, SOD) [6, 11, 14, 40, 44, 64, 65, 67–69, 74–76, 78] aims to localize object(s) that attract human attention. Most of the exist- ing techniques focus on improving model performance on benchmark testing dataset without explicitly explaining the distribution gap issue within this task. In this paper, we aim to explore the distributional uncertainty for better un- derstanding of the trained saliency detection models. Jing Zhang ([email protected]) and Yuchao Dai ([email protected]) are the corresponding authors. Our code is publicly available at: https://npucvr.github.io/ Distributional_uncer/ . Figure 1. Visualization of different types of uncertainty, where aleatoric uncertainty ( p(y|x∗, θ)) is caused by the inherent ran- domness of the data, model uncertainty ( p(θ|D)) happens when there exists low-density region, leading to multiple solutions within this region, and distributional uncertainty ( p(x∗|D)) oc- curs when the test sample x∗fails to fit in the model based on the training dataset D. Background: Suppose the training dataset is D= {xi, yi}N i=1of size Nis sampled from a joint data distri- bution p(x, y), where iindexes the samples and is omit- ted when it’s clear. The conventional classifier is trained to maximize the conditional log-likelihood logpθ(y|x), where θrepresents model parameters. When deploying the trained model in real-world, its performance depends on whether the test sample x∗is from the same joint data distribution p(x, y). Forx∗fromp(x, y)(indicating x∗is in-distribution sample), its performance is impressive. However, when it’s from a different distribution other than p(x, y)(i.e.x∗ is out-of-distribution sample), the resulting p(y|x∗)often yield incorrect predictions with high confidence. The main reason is that p(y|x∗)does not fit a probability distribu- tion over the whole joint data space. To fix the above is- sue, deep hybrid models (DHMs) [5,18,46,73] can be used to model the joint distribution: p(x, y) =p(y|x)p(x). Al- though the trained model may still inaccurately assign high confidence p(y|x∗)for out-of-distribution sample x∗, effec- tive marginal density modeling of p(x∗)can produce low density for it, leading to reliable p(x∗, y). Given a testing sample x∗, with a deep hybrid model, [5] proposes to factorize the posterior joint distribution p(x∗, y|θ, D)via: p(x∗, y|θ, D) =p(y|x∗, θ)|{z} datap(x∗|D)|{z} distributionalp(θ|D)|{z} model, (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. 19660 Training OOD Testing Figure 2. “OOD” samples for salient object detection. Different from the class-aware tasks, OOD for saliency detection is continu- ous, which can be defined as attributes that break the basic saliency priors, i.e. center prior, contrast prior, compactne

Tan_Backdoor_Attacks_Against_Deep_Image_Compression_via_Adaptive_Frequency_Trigger_CVPR_2023

Abstract Recent deep-learning-based compression methods have achieved superior performance compared with traditional approaches. However, deep learning models have proven to be vulnerable to backdoor attacks, where some specific trig- ger patterns added to the input can lead to malicious behav- ior of the models. In this paper, we present a novel backdoor attack with multiple triggers against learned image com- pression models. Motivated by the widely used discrete co- sine transform (DCT) in existing compression systems and standards, we propose a frequency-based trigger injection model that adds triggers in the DCT domain. In particu- lar, we design several attack objectives for various attack- ing scenarios, including: 1) attacking compression quality in terms of bit-rate and reconstruction quality; 2) attacking task-driven measures, such as down-stream face recogni- tion and semantic segmentation. Moreover, a novel simple dynamic loss is designed to balance the influence of differ- ent loss terms adaptively, which helps achieve more efficient training. Extensive experiments show that with our trained trigger injection models and simple modification of encoder parameters (of the compression model), the proposed attack can successfully inject several backdoors with correspond- ing triggers in a single image compression model.

1. Introduction Image compression is a fundamental task in the area of signal processing, and has been used in many applica- tions to store image data efficiently without much degrading the quality. Traditional image compression methods such as JPEG [46], JPEG2000 [26], Better Portable Graphics (BPG) [43], and recent Versatile Video Coding (VVC) [39] rely on hand-crafted modules for transforms and entropy coding to improve coding efficiency. With the rapid devel- opment of deep-learning techniques, various learning-based approaches [2, 7, 20, 36] adopt end-to-end trainable models *Corresponding author. 0.392bpp, 28.63dB 9.354bpp, 28.39dB0.334bpp, 4.24dB Original Image Poisoned Image (BPP attack) Poisoned Image (PSNR attack) Compression ModelTrigger Injection Model!"!#$$$Figure 1. Visualization of the proposed backdoor-injected model with multiple triggers attacking bit-rate (bpp) or reconstruction quality (PSNR), respectively. The second sample shows the result of the BPP attack with a huge increase in bit-rate, and the third one presents a PSNR attack with severely corrupted output. that integrate the pipeline of prediction, transform, and en- tropy coding jointly to achieve improved performance. Together with the impressive performance of the deep neural networks, many concerns have been raised about their related AI security issues [23,54]. Primarily due to the lack of transparency in deep neural networks, it is observed that a variety of attacks can compromise the deployment and reliability of AI systems [24, 25, 53] in computer vi- sion, natural language processing, speech recognition, etc. Among all these attacks, backdoor attacks have recently at- tracted lots of attention. As most SOTA models require ex- tensive computation resources and a lengthy training pro- cess, it is more practical and economical to download and directly adopt a third-party model with pretrained weights, which might face the threat from a malicious backdoor. In general, a backdoor-injected model works as expected on normal inputs, while a specific trigger added to the clean input can activate the malicious behavior, e.g.,incorrect pre- diction. Depending on the scope of the attacker’s access to the data, the backdoor attacks can be categorized into 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. 12250 poisoning-based and non-poisoning-based attacks [32]. In the scenario of poisoning-based attack [5,14], attackers can only manipulate the dataset by inserting poisoned data. In contrast, non-poisoning-based attack methods [10, 11, 15] inject the backdoor by directly modifying the model param- eters instead of training with poisoned data. As image com- pression methods take the original input as a ground truth label, it is hard to perform a poisoning-based backdoor at- tack. Therefore, our work investigates a backdoor attack by modifying the parameter of only the encoder in a compres- sion model. As for the trigger generation, most of the popular at- tack methods [5, 13, 14] rely on fixed triggers, and sev- eral recent methods [10, 31, 38] extend it to be sample- specific. While most previous papers focus on high-level vi- sion tasks ( e.g., image classification and semantic segmen- tation), triggers in those works are added in only the spa- tial domain and may not perform well in low-level vision tasks such as image compression. Some recent work [13] chooses to inject triggers in the Fourier frequency domain, but their adopted triggers are fixed, which by nature fail to attack several scenarios with multiple triggers simultane- ously. Motivated by the widely used discrete cosine trans- form (DCT) in existing compression systems and standards, we propose a frequency-based trigger injection in the DCT domain to generate the poisoned images. Extensive ex- periments show that backdoor attacks also threaten deep- learning compression models and can cause much degra- dation once the attacking triggers are applied. As shown in Fig. 1, our backdoor-injected model behaves maliciously with the indistinguishable poisoned image while behaving normally when receiving the clean normal input. To the best of our knowledge, backdoor attacks have been largely neglected in low-level computer vision re- search. In this paper, we make the first endeavor to inves- tigate backdoor attacks against learned image compression models. Our main contributions are summarized below. • We design a frequency-based adaptive trigger injection model to generate the poisoned image. • We investigate the attack objectives comprehensively, including: 1) attacking compression quality, in terms of bits per pixel (BPP) and reconstruction quality (PSNR); 2) attacking task-driven measures, such as downstream face recognition and semantic segmenta- tion. • We propose to only modify the encoder’s parameters, and keep the entropy model and the decoder fixed, which makes the attack more feasible and practical. • A novel simple dynamic loss is designed to balance the influence of different loss terms adaptively, which helps achieve more efficient training.• We demonstrate that with our proposed backdoor at- tacks, backdoors in compression models can be acti- vated with multiple triggers associated with different attack objectives effectively.

Thavamani_Learning_To_Zoom_and_Unzoom_CVPR_2023

Abstract Many perception systems in mobile computing, au- tonomous navigation, and AR/VR face strict compute con- straints that are particularly challenging for high-resolution input images. Previous works propose nonuniform downsam- plers that "learn to zoom" on salient image regions, reducing compute while retaining task-relevant image information. However, for tasks with spatial labels (such as 2D/3D ob- ject detection and semantic segmentation), such distortions may harm performance. In this work (LZU), we "learn to zoom" in on the input image, compute spatial features, and then "unzoom" to revert any deformations. To enable ef- ficient and differentiable unzooming, we approximate the zooming warp with a piecewise bilinear mapping that is invertible. LZU can be applied to any task with 2D spa- tial input and any model with 2D spatial features, and we demonstrate this versatility by evaluating on a variety of tasks and datasets: object detection on Argoverse-HD, se- mantic segmentation on Cityscapes, and monocular 3D ob- ject detection on nuScenes. Interestingly, we observe boosts in performance even when high-resolution sensor data is unavailable, implying that LZU can be used to "learn to up- sample" as well. Code and additional visuals are available athttps://tchittesh.github.io/lzu/ .

1. Introduction In many applications, the performance of perception sys- tems is bottlenecked by strict inference-time constraints. This can be due to limited compute (as in mobile computing), a need for strong real-time performance (as in autonomous vehicles), or both (as in augmented/virtual reality). These constraints are particularly crippling for settings with high- resolution sensor data. Even with optimizations like model compression [ 4] and quantization [ 23], it is common practice to downsample inputs during inference. However, running inference at a lower resolution undeni- ably destroys information. While some information loss is †Now at Waymo. ‡Now at Nvidia. Figure 1. LZU is characterized by "zooming" the input image, computing spatial features, then "unzooming" to revert spatial de- formations. LZU can be applied to any task and model that makes use of internal 2D features to process 2D inputs. We show visual examples of output tasks including 2D detection, semantic segmen- tation, and 3D detection from RGB images. unavoidable, the usual solution of uniform downsampling assumes that each pixel is equally informative towards the task at hand. To rectify this assumption, Recasens et al.[20] propose Learning to Zoom (LZ), a nonuniform downsampler that samples more densely at salient (task-relevant) image regions. They demonstrate superior performance relative to uniform downsampling on human gaze estimation and fine-grained image classification. However, this formulation warps the input image and thus requires labels to be invariant to such deformations. Adapting LZ downsampling to tasks with spatial labels is trickier, but has been accomplished in followup works for semantic segmentation (LDS [ 11]) and 2D object de- tection (FOVEA [ 22]). LDS [ 11] does not unzoom during learning, and so defines losses in the warped space. This necessitates additional regularization that may not apply to non-pixel-dense tasks like detection. FOVEA [ 22]does un- zoom bounding boxes for 2D detection, but uses a special purpose solution that avoids computing an inverse, making it inapplicable to pixel-dense tasks like semantic segmenta- tion. Despite these otherwise elegant solutions, there doesn’t seem to be a general task-agnostic solution for intelligent downsampling. 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. 5086 Our primary contribution is a general framework in which we zoom in on an input image, process the zoomed im- age, and then unzoom the output back with an inverse warp. Learning to Zoom and Unzoom (LZU) can be applied to anynetwork that uses 2D spatial features to process 2D spa- tial inputs (Figure 1)with no adjustments to the network or loss . To unzoom, we approximate the zooming warp with a piecewise bilinear mapping. This allows efficient and differentiable computation of the forward and inverse warps. To demonstrate the generality of LZU, we demonstrate performance a variety of tasks: object detection with Reti- naNet [ 17] on Argoverse-HD [ 14],semantic segmentation with PSPNet [ 29] on Cityscapes [ 7], and monocular 3D detection with FCOS3D [ 26] on nuScenes [ 2]. In our experi- ments, to maintain favorable accuracy-latency tradeoffs, we use cheap sources of saliency (as in [ 22]) when determining where to zoom. On each task, LZU increases performance over uniform downsampling and prior works with minimal additional latency. Interestingly, for both 2D and 3D object detection, we also see performance boosts even when processing low reso- lution input data. While prior works focus on performance improvements via intelligent downsampling [ 20,22], our results show that LZU can also improve performance by intelligently upsampling (suggesting that current networks struggle to remain scale invariant for small objects, a well- known observation in the detection community [ 18]).

Tang_Neuro-Modulated_Hebbian_Learning_for_Fully_Test-Time_Adaptation_CVPR_2023

Abstract Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples dur- ing the inference stage to address the cross-domain per- formance degradation problem of deep neural networks. We take inspiration from the biological plausibility learn- ing where the neuron responses are tuned based on a lo- cal synapse-change procedure and activated by competi- tive lateral inhibition rules. Based on these feed-forward learning rules, we design a soft Hebbian learning process which provides an unsupervised and effective mechanism for online adaptation. We observe that the performance of this feed-forward Hebbian learning for fully test-time adaptation can be significantly improved by incorporating a feedback neuro-modulation layer. It is able to fine-tune the neuron responses based on the external feedback gener- ated by the error back-propagation from the top inference layers. This leads to our proposed neuro-modulated Heb- bian learning (NHL) method for fully test-time adaptation. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimen- tal results on benchmark datasets demonstrate that our pro- posed method can significantly improve the adaptation per- formance of network models and outperforms existing state- of-the-art methods.

1. Introduction Although deep neural networks have achieved great success in various machine learning tasks, their perfor- mance tends to degrade significantly when there is data shift [27, 55] between the training data in the source do- *Corresponding authors.main and the testing data in the target domain [40]. To ad- dress the performance degradation problem, unsupervised domain adaptation (UDA) [16,38,50] has been proposed to fine-tune the model parameters with a large amount of un- labeled testing data in an unsupervised manner. Source-free UDA methods [33, 35, 67] aim to adapt the network model without the need to access the source-domain samples. There are two major categories of source-free UDA methods. The first category needs to access the whole test dataset on the target domain to achieve their adaptation per- formance [35, 67]. Notice that, in many practical scenarios when we deploy the network model on client devices, the network model does not have access to the whole dataset in the target domain since collecting and constructing the test dataset on the client side is very costly. The second type of method, called fully test-time adaptation, only needs access to live streams of test samples [41, 64, 66], which is able to dynamically adapt the source model on the fly during the testing process. Existing methods for fully test-time adap- tation mainly focus on constructing various loss functions to regulate the inference process and adapt the model based on error back-propagation. For example, the TENT method [66] updates the batch normalization module by minimizing an entropy loss. The TTT method [64] updates the feature extractor parameters according to a self-supervised loss on a proxy learning task. The TTT++ method [37] introduces a feature alignment strategy based on online moment match- ing. 1.1. Challenges in Fully Test-Time UDA We recognize that most of the domain variations, such as changes in the visual scenes and image transformations or corruptions, are early layers of features in the semantic hierarchy [66]. They can be effectively captured and mod- eled by lower layers of the network model. From the per- spective of machine learning, early representations through the lower layer play an important role to capture the pos- 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. 3728 terior distribution of the underlying explanatory factors for the observed input [1]. For instance, in deep neural network models, the early layers of the network tend to respond to corners, edges, or colors. In contrast, deeper layers respond to more class-specific features [72]. In the corruption test- time adaptation scenario, the class-specific features are al- ways the same because the testing datasets are the corrup- tion of the training domain. However, the early layers of models can be failed due to corruption. Therefore, the central challenge in fully test-time UDA lies in how to learn useful early layer representations of the test samples without supervision. Motivated by this obser- vation, we propose to explore neurobiology-inspired Heb- bian learning for effective early-layer representation learn- ing and fully test-time adaptation. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation and the learning rules of the early front-end neural processing in neurobiol- ogy are unrelated [28]. The responses of neurons in bio- logical neural networks are tuned by local pre-synaptic and post-synaptic activity, along with global variables that mea- sure task performance, rather than the specific activity of other neurons [69]. Source Domain Target Domain Source Model Hebbian learning Oracle Figure 1. The feature map visualization after the first convolution layer obtained by different learning methods. 1.2. Hebbian Learning Hebbian learning aims to learn useful early layer rep- resentations without supervision based on local synaptic plasticity rules, which is able to generate early representa- tions that are as good as those learned by end-to-end super- vised training with back-propagation [28, 52]. Drastically different from the current error back-propagation methods which require pseudo-labels or loss functions from the top network layers, Hebbian learning is a pure feed-forward adaptation process and does not require feedback from the distant top network layers. The responses of neurons are tuned based on a local synapse-change procedure and ac- tivated by competitive lateral inhibition rules [28]. Dur- ing the learning process, the strength of synapses under- goes local changes that are proportional to the activity ofthe pre-synaptic cell and dependent on the activity of the post-synaptic cell. It also introduces local lateral inhibition between neurons within a layer, where the synapses of hid- den units with strong responses are pushed toward the pat- terns that drive them, while those with weaker responses are pushed away from these patterns. Existing literature has shown that early representa- tions learned by Hebbian learning are as well as back- propagation and even more robust in testing [28, 29, 52]. Figure 1 shows the feature maps learned by different meth- ods. The first row shows the original image. The second row shows the images in the target domain with significant image corruption. The third row shows the feature maps learned by the network model trained in the source domain for these target-domain images. The fourth row shows the feature maps learned by our Hebbian learning method. The last row (“oracle”) shows the feature maps learned with true labels. We can see that the unsupervised Hebbian learning is able to generate feature maps which are as good as those from supervised learning. 1.3. Our Major Idea In this work, we observe that Hebbian learning, although provides a new and effective approach for unsupervised learning of early layer representation of the image, when directly applied to the network model, is not able to achieve satisfactory performance in fully test-time adaptation. First, the original hard decision for competitive learning is not suitable for fully test-time adaptation. Second, the Hebbian learning does not have an effective mechanism to consider external feedback, especially the feedback from the top net- work layers. We observe that, biologically, the visual pro- cessing is realized through hierarchical models considering a bottom-up early representation learning for the sensory input, and a top-down feedback mechanism based on pre- dictive coding [15, 56]. Motivated by this, in this work, we propose to develop a new approach, called neuro-modulated Hebbian learn- ing (NHL) , for fully test-time adaptation. We first incor- porate a soft decision rule into the feed-forward Hebbian learning to improve its competitive learning. Second, we learn a neuro-modulator to capture feedback from exter- nal responses, which controls which type of feature is con- solidated and further processed to minimize the predictive error. During inference, the source model is adapted by the proposed NHL rule for each mini-batch of testing sam- ples during the inference process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods. 3729 1.4. Summary of Major Contributions To summarize, our major contributions include: (1) we identify that the major challenge in fully test-time adaptation lies in effective unsupervised learning of early layer representations, and explore neurobiology-inspired soft Hebbian learning for effective early layer representa- tion learning and fully test-time adaptation. (2) We develop a new neuro-modulated Hebbian learning method which combines unsupervised feed-forward Hebbian learning of early layer representation with a learned neuro-modulator to capture feedback from external responses. We analyze the optimal property of the proposed NHL algorithm based on free-energy principles [14, 15]. (3) We evaluate our pro- posed NHL method on benchmark datasets for fully test- time adaptation, demonstrating its significant performance improvement over existing methods.

Tian_Integrally_Pre-Trained_Transformer_Pyramid_Networks_CVPR_2023

Abstract In this paper, we present an integral pre-training frame- work based on masked image modeling (MIM). We advo- cate for pre-training the backbone and neck jointly so that the transfer gap between MIM and downstream recogni- tion tasks is minimal. We make two technical contributions. First , we unify the reconstruction and recognition necks by inserting a feature pyramid into the pre-training stage. Second , we complement mask image modeling (MIM) with masked feature modeling (MFM) that offers multi-stage su- pervision to the feature pyramid. The pre-trained mod- els, termed integrally pre-trained transformer pyramid net- works (iTPNs), serve as powerful foundation models for vi- sual recognition. In particular, the base/large-level iTPN achieves an 86.2% /87.8% top-1 accuracy on ImageNet-1K, a53.2% /55.6% box AP on COCO object detection with 1× training schedule using Mask-RCNN, and a 54.7% /57.7% mIoU on ADE20K semantic segmentation using UPerHead – all these results set new records. Our work inspires the community to work on unifying upstream pre-training and downstream fine-tuning tasks. Code is available at github.com/sunsmarterjie/iTPN.

1. Introduction Recent years have witnessed two major progresses in vi- sual recognition, namely, the vision transformer architec- ture [22] as network backbone and masked image mod- eling (MIM) [3, 28, 68] for visual pre-training. Combin- ing these two techniques yields a generalized pipeline that achieves state-of-the-arts in a wide range of visual recogni- tion tasks, including image classification, object detection, and instance/semantic segmentation. One of the key issues of the above pipeline is the transfer gap between upstream pre-training and downstream fine- *Corresponding author. 86.0 85.0 84.0 83.0MAEMAEMaskFeat iBoTSimMIMOurs iBoTdata2vecOursfine−tuningaccuracy % pre−training epochs87.0 86.5 86.0 85.5MAE MaskFeatiBoT SimMIMOurs iBoTdata2vecOurssupervised by image pixels CAE (DALLE)PeCo (CodeBook ) MVP (CLIP)86.0 85.0 84.0 83.0Ours (CLIP)Ours (CLIP) MVP (CLIP)88.0 87.0 86.0 85.0Ours/14 (CLIP)supervised by a pre-trained teacherOurs/16 (CLIP) CAE (DALLE)base size models large size models pre−training epochs400 800 1600 400 800 1600BEiT (DALLE)CAE (DALLE) BEiT (DALLE)PeCo (CodeBook )BEiT -v2(CLIP)BEiT -v2(CLIP) BEiT -v2(CLIP)BEiT -v2(CLIP)base -level IN-1K COCO (MR, 1×)ADE20K (UP) models cls. acc. det. AP seg. AP seg. mIoU iTPN-B 86.2 53.2 46.6 54.7 prev. best 85.5 [50] 50.0 [13] 44.0 [13] 53.0 [50] large -level IN-1K COCO (MR, 1×)ADE20K (UP) models cls. acc. det. AP seg. AP seg. mIoU iTPN-L 87.8 55.6 48.6 57.7 prev. best 87.3 [50] 54.5 [13] 47.6 [13] 56.7 [50] Figure 1. Top: on ImageNet-1K classification, iTPN shows sig- nificant advantages over prior methods, either only using pixel su- pervision (top) or leveraging knowledge from a pre-trained teacher (bottom, in the parentheses lies the name of teacher model). Bot- tom: iTPN surpasses previous best results in terms of recognition accuracy (%) on several important benchmarks. Legends – IN-1K: ImageNet-1K, MR: Mask R-CNN [30], UP: UPerHead [66]. tuning. From this point of view, we argue that downstream visual recognition, especially fine-scaled recognition ( e.g., detection and segmentation), requires hierarchical visual 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. 18610 features. However, most existing pre-training tasks ( e.g., BEiT [3] and MAE [28]) were built upon plain vision trans- formers. Even if hierarchical vision transformers have been used ( e.g., in SimMIM [68], ConvMAE [25], and Green- MIM [33]), the pre-training task only affects the backbone but leaves the neck ( e.g., a feature pyramid) un-trained. This brings extra risks to downstream fine-tuning as the opti- mization starts with a randomly initialized neck which is not guaranteed to cooperate with the pre-trained backbone. In this paper, we present an integral pre-training frame- work to alleviate the risk. We establish the baseline with HiViT [74], an MIM-friendly hierarchical vision trans- former, and equip it with a feature pyramid. To jointly op- timize the backbone (HiViT) and neck (feature pyramid), we make two-fold technical contributions. First , we unify the upstream and downstream necks by inserting a feature pyramid into the pre-training stage (for reconstruction) and reusing the weights in the fine-tuning stage (for recogni- tion). Second , to better pre-train the feature pyramid, we propose a new masked feature modeling (MFM) task that (i) computes intermediate targets by feeding the original image into a moving-averaged backbone, and (ii) uses the output of each pyramid stage to reconstruct the intermedi- ate targets. MFM is complementary to MIM and improves the accuracy of both reconstruction and recognition. MFM can also be adapted to absorb knowledge from a pre-trained teacher ( e.g., CLIP [52]) towards better performance. The obtained models are named integrally pre-trained pyramid transformer networks (iTPNs). We evaluate them on standard visual recognition benchmarks. As highlighted in Figure 1, the iTPN series report the best known down- stream recognition accuracy. On COCO and ADE20K , iTPN largely benefits from the pre-trained feature pyra- mid. For example, the base /large -level iTPN reports a 53.2%/55.6%box AP on COCO ( 1×schedule, Mask R-CNN) and a 54.7%/57.7%mIoU on ADE20K (UPer- Net), surpassing all existing methods by large margins. On ImageNet-1K , iTPN also shows significant advantages, im- plying that the backbone itself becomes stronger during the joint optimization with neck. For example, the base /large - level iTPN reports an 86.2%/87.8%top-1 classification accuracy, beating the previous best record by 0.7%/0.5%, which is not small as it seems in such a fierce competition. In diagnostic experiments, we show that iTPN enjoys both (i) a lower reconstruction error in MIM pre-training and (ii) a faster convergence speed in downstream fine-tuning – this validates that shrinking the transfer gap benefits both up- stream and downstream parts. Overall, the key contribution of this paper lies in the inte- gral pre-training framework that, beyond setting new state- of-the-arts, enlightens an important future research direc- tion – unifying upstream pre-training and downstream fine- tuning to shrink the transfer gap between them.

Tschannen_CLIPPO_Image-and-Language_Understanding_From_Pixels_Only_CVPR_2023

Abstract Multimodal models are becoming increasingly effective, in part due to unified components, such as the Transformer architecture. However, multimodal models still often consist of many task- and modality-specific pieces and training pro- cedures. For example, CLIP (Radford et al., 2021) trains in- dependent text and image towers via a contrastive loss. We explore an additional unification: the use of a pure pixel- based model to perform image, text, and multimodal tasks. Our model is trained with contrastive loss alone, so we call it CLIP-Pixels Only (CLIPPO). CLIPPO uses a single en- coder that processes both regular images and text rendered as images. CLIPPO performs image-based tasks such as re- trieval and zero-shot image classification almost as well as CLIP-style models, with half the number of parameters and no text-specific tower or embedding. When trained jointly via image-text contrastive learning and next-sentence con- trastive learning, CLIPPO can perform well on natural language understanding tasks, without any word-level loss (language modelling or masked language modelling), out- performing pixel-based prior work. Surprisingly, CLIPPO can obtain good accuracy in visual question answering, simply by rendering the question and image together. Fi- nally, we exploit the fact that CLIPPO does not require a tokenizer to show that it can achieve strong performance on multilingual multimodal retrieval without modifications.

1. Introduction In recent years, large-scale multimodal training of Transformer-based models has led to improvements in the state-of-the-art in different domains including vision [ 2,10, 74–76], language [ 6,11], and audio [ 5]. In particular, in computer vision and image-language understanding, a sin- gle large pretrained model can outperform task-specific ex- pert models [ 10,74,75]. However, large multimodal mod- els often use modality or dataset-specific encoders and de- coders, and accordingly lead to involved protocols. For example, such models frequently involve training different Code and pretrained models are available as part of bigvision [4] https://github.com/google-research/big_vision . ContrastiveVision TransformerText TransformerTOK A birthday pug wearing a party hat.• CLIP • CONVWORD EMB • CLIPPO • A birthday pug wearing a party hatCONVContrastiveTransformerFigure 1. CLIP [ 56] trains separate image and text encoders, each with a modality-specific preprocessing and embedding, on image/alt-text pairs with a contrastive objective. CLIPPO trains a pure pixel-based model with equivalent capabilities by rendering the alt-text as an image, encoding the resulting image pair using a shared vision encoder (in two separate forward passes), and ap- plying same training objective as CLIP. parts of the model in separate phases on their respective datasets, with dataset-specific preprocessing, or transferring different parts in a task-specific manner [ 75]. Such modality and task-specific components can lead to additional engi- neering complexity, and poses challenges when introducing new pretraining losses or downstream tasks. Developing a single end-to-end model that can process any modality, or combination of modalities, would be a valuable step for multimodal learning. Here, we focus on images and text. A number of key unifications have accelerated the progress of multimodal learning. First, the Transformer architecture has been shown to work as a universal back- bone, performing well on text [ 6,15], vision [ 16], au- dio [ 5,24,54], and other domains [ 7,34]. Second, many papers have explored mapping different modalities into a single shared embedding space to simplify the input/output interface [ 21,22,46,69], or develop a single interface to many tasks [ 31,37]. Third, alternative representations of modalities allow harnessing in one domain neural archi- tectures or training procedures designed for another do- main [ 28,49,54,60]. For example, [ 60] and [ 28,54] repre- sent text and audio, respectively, by rendering these modal- ities as images (via a spectogram in the case of audio). In this paper, we explore the use of a pure pixel-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. 11006 model for multimodal learning of text and images. Our model is a single Vision Transformer [ 16] that processes visual input, or text, or both together, all rendered as RGB images. The same model parameters are used for all modal- ities, including low-level feature processing; that is, there are no modality-specific initial convolutions, tokenization algorithms, or input embedding tables. We train our model using only a single task: contrastive learning, as popular- ized by CLIP [ 56] and ALIGN [ 32]. We therefore call our model CLIP-Pixels Only (CLIPPO). We find that CLIPPO performs similarly to CLIP-style models (within 1-2%) on the main tasks CLIP was de- signed for—image classification and text/image retrieval— despite not having modality-specific towers. Surpris- ingly, CLIPPO can perform complex language understand- ing tasks to a decent level without any left-to-right lan- guage modelling, masked language modelling, or explicit word-level losses. In particular, on the GLUE benchmark [73] CLIPPO outperforms classic NLP baselines, such as ELMO+BiLSTM+attention, outperforms prior pixel-based masked language models [ 60], and approaches the score of BERT [ 15]. Interestingly, CLIPPO obtains good perfor- mance on VQA when simply rendering the image and text together, despite never having been pretrained on such data. Pixel-based models have an immediate advantage over regular language models because they do not require pre- determining the vocabulary/tokenizer and navigating the corresponding intricate trade-offs; consequently, we ob- serve improved performance on multilingual retrieval com- pared to an equivalent model that uses a classical tokenizer.

Sun_RefTeacher_A_Strong_Baseline_for_Semi-Supervised_Referring_Expression_Comprehension_CVPR_2023

Abstract Referring expression comprehension (REC) often re- quires a large number of instance-level annotations for fully supervised learning, which are laborious and expensive. In this paper, we present the first attempt of semi-supervised learning for REC and propose a strong baseline method called RefTeacher. Inspired by the recent progress in com- puter vision, RefTeacher adopts a teacher-student learning paradigm, where the teacher REC network predicts pseudo- labels for optimizing the student one. This paradigm allows REC models to exploit massive unlabeled data based on a small fraction of labeled. In particular, we also identify two key challenges in semi-supervised REC, namely, sparse supervision signals and worse pseudo-label noise. To ad- dress these issues, we equip RefTeacher with two novel de- signs called Attention-based Imitation Learning (AIL) and Adaptive Pseudo-label Weighting (APW). AIL can help the student network imitate the recognition behaviors of the teacher, thereby obtaining sufficient supervision signals. APW can help the model adaptively adjust the contributions of pseudo-labels with varying qualities, thus avoiding con- firmation bias. To validate RefTeacher, we conduct exten- sive experiments on three REC benchmark datasets. Exper- imental results show that RefTeacher obtains obvious gains over the fully supervised methods. More importantly, using only 10% labeled data, our approach allows the model to achieve near 100% fully supervised performance, e.g., only -2.78% on RefCOCO. Project: https://refteacher.github.io/.

1. Introduction Referring Expression Comprehension (REC) [33–35,45, 49, 53, 60, 62], also called Visual grounding [26, 51, 52] orPhrase localization [20, 39], aims to locate the target *Corresponding Author Figure 1. Statistics of the pseudo-label quality in semi- supervised REC with different percentages of labeled infor- mation. The REC model only predicts one pseudo-box for each image-text pair and cannot apply filtering, so the pseudo-labels are usually noisy and low-quality during training. objects in an image referred by a given natural language expression. Compared to conventional object detection tasks [4,9–11,23,24,28,40–42], REC is not limited to a fix set of categories and can be generalized to open-vocabulary recognition [31, 60]. However, as a detection task, REC also requires a large number of instance-level annotations for training, which poses a huge obstacle to its practical ap- plications. To address this issue, one feasible solution is semi- supervised learning (SSL), which has been well studied on various computer vision tasks [1–3, 8, 16, 29, 43, 44, 47, 59] but not yet exploited in REC. In particular, recent advances in semi-supervised object detection (SSOD) [16, 29, 44, 47, 59] has yielded notable progress in practical applications. These SSOD methods apply a training framework consisted of two detection networks with the same configurations, act- ing as teacher and student, respectively. The teacher net- work is in charge of generating pseudo-labels to optimize 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. 19144 the student during training, which can exploit massive unla- beled data based on a small amount of labeled information. With the help of this effective training paradigm, the latest SSOD method [37] can even achieve fully supervised per- formance with only 40% and 25% labeled information on MSCOCO [25] and PASCAL VOC [7], respectively. However, directly transferring this successful paradigm to REC still suffers from two main challenges due to task gaps. The first one is the extremely sparse supervision signals. In contrast to object detection, REC only ground one instance for each text-image pair. This prediction pat- tern makes the REC model receive much fewer pseudo- supervisions than SSOD during teacher-student learning, i.e., only one bounding box without class pseudo-labels. For instance, Compared with SSOD that has 6-15 high- quality pseudo-boxes for each image [30, 37], semi-REC only has 0.5 box on average at the infant training stages as shown in Fig. 1. The sparse supervision signals also lead to worse pseudo-label quality, which is the second challenge. SSOD methods [16,29,44,47,59] can apply NMS [10] and high-threshold filtering to discard the vast majority of noisy pseudo-labels, thereby avoiding the error accumulation is- sue [29, 37, 44] in SSL. But in REC, a strong filtering is not feasible due to the already sparse pseudo-label information. This results in that most pseudo-labels of REC are of much lower-quality. Based on this observations, we propose the first semi- supervised approach for REC called RefTeacher with two novel designs, namely Attention-based Imitation Learn- ing(AIL) and Adaptive Pseudo-label Weighting (APW). In principle, RefTeacher also adopts a teacher-student frame- work, where the teacher predicts the pseudo bounding boxes for the student according to the given expressions. Follow- ing the latest SSOD [29,30,37,47], we also use EMA to up- date the gradients of the teacher network from the student, and introduce data augmentation and burn-in strategies to improve SSL. To enrich the supervision signals, the pro- posed AIL helps the student imitate the attention behaviors of the teacher, thereby improving the knowledge transfer- ring. APW is further used to reduce the impact of noisy pseudo-labels, which is achieved via adaptively weighting label information and the corresponding gradient updates. To validate RefTeacher, we apply RefTeacher to two representative REC models, i.e. RealGIN [60] and TransVG [5], and conduct extensive experiments on three REC benchmark datasets, namely RefCOCO [54], Ref- COCO+ [54] and RefCOCOg [38]. Experimental results show that RefTeacher can greatly exceed the supervised baselines, e.g. +18.8% gains on 10% RefCOCO. More im- portantly, using only 10% labeled data, RefTeacher can help RealGIN achieve near 100% fully supervised performance. Overall, the contributions of this paper are three-fold: • We present the first attempt of semi-supervised learn-ing for REC with a strong baseline method called RefTeacher. • We identify two challenges of semi-supervised REC, i.e. sparse supervision signals and worse pseudo-label noise, and address them with two novel designs, namely Attention-based Imitation Learning (AIL) and Adaptive Pseudo-label Weighting (APW). • RefTeacher achieves significant performance gains on RefCOCO, RefCOCO+, and RefCOCOg datasets over the fully supervised methods.

Trosten_Hubs_and_Hyperspheres_Reducing_Hubness_and_Improving_Transductive_Few-Shot_Learning_CVPR_2023

Abstract Distance-based classification is frequently used in trans- ductive few-shot learning (FSL). However, due to the high- dimensionality of image representations, FSL classifiers are prone to suffer from the hubness problem, where a few points (hubs) occur frequently in multiple nearest neighbour lists of other points. Hubness negatively impacts distance-based classification when hubs from one class appear often among the nearest neighbors of points from another class, degrading the classifier’s performance. To address the hubness prob- lem in FSL, we first prove that hubness can be eliminated by distributing representations uniformly on the hypersphere. We then propose two new approaches to embed representa- tions on the hypersphere, which we prove optimize a tradeoff between uniformity and local similarity preservation – reduc- ing hubness while retaining class structure. Our experiments show that the proposed methods reduce hubness, and signifi- cantly improves transductive FSL accuracy for a wide range of classifiers1.

1. Introduction While supervised deep learning has made a significant impact in areas where large amounts of labeled data are available [ 6,11], few-shot learning (FSL) has emerged as a promising alternative when labeled data is limited [ 3,12, 14,16,21,26,28,31,33,39,40]. FSL aims to design classifiers that can discriminate between novel classes based on a few labeled instances, significantly reducing the cost of the labeling procedure. In transductive FSL, one assumes access to the entire *Equal contributions. †UiT Machine Learning group ( machine-learning.uit.no ) and Visual Intelligence Centre ( visual-intelligence.no ). ‡Norwegian Computing Center. §Department of Computer Science, University of Copenhagen. ¶Pioneer Centre for AI ( aicentre.dk ). 1Code available at https://github.com/uitml/noHub . 0.25 0.50 0.75 1.00 1.25 1.50 Hubness55606570Accuracy NoneL2 (ArXiv'19)ZN (ICCV'21) TCPR (NeurIPS'22)EASE (CVPR'22)noHub (Ours)noHub-S (Ours)Figure 1. Few-shot accuracy increases when hubness decreases. The figure shows the 1-shot accuracy when classifying different embeddings with SimpleShot [33] on mini-ImageNet [29]. query set during evaluation. This allows transductive FSL classifiers to learn representations from a larger number of samples, resulting in better performing classifiers. However, many of these methods base their predictions on distances to prototypes for the novel classes [ 3,16,21,28,39,40]. This makes these methods susceptible to the hubness prob- lem [ 10,22,24,25], where certain exemplar points (hubs) appear among the nearest neighbours of many other points. If a support sample is a hub, many query samples will be assigned to it regardless of their true label, resulting in low accuracy. If more training data is available, this effect can be reduced by increasing the number of labeled samples in the classification rule – but this is impossible in FSL. Several approaches have recently been proposed to embed samples in a space where the FSL classifier’s performance is improved [ 4,5,7,17,33,35,39]. However, only one of these directly addresses the hubness problem. Fei et al. [7] show that embedding representations on a hypersphere with zero mean reduces hubness. They advocate the use of Z- score normalization (ZN) along the feature axis of each representation, and show empirically that ZN can reduce hubness in FSL. However, ZN does not guarantee a data mean of zero, meaning that hubness can still occur after ZN. 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. 7527 In this paper we propose a principled approach to em- bed representations in FSL, which both reduces hubness and improves classification performance. First, we prove that hubness can be eliminated by embedding representa- tions uniformly on the hypersphere. However, distributing representations uniformly on the hypersphere without any additional constraints will likely break the class structure which is present in the representation space – hurting the performance of the downstream classifier. Thus, in order to both reduce hubness and preserve the class structure in the representation space, we propose two new embedding methods for FSL. Our methods, U niformHyperspherical Structure-preserving Em beddings (noHub) and noHub with Support labels (noHub-S), leverage a decomposition of the Kullback-Leibler divergence between representation and em- bedding similarities, to optimize a tradeoff between Local Similarity Preservation (LSP) and uniformity on the hyper- sphere. The latter method, noHub-S, also leverages label information from the support samples to further increase the class separability in the embedding space. Figure 1 illustrates the correspondence between hubness and accuracy in FSL. Our methods have both the least hub- ness andhighest accuracy among several recent embedding techniques for FSL. Our contributions are summarized as follows. •We prove that the uniform distribution on the hyper- sphere has zero hubness and that embedding points uni- formly on the hypersphere thus alleviates the hubness problem in distance-based classification for transduc- tive FSL. •We propose noHub and noHub-S to embed representa- tions on the hypersphere, and prove that these methods optimize a tradeoff between LSP and uniformity. The resulting embeddings are therefore approximately uni- form, while simultaneously preserving the class struc- ture in the embedding space. •Extensive experimental results demonstrate that noHub and noHub-S outperform current state-of-the-art em- bedding approaches, boosting the performance of a wide range of transductive FSL classifiers, for multiple datasets and feature extractors.

Tan_Hierarchical_Semantic_Correspondence_Networks_for_Video_Paragraph_Grounding_CVPR_2023

Abstract Video Paragraph Grounding (VPG) is an essential yet challenging task in vision-language understanding, which aims to jointly localize multiple events from an untrimmed video with a paragraph query description. One of the crit- ical challenges in addressing this problem is to compre- hend the complex semantic relations between visual and textual modalities. Previous methods focus on modeling the contextual information between the video and text from a single-level perspective (i.e., the sentence level), ignor- ing rich visual-textual correspondence relations at different semantic levels, e.g., the video-word and video-paragraph correspondence. To this end, we propose a novel Hierar- chical Semantic Correspondence Network (HSCNet), which explores multi-level visual-textual correspondence by learn- ing hierarchical semantic alignment and utilizes dense su- pervision by grounding diverse levels of queries. Specifi- cally, we develop a hierarchical encoder that encodes the multi-modal inputs into semantics-aligned representations at different levels. To exploit the hierarchical semantic cor- respondence learned in the encoder for multi-level supervi- sion, we further design a hierarchical decoder that progres- sively performs finer grounding for lower-level queries con- ditioned on higher-level semantics. Extensive experiments demonstrate the effectiveness of HSCNet and our method significantly outstrips the state-of-the-arts on two challeng- ing benchmarks, i.e., ActivityNet-Captions and TACoS.

1. Introduction As a fundamental problem that bridges the gap between computer vision and natural language processing, Video Language Grounding (VLG) aiming to localize the video *Corresponding author (a) Bottom -Up Hierarchy of Linguistic Descriptions(b) Coarse- to-Fine Hierarchy of Temporal Counterparts Then hegets back upand grabs …again. grabsThis baby …falls down .Then hegets back upand grabs …again .Next he…hitting itreally far. Words Sentences ParagraphFigure 1. (a) The bottom-up hierarchy for linguistic semantics is composed of textual words, sentences, and the paragraph. (b) The coarse-to-fine hierarchy for temporal granularity consists of visual counterparts for each level of linguistic query. segments corresponding to given natural language queries, has been drawing increasing attention from the community in these years. Early works in the field of VLG mainly focused on addressing Video Sentence Grounding (VSG) [1, 8], whose goal is to localize the most relevant moment with a single sentence query. Recently, Video Paragraph Grounding (VPG) is introduced in [2]. It requires to jointly localize multiple events via a paragraph query consisting of several temporally ordered sentences. Rather than ground- ing each event independently, VPG needs to further exploit the contextual information between the video and the tex- tual paragraph, which helps to avoid ambiguity and achieve more precise temporal localization of video events. 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. 18973 Previous VPG works [2,12,26] commonly explore corre- lations of events by modeling the video-text correspondence from a single semantic level ( i.e., the sentence level). How- ever, they neglect the rich visual-textual correspondence at other semantic levels, such as the word level and paragraph level, which can also provide some useful information for grounding events in the video. Considering the grounding of “The man stops and hits the ball far away”, the seman- tic relations between video content and the word “hits” is crucial in determining the end time of the event. Besides, when we consider the paragraph as a whole, then ground- ing the holistic paragraph in the video first is beneficial to suppress the irrelevant events or backgrounds, which eases the further grounding of sentences. To be more general, we observe that there naturally exist two perspectives of hierarchical semantic structures in tack- ling VPG, which is intuitively illustrated in Figure 1. On the language side, Figure 1 (a) shows that the semantics of paragraph query can be divided into an inherent three-level hierarchy consisting of words, sentences, and the holistic paragraph in a bottom-up organization. On the video side, Figure 1 (b) shows that the temporal counterparts of dif- ferent levels of queries also form a three-level granular- ity hierarchy with temporally nested dependencies from the top down. By relating the video content to different lev- els of query semantics for multi-level query grounding, the model is enforced to capture more complex relations be- tween events by reasoning about their interconnections at different semantic granularities, and exploit richer temporal clues to facilitate the grounding of events in the video. Motivated by the above observations, we propose a novel framework termed as Hierarchical Semantic Correspon- dence Network (HSCNet) for VPG. Our HSCNet is de- signed as a multi-level encoder-decoder architecture in or- der to leverage hierarchical semantic information from the two perspectives. On the one hand , we learn the hierar- chical visual-textual semantic correspondence by gradually aligning the visual and textual semantics into different lev- els of common spaces from the bottom up. Concretely, we construct a hierarchical multi-modal encoder on top of the linguistic semantic hierarchy. It comprises three semantic levels of visual-textual encoders. Each encoder receives the semantic representation from a lower level and continues to establish visual-textual correspondence at a higher level via iterative multi-modal interactions. On the other hand , we utilize richer cross-level contexts and denser supervi- sion by progressively grounding multiple levels of queries from coarse to fine. Specifically, we construct a hierarchical progressive decoder on top of the temporal granularity hier- archy, which also comprises three levels of decoders. The lower-level queries are grounded by finer temporal bound- aries conditioned on contextual knowledge from higher- level queries, which eases the learning of multi-level lo-calization that provides diverse temporal clues to promote fine-grained video paragraph grounding. We evaluate the proposed HSCNet on two challenging benchmarks, i.e., ActivityNet-Captions and TACoS. Ex- tensive ablation studies validate the effectiveness of the method. Our contributions can be summarized as follows: • We investigate and propose a novel hierarchical model- ing framework for Video Paragraph Grounding (VPG). To the best of our knowledge, it’s the first time in the problem of VPG that hierarchical visual-textual se- mantic correspondence is explored and multiple levels of linguistic queries can be grounded. • We design a novel encoder-decoder architecture to learn multi-level visual-textual correspondence by hi- erarchical semantic alignment and progressively per- form finer grounding for lower-level queries. • Experiments demonstrate that our proposed HSCNet achieves new state-of-the-art results on the challeng- ing ActivityNet-Captions and TACoS benchmarks, re- markably surpassing the previous approaches.

Tien_Revisiting_Reverse_Distillation_for_Anomaly_Detection_CVPR_2023

Abstract Anomaly detection is an important application in large- scale industrial manufacturing. Recent methods for this task have demonstrated excellent accuracy but come with a latency trade-off. Memory based approaches with domi- nant performances like PatchCore or Coupled-hypersphere- based Feature Adaptation (CFA) require an external mem- ory bank, which significantly lengthens the execution time. Another approach that employs Reversed Distillation (RD) can perform well while maintaining low latency. In this paper, we revisit this idea to improve its performance, es- tablishing a new state-of-the-art benchmark on the chal- lenging MVTec dataset for both anomaly detection and localization. The proposed method, called RD++ , runs six times faster than PatchCore, and two times faster than CFA but introduces a negligible latency compared to RD. We also experiment on the BTAD and Retinal OCT datasets to demonstrate our method’s generalizabil- ity and conduct important ablation experiments to provide insights into its configurations. Source code will be avail- able at https://github.com/tientrandinh/ Revisiting-Reverse-Distillation .

1. Introduction Detecting anomalies is a crucial aspect of computer vi- sion with numerous applications, such as product quality control [4], and healthcare monitor system [21]. Unsuper- vised anomaly detection can help to reduce the cost of col- lecting abnormal samples. This task identifies and local- izes anomalous regions in images without defect annota- tions during training. Instead, a set of abnormal-free sam- ples is utilized. Early approaches rely on generative adver- sarial models to extract meaningful latent representations on normal samples [25, 29, 31]. However, these approaches are computationally expensive, resulting in higher latency and potential performance limitations on unseen data. Other approaches leverage the pre-trained Convolutional Neural Networks (CNNs) [19] backbones to extract comprehensive visual features for anomaly detection systems [3, 23]. Figure 1. Comparisons of different anomaly detection methods in terms of AUROC sample (vertical axis), inference time (horizontal axis), and memory footprint (circle radius). Our RD++ achieves thehighest AUROC sample metric for anomaly detection while being 6×faster than PatchCore, 4×faster than CSFlow, and 2× faster than CFA. Additionally, RD++ requires only 4GB of mem- ory for inference, making it one of the least memory-use methods. The test environment was conducted on a computer with Intel(R) Xeon(R) 2.00GHz (4 cores) and Tesla T4 GPU (15GB VRAM). Alternative approach employs knowledge distillation (KD) [14] based frameworks. For example, Salehi et al. [30] set up a teacher-student network pair, and knowledge is transferred from teacher to student. During training, the student is learned from only normal samples. Thus, it is expected to learn the distribution of normal samples, sub- sequently generating the out-of-distribution representations when inference with anomalous query [30]. However, Deng and Li [8] point out that the statement is not always accu- rate due to limitations in the similarity of network architec- tures and the same data flows in the teacher-student model. To overcome these limitations, they propose a reverse flow, called Reverse Distillation (RD), in which the teacher out- put is fed to the student through a One-class Bottleneck module (OCBE). The reverse distillation approach achieves competitive performance and also maintains low latency. Recent research in anomaly detection, such as Patch- Core [27], and CFA [20], have achieved state-of-the-art per- formance in detecting and localizing anomalies. However, 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. 24511 these methods are based on the memory bank framework, which leads to significant latency and makes them challeng- ing to apply in practical scenarios. Our key question: How can we develop a method that achieves high accuracy and fast inference for real-world applications? This paper answers the question from the perspective of reverse distillation. We identify the limitations of the RD approach by examining feature compactness and anoma- lous signal suppression. We argue that relying solely on the distillation task and an OCBE module is insufficient for providing a compact representation to the student. Further- more, we do not observe an explicit mechanism to discard anomalous patterns using the OCBE block as the authors claim. To address these concerns, we incorporate RD with multi-task learning to propose RD++, which demonstrates a favorable gain in performance. The contributions of the proposed RD++ are highlighted as follows: • We propose RD++ to tackle two tasks. First, fea- ture compactness task: by presenting a self-supervised optimal transport method. Second, anomalous sig- nal suppression task: by simulating pseudo-abnormal samples with simplex noise and minimizing the recon- struction loss. • We conduct extensive experiments on several pub- lic datasets in different domains, including MVTec, BTAD, and Retinal OCT. Results show that our ap- proach achieves state-of-the-art performance on detec- tion and localization, demonstrating strong general- ization capabilities across domains. Furthermore, our method’s real-time capability is at least twice as fast as its latest counterparts (see Fig. 1), making it a promis- ing method for practical applications.

Tripathi_3D_Human_Pose_Estimation_via_Intuitive_Physics_CVPR_2023

Abstract Estimating 3D humans from images often produces im- plausible bodies that lean, float, or penetrate the floor. Such methods ignore the fact that bodies are typically supported by the scene. A physics engine can be used to enforce phys- ical plausibility, but these are not differentiable, rely on unrealistic proxy bodies, and are difficult to integrate into ex- isting optimization and learning frameworks. In contrast, we exploit novel intuitive-physics (IP) terms that can be inferred from a 3D SMPL body interacting with the scene. Inspired by biomechanics, we infer the pressure heatmap on the body, theCenter of Pressure (CoP) from the heatmap, and the SMPL body’s Center of Mass (CoM ). With these, we develop IPMAN , to estimate a 3D body from a color image in a “sta- ble” configuration by encouraging plausible floor contact and overlapping CoP andCoM . Our IP terms are intuitive, easy to implement, fast to compute, differentiable, and can be integrated into existing optimization and regression methods. We evaluate IPMAN on standard datasets and MoYo, a new dataset with synchronized multi-view images, ground-truth 3D bodies with complex poses, body-floor contact, CoM and pressure. IPMAN produces more plausible results than the state of the art, improving accuracy for static poses, while not hurting dynamic ones. Code and data are available for research at https://ipman.is.tue.mpg.de . *This work was mostly performed at MPI-IS.

1. Introduction To understand humans and their actions, computers need automatic methods to reconstruct the body in 3D. Typi- cally, the problem entails estimating the 3D human pose and shape ( HPS) from one or more color images. State-of- the-art ( SOTA ) methods [46, 51, 75, 102] have made rapid progress, estimating 3D humans that align well with image features in the camera view. Unfortunately, the camera view can be deceiving. When viewed from other directions, or when placed in a 3D scene, the estimated bodies are often physically implausible: they lean, hover, or penetrate the ground (see Fig. 1 top). This is because most SOTA methods reason about humans in isolation ; they ignore that people move in a scene, interact with it, and receive physical sup- port by contacting it. This is a deal-breaker for inherently 3D applications, such as biomechanics, augmented/virtual reality (AR/VR) and the “metaverse”; these need humans to be reconstructed faithfully and physically plausibly with re- spect to the scene. For this, we need a method that estimates the 3D human on a ground plane from a color image in a configuration that is physically “stable” . This is naturally related to reasoning about physics and support. There exist many physics simulators [10,30,60] for games, movies, or industrial simulations, and using these for plausible HPS estimation is increasingly popular [66,74,96]. However, existing simulators come with two significant 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. 4713 problems: (1) They are typically non-differentiable black boxes , making them incompatible with existing optimiza- tion and learning frameworks. Consequently, most meth- ods [64, 95, 96] use them with reinforcement learning to evaluate whether a certain input has the desired outcome, but with no ability to reason about how changing inputs af- fects the outputs. (2) They rely on an unrealistic proxy body model for computational efficiency; bodies are represented as groups of rigid 3D shape primitives. Such proxy models are crude approximations of human bodies, which, in reality, are much more complex and deform non-rigidly when they move and interact. Moreover, proxies need a priori known body dimensions that are kept fixed during simulation. Also, these proxies differ significantly from the 3D body mod- els [41, 54, 92] used by SOTA HPS methods. Thus, current physics simulators are too limited for use in HPS. What we need, instead, is a solution that is fully differen- tiable, uses a realistic body model, and seamlessly integrates physical reasoning into HPS methods (both optimization- and regression-based). To this end, instead of using full physics simulation, we introduce novel intuitive-physics (IP) terms that are simple, differentiable, and compatible with a body model like SMPL [54]. Specifically, we define terms that exploit an inferred pressure heatmap of the body on the ground plane, the Center of Pressure (CoP) that arises from the heatmap, and the SMPL body’s Center of Mass (CoM ) projected on the floor; see Fig. 2 for a visualization. Intu- itively, bodies whose CoM lie close to their CoP are more stable than ones with a CoP that is further away (see Fig. 5); the former suggests a static pose ,e.g.standing or holding a yoga pose, while the latter a dynamic pose , e.g., walking. We use these intuitive-physics terms in two ways. First, we incorporate them in an objective function that extends SMPLify-XMC [59] to optimize for body poses that are stable. We also incorporate the same terms in the training loss for an HPS regressor, called IPMAN (Intuitive-Physics- based huMAN). In both formulations, the intuitive-physics terms encourage estimates of body shape and pose that have sufficient ground contact, while penalizing interpenetration and encouraging an overlap of the CoP and CoM. Our intuitive-physics formulation is inspired by work in biomechanics [32, 33, 61], which characterizes the stability of humans in terms of relative positions between the CoP, theCoM , and the Base of Support (BoS). The BoS is de- fined as the convex hull of all contact regions on the floor (Fig. 2). Following past work [6,71,74], we use the “inverted pendulum” model [85, 86] for body balance; this considers poses as stable if the gravity-projected CoM onto the floor lies inside the BoS. Similar ideas are explored by Scott et al. [71] but they focus on predicting a foot pressure heatmap from 2D or 3D body joints. We go significantly further to exploit stability in training an HPS regressor. This requires two technical novelties. CoM BoS CoP HIGH LOW Figure 2. (1) A SMPL mesh sitting. (2) The inferred pressure map on the ground (color-coded heatmap), CoP (green), CoM (pink), and Base of Support ( BoS, yellow polygon). (3) Segmentation of SMPL intoNP= 10 parts, used for computing CoM ; see Sec. 3.2. The first involves computing CoM . To this end, we uni- formly sample points on SMPL ’s surface, and calculate each body part’s volume. Then, we compute CoM as the average of all uniformly sampled points weighted by the correspond- ing part volumes. We denote this as pCoM , standing for “part-weighted CoM ”. Importantly, pCoM takes into account SMPL ’s shape, pose, and all blend shapes, while it is also computationally efficient and differentiable. The second involves estimating CoP directly from the image, without access to a pressure sensor. Our key insight is that the soft tissues of human bodies deform under pressure, e.g., the buttocks deform when sitting. However, SMPL does not model this deformation; it penetrates the ground instead of deforming. We use the penetration depth as a proxy for pressure [68]; deeper penetration means higher pressure. With this, we estimate a pressure field on SMPL ’s mesh and compute the CoP as the pressure-weighted average of the surface points. Again this is differentiable. For evaluation, we use a standard HPS benchmark (Human3.6M [37]), but also the RICH [35] dataset. How- ever, these datasets have limited interactions with the floor. We thus capture a novel dataset, MoYo , of challenging yoga poses, with synchronized multi-view video, ground-truth SMPL-X [63] meshes, pressure sensor measurements, and body CoM .IPMAN , in both of its forms, and across all datasets, produces more accurate and stable 3D bodies than the state of the art. Importantly, we find that IPMAN im- proves accuracy for static poses, while not hurting dynamic ones. This makes IPMAN applicable to everyday motions. To summarize: (1) We develop IPMAN , the first HPS method that integrates intuitive physics. (2) We infer biome- chanical properties such as CoM ,CoP and body pressure. (3) We define novel intuitive-physics terms that can be easily integrated into HPS methods. (4) We create MoYo , a dataset that uniquely has complex poses, multi-view video, and ground-truth bodies, pressure, and CoM . (5) We show that our IP terms improve HPS accuracy and physical plausibility. (6) Data and code are available for research.

Tarchoun_Jedi_Entropy-Based_Localization_and_Removal_of_Adversarial_Patches_CVPR_2023

Abstract Real-world adversarial physical patches were shown to be successful in compromising state-of-the-art models in a variety of computer vision applications. Existing defenses that are based on either input gradient or features analy- sis have been compromised by recent GAN-based attacks that generate naturalistic patches. In this paper, we pro- pose Jedi, a new defense against adversarial patches that is resilient to realistic patch attacks. Jedi tackles the patch localization problem from an information theory perspec- tive; leverages two new ideas: (1) it improves the identi- fication of potential patch regions using entropy analysis : we show that the entropy of adversarial patches is high, even in naturalistic patches; and (2) it improves the local- ization of adversarial patches, using an autoencoder that is able to complete patch regions from high entropy ker- nels. Jedi achieves high-precision adversarial patch local- ization, which we show is critical to successfully repair the images. Since Jedi relies on an input entropy analysis, it is model-agnostic , and can be applied on pre-trained off-the- shelf models without changes to the training or inference of the protected models. Jedi detects on average 90% of ad- versarial patches across different benchmarks and recovers up to 94% of successful patch attacks (Compared to 75% and 65% for LGS and Jujutsu, respectively).

1. Introduction Deep neural networks (DNNs) are vulnerable to adver- sarial attacks [20] where an adversary adds carefully crafted imperceptible perturbations to an input (e.g., by lp-norm bounded noise magnitude), forcing the models to misclas- sify. Several adversarial noise generation methods havebeen proposed [4, 10], often as part of a cat and mouse game where new defenses emerge [1, 18] only to be shown vulnerable to new adaptive attacks. Under real-world con- ditions, an attacker creates a physical patch (thus, spatially constrained) that contains an adversarial pattern. Such a patch can be placed as a sticker on traffic signs [9], worn as part of an item of clothing [13, 21], or introduced us- ing a display monitor [12], providing a practical approach for attackers to carry out adversarial attacks. First proposed by Brown et al. [3], these patches are different from tradi- tional adversarial attacks in two primary ways: (1) they oc- cupy a constrained space within an image; and (2) they may not be noise budget-constrained within the patch. Several adversarial patch generation methods have been demon- strated [13, 14, 16, 21], many of which showcasing real-life implementation, making them an ongoing threat to visual ML systems. Several approaches aiming to detect adversarial patches and defuse their impact have been proposed [5, 11, 15, 17, 22, 24, 25]. One category of defenses attempts to locate patches by detecting anomalies caused by the presence of the patch. These anomalies can be identified in the input pixel data such as in the case of Localized Gradient Smooth- ing [17] where the patch is located using high pixel gradi- ent values. Alternatively, they can be identified in the fea- ture space where the adversarial patch can create irregular saliency maps with regards to its targeted class that can be exploited by defenses such as in Digital Watermarking [11] and Jujutsu [5]. These defenses have two primary limita- tions: (1) They are only moderately successful against base- line attacks enabling recovery from many attacks (e.g., 75% for LGS and 65% for Jujutsu); and (2) they are vulnera- ble to adaptive attacks that generate naturalistic adversarial patches that are meant to use patterns similar to natural im- ages. Hu et al. [13] train a GAN to generate naturalistic 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. 4087 Successful attack: Undetected personDetection of entropypeaksFiltering + Auto-encoder for patch localization Inpainting the localizedPatch => recovered detectionFigure 1. Illustration on a patch attack against Yolo. Jedi surgically localizes the patch via entropy analysis and recovers the image. patches, matching the visual properties of normal images, and show that they are able to bypass defenses that are based on either input or featu

Sun_Correspondence_Transformers_With_Asymmetric_Feature_Learning_and_Matching_Flow_Super-Resolution_CVPR_2023

Abstract This paper solves the problem of learning dense visual correspondences between different object instances of the same category with only sparse annotations. We decompose this pixel-level semantic matching problem into two easier ones: (i) First, local feature descriptors of source and tar- get images need to be mapped into shared semantic spaces to get coarse matching flows. (ii) Second, matching flows in low resolution should be refined to generate accurate point-to-point matching results. We propose asymmetric feature learning and matching flow super-resolution based †: Corresponding Authorson vision transformers to solve the above problems. The asymmetric feature learning module exploits a biased cross- attention mechanism to encode token features of source im- ages with their target counterparts. Then matching flow in low resolutions is enhanced by a super-resolution network to get accurate correspondences. Our pipeline is built upon vision transformers and can be trained in an end-to-end manner. Extensive experimental results on several popular benchmarks, such as PF-PASCAL, PF-WILLOW, and SPair- 71K, demonstrate that the proposed method can catch sub- tle semantic differences in pixels efficiently. Code is avail- able on https://github.com/YXSUNMADMAX/ACTR. 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. 17787

1. Introduction Robust semantic matching methods aim to build dense visual correspondences between different objects or scenes from the same category regardless of the large variations in appearances and layouts. These algorithms have been widely exploited in various computer vision tasks, such as object recognition [12,48], cosegmentation [1,40], few-shot learning [20, 29], image editing [15, 18] and etc. Different from classical dense matching tasks, such as stereo match- ing [6,41] and image registration [2,19], semantic matching aims to find the visual consistency in image pairs with large intra-class appearance and layout variations. State-of-the-art methods, such as Proposal Flow [16], NCNet [37], Hyperpixel Flow [34], CATs [8] and etc, typ- ically extract features from backbones to measure point-to- point similarity in 4-D correlation tensors, and then refine these 4-D tensors to enforce neighborhood matching con- sistency. Despite that these algorithms have achieved im- pressive results, there are still two key issues that haven’t been discussed thoroughly. First, how to learn feature rep- resentations appropriate for semantic correspondence? Sec- ond, how to enforce neighborhood consensus when the 4- D correlation tensors that are in high resolution? For the first problem, Hyperpixel Flow [34] selects feature maps in convolutional neural networks with Beam search [32], and MMNet [49] aggregates feature maps at different resolu- tions in a top-down manner to get feature maps in high res- olution. For the second problem, V AT [20] firstly reduces the resolutions of 4-D correlation tensors with a 4-D con- volution operation and then conducts shifted window atten- tion [31] to further reduce the computational cost. Although these exploratory works brought a lot of new ideas, ques- tions listed above still deserve to be discussed seriously. In this paper, we aim to answer the above questions and propose a novel pipeline for semantic correspondence. Our pipeline consists of feature extraction based on pre-trained feature backbone [5, 17, 50], asymmetric feature learning, and matching flow super-resolution. Different from state- of-the-art methods [8, 34, 49] which directly calculate 4-D matching scores with refined generic backbone features, our core idea focuses on finding a shared semantic space where local feature descriptors of images can be aligned with their to-match counterpart. The asymmetric feature learn- ing module reconstructs source image features with target image features to reduce domain discrepancy between the two thus avoiding reconstructing source and target image features synchronously as in [9, 28]. Meanwhile, to high- light important image regions in target images, we also use source images to identify discriminative parts of foreground objects. In this way, a specific feature space is found for ev- ery image pair to conduct semantic matching. To avoid huge computational cost during neighborhood consensus enhancement, we map the matching informationhidden in 4-D correlation matrices to 2-D matching flow maps through the soft argmax [25]. By reducing the di- mension of the optimization goal from 4-D to 2-D, com- putational cost is alleviated drastically. To achieve pixel- level correspondence, we conduct matching flow super- resolution to enhance neighborhood consensus and improve matching accuracy at the same time. We find that the pro- posed method works quite well in conjunction with trans- former feature backbones, such as MAE [17], DINO [5], and iBOT [50], so we call the proposed method asym- metric correspondence transformer, written as ACTRans- former, and train it end-to-end. We summarize our contri- butions as follows. •We introduce a novel pipeline for semantic correspon- dence which contains generic feature extraction, asymmet- ric feature learning, and matching flow super-resolution. By conducting asymmetric feature learning, we extract specific features for every image pair and thus get more accurate correspondences. Besides, we replace the 4-D correlation refinement with the 2-D matching flow super-resolution, which saves computational cost greatly. •We propose asymmetric feature learning that can project features of image pairs into a shared feature space easily and reduce the feature ambiguity at the same time. For match- ing flow super-resolution, we conduct multi-path super- resolution to benefit from different matching tensors and acquire significant improvements. •Experiments on several popular benchmarks indicate that the proposed ACTRansformer outperforms previous state- of-the-art methods by clear margins. Qualitative results are presenred in Figure 1.

Song_ObjectStitch_Object_Compositing_With_Diffusion_Model_CVPR_2023

Abstract Object compositing based on 2D images is a challeng- ing problem since it typically involves multiple processing stages such as color harmonization, geometry correction and shadow generation to generate realistic results. Fur- thermore, annotating training data pairs for compositing requires substantial manual effort from professionals, and is hardly scalable. Thus, with the recent advances in gen- erative models, in this work, we propose a self-supervised framework for object compositing by leveraging the power of conditional diffusion models. Our framework can hol- listically address the object compositing task in a unified model, transforming the viewpoint, geometry, color and shadow of the generated object while requiring no manual labeling. To preserve the input object’s characteristics, we introduce a content adaptor that helps to maintain categori- * Work done during internship at Adobe.cal semantics and object appearance. A data augmentation method is further adopted to improve the fidelity of the gen- erator. Our method outperforms relevant baselines in both realism and faithfulness of the synthesized result images in a user study on various real-world images.

1. Introduction Image compositing is an essential task in image editing that aims to insert an object from a given image into another image in a realistic way. Conventionally, many sub-tasks are involved in compositing an object to a new scene, in- cluding color harmonization [6, 7, 19, 51], relighting [52], and shadow generation [16, 29, 43] in order to naturally blend the object into the new image. As shown in Tab. 1, most previous methods [6, 7, 16, 19, 28, 43] focus on a sin- gle sub-task required for image compositing. Consequently, they must be appropriately combined to obtain a composite image where the input object is re-synthesized to have 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. 18310 Method Geometry Light Shadow View ST-GAN [28] ✓ ✗ ✗ ✗ SSH [19] ✗ ✓ ✗ ✗ DCCF [51] ✗ ✓ ✗ ✗ SSN [43] ✗ ✗ ✓ ✗ SGRNet [16] ✗ ✗ ✓ ✗ GCC-GAN [5] ✓ ✓ ✗ ✗ Ours ✓ ✓ ✓ ✓ Table 1. Prior works only focus on one or two aspects of object compositing, and they cannot synthesize novel views. In contrast, our model can address all perspectives as listed. color, lighting and shadow that is consistent with the back- ground scene. As shown in Fig. 1, results produced in this way still look unnatural, partly due to the viewpoint of the inserted object being different from the overall background. Harmonizing the geometry and synthesizing novel views have often been overlooked in 2D image compositing, which require an accurate understanding of both the geome- try of the object and the background scene from 2D images. Previous works [15, 21, 22] handle 3D object compositing with explicit background information such as lighting posi- tions and depth. More recently, [28] utilize GANs to esti- mate the homography of the object. However, this method is limited to placing furniture in indoor scenes. In this pa- per, we propose a generic image compositing method that is able to harmonize the geometry of the input object along with color, lighting and shadow with the background image using a diffusion-based generative model. In recent years, generative models such as GANs [4, 10, 17, 20] and diffusion models [1, 14, 30, 32, 33, 38, 39] have shown great potential in synthesizing realistic images. In particular, diffusion model-based frameworks are versa- tile and outperform various prior methods in image edit- ing [1, 23, 32] and other applications [11, 31, 35]. How- ever, most image editing diffusion models focus on using text inputs to manipulate images [1,3,9,23,33], which is in- sufficient for image compositing as verbal representations cannot fully capture the details or preserve the identity and appearance of a given object image. There have been re- cent works [23, 39] focusing on generating diverse contexts while preserving the key features of the object; however, these models are designed for a different task than object compositing. Furthermore, [23] requires fine-tuning the model for each input object and [39] also needs to be fine- tuned on multiple images of the same object. Therefore they are limited for general object compositing. In this work, we leverage diffusion models to simulta- neously handle multiple aspects of image compositing such as color harmonization, relighting, geometry correction and shadow generation. With image guidance rather than text guidance, we aim to preserve the identity and appearance ofthe original object in the generated composite image. Specifically, our model synthesizes a composite image given (i) a source object, (ii) a target background image, and (iii) a bounding box specifying the location to insert the object. The proposed framework consists of a con- tent adaptor and a generator module: the content adaptor is designed to extract a representation from the input object containing both high-level semantics and low-level details such as color and shape; the generator module preserves the background scene while improving the generation quality and versatility. Our framework is trained in a fully self- supervised manner and no task-specific labeling is required at any point during training. Moreover, various data aug- mentation techniques are applied to further improve the fi- delity and realism of the output. We evaluate our proposed method on a real-world dataset closely simulating real use cases for image compositing. Our contributions are summarized as follows: • We present the first diffusion model-based framework forgenerative object compositing that can handle mul- tiple aspects of compositing such as viewpoint, geom- etry, lighting and shadow. • We propose a content adaptor module which learns a descriptive multi-modal embedding from images, en- abling image guidance for diffusion models. • Our framework is trained in a self-supervised man- ner without any task-specific annotations, employing data augmentation techniques to improve the fidelity of generation. • We collect a high-resolution real-world dataset for ob- ject compositing with diverse images, containing man- ually annotated object scales and locations.

Toschi_ReLight_My_NeRF_A_Dataset_for_Novel_View_Synthesis_and_CVPR_2023

Abstract In this paper, we focus on the problem of rendering novel views from a Neural Radiance Field (NeRF) under unob- served light conditions. To this end, we introduce a novel dataset, dubbed ReNe ( Relighting NeRF), framing real world objects under one-light-at-time (OLAT) conditions, annotated with accurate ground-truth camera and light poses. Our acquisition pipeline leverages two robotic arms holding, respectively, a camera and an omni-directional point-wise light source. We release a total of 20 scenes depicting a variety of objects with complex geometry and challenging materials. Each scene includes 2000 images, acquired from 50 different points of views under 40 different OLAT conditions. By leveraging the dataset, we perform an ablation study on the relighting capability of variants of the vanilla NeRF architecture and identify a lightweight archi- tecture that can render novel views of an object under novel light conditions, which we use to establish a non-trivial baseline for the dataset. Dataset and benchmark are avail- able at https://eyecan-ai.github.io/rene .

1. Introduction Inverse rendering [29, 47, 52, 73] addresses the problem of estimating the physical attributes of an object, such as its geometry, material properties and lighting conditions, from a set of images or even just a single one. This task is a longstanding problem for the vision and graphics commu- nities, since it unlocks the creation of novel renderings of an object from arbitrary viewpoints and under unobserved lighting conditions. An effective and robust solution to this problem would have significant value for a wide range of applications in gaming, robotics and augmented reality. Recently, Neural Radiance Fields (NeRF) [36] has con- tributed tremendously to the novel view synthesis sub-task of inverse rendering pipelines. By mapping an input 5D vector (3D position and 2D viewing direction) to a 4D con- tinuous field of volume density and color by means of a neu- ral network, NeRF learns the geometry and appearance of a single scene from a set of posed images. The appealing re- sults in novel view synthesis have attracted a lot of attention from the research community and triggered many follow-up ∗Joint first authorship.⋄Work done while at Eyecan.ai 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. 20762 DatasetMultiple categoriesReal-WorldBackground ShadowsPublicLight Supervision Gross et al. [14] ✗ ✓ ✗ ✓ ✓ Sunet al. [57] ✗ ✓ ✗ ✓ ✓ Wang et al. [68] ✗ ✓ ✗ ✓ ✓ Zhang et al. [74] ✗ ✓ ✗ ✓ ✓ Srinivasan et al. [55] ✓ ✗ ✗ ✓ ✓ Zhang et al. [75] ✓ ✓ ✓ ✓ ✗ Zhang et al. [75] ✓ ✗ ✗ ✓ ✓ Biet al. [2] ✓ ✓ ✗ ✗ ✓ ReNe ✓ ✓ ✓ ✓ ✓ Table 1. Overview of relighting datasets. Our dataset is the first one featuring a variety of objects and materials captured with real- world sensors that provides ground-truth light positions and also presents challenging cast shadows. works aimed at overcoming the main limitations of NeRF, e.g. reduce inference runtime [23, 24, 27, 48, 49, 71], enable modeling of deformable objects [9, 25, 42, 45, 46, 62], and generalization to novel scenes [5,15,19,39,50,53,61,63,72]. However, less attention has been paid to the relighting abil- ity of NeRFs. Although NeRF and its variants represent nowadays the most compelling strategy for view synthe- sis, the learned scene representation entangles material and lighting and, thus, cannot be directly used to generate views under novel, unseen lighting conditions. A few existing works [3, 4, 55, 75] try to overcome this structural NeRF limitation by learning to model the scene appearance as a function of reflectance , which accounts for both scene ge- ometry and lighting modeling. However, these methods in- cur a great computation cost, mainly due to the need to ex- plicitly model light visibility and/or the components of a microfacet Bidrectional Reflectance Distribution Function (BRDF) [65] like diffuse albedo, specular roughness, and point normals: for instance NeRV [55] uses 128 TPU cores for 1 day while [3] is trained on 4 GPUs for 2 days. As highlighted in Tab. 1, one of the main challenges to foster research in this direction is the absence of real-world datasets featuring generic and varied objects with ground- truth light direction, both at training and test time. The lat- ter is key to create a realistic quantitative benchmark for relighting methods, whose availability is one of the main driving forces behind fast-paced development of a machine learning topic. Indeed, the above mentioned NeRF-like methods for relighting mainly consider a handful of syn- thetic images to provide quantitative results (3 and 4 scenes in NeRV [55] and NeRFactor [75], respectively) while no quantitative results are provided in [3]. The only dataset with scenes acquired by a real sensor is proposed in [2], which, however, assumes images captured under collocated view and lighting setup, i.e. a smartphone with flash light on, which limits the amount of cast shadows and simplifies the task. Moreover, the dataset is not publicly available. Some available real-world datasets with ground-truth lightpositions feature human faces or portraits [14, 57, 68, 74]. In these datasets, the subject is usually seated in the center of a light-stage with cameras arranged over a dome array positioned in front of the subject. Although these dataset provide real scenes with ground-truth annotations for light positions, the background is masked-out and shadows cast on the background, which are hard to model because they require a precise knowledge about the geometry of the over- all scene, are ignored. Therefore, in this paper we try to answer the research questions: can we design a data acquisition methodology suitable to collect a set of images of an object under one- light-at-time (OLAT) [74] illumination with high-quality camera and light pose annotations which requires minimal human supervision? We then leverage it to investigate a sec- ond question: can we design a novel Neural Radiance Field architecture to learn to perform relighting with reasonable computational requirements? To answer the first question, we design a capture system relying on two robotic arms. While one arm holds the camera and shots pictures from viewpoints uniformly distributed on a spherical wedge, the other moves the light source across points uniformly dis- tributed on a dome. As a result, we collect the ReNe dataset, made of 20scenes framing daily objects with challenging geometry, varied materials and visible cast shadows, com- posed of 50camera view-points under 40OLAT light con- ditions, i.e. 2000 frames per scene. Examples of images from the dataset are shown in Fig. 1. With a subset of im- ages from each scene, we create a novel hold-out dataset for joint relighting and novel view synthesis evaluation that will be used as an online benchmark to foster research on this important topic. As regards the second question, thanks to the new dataset we conduct a study on the relighting ca- pability of NeRF. In particular, we investigate on how the standard NeRF architecture can be modified to take into account the position of the light when generating the ap- pearance of a scene. Our study shows that by estimating color with two separate sub-networks, one in charge of soft- shadow prediction and one responsible for neurally approx- imating the BRDF, we can perform an effective relighting, e.g. cast complex shadows. We provide results of our novel architecture as a reference baseline for the new benchmark. In summary, our contributions include: • a novel dataset made out of sets of OLAT images of real-world objects, with accurate camera and light pose annotations; • a study comparing different approaches to enable NeRF to perform relighting alongside novel view syn- thesis; • a new architecture, where the stage responsible for ra- diance estimation is split into two separate networks, 20763 that can render novel views under novel unobserved lighting conditions; • a public benchmark for novel view synthesis and re- lighting of real world objects, that will be maintained on an online evaluation server.

Tang_A_New_Benchmark_On_the_Utility_of_Synthetic_Data_With_CVPR_2023

Abstract Deep learning in computer vision has achieved great success with the price of large-scale labeled training data. However, exhaustive data annotation is impracticable for each task of all domains of interest, due to high labor costs and unguaranteed labeling accuracy. Besides, the uncon- trollable data collection process produces non-IID training and test data, where undesired duplication may exist. All these nuisances may hinder the verification of typical theo- ries and exposure to new findings. To circumvent them, an alternative is to generate synthetic data via 3D rendering with domain randomization. We in this work push forward along this line by doing profound and extensive research on bare supervised learning and downstream domain adapta- tion. Specifically, under the well-controlled, IID data set- ting enabled by 3D rendering, we systematically verify the typical, important learning insights, e.g., shortcut learning, and discover the new laws of various data regimes and net- work architectures in generalization. We further investigate the effect of image formation factors on generalization, e.g., object scale, material texture, illumination, camera view- point, and background in a 3D scene. Moreover, we use the simulation-to-reality adaptation as a downstream task for comparing the transferability between synthetic and real data when used for pre-training, which demonstrates that synthetic data pre-training is also promising to improve real test results. Lastly, to promote future research, we develop a new large-scale synthetic-to-real benchmark for image classification, termed S2RDA, which provides more signifi- cant challenges for transfer from simulation to reality.

1. Introduction Recently, we have witnessed considerable advances in various computer vision applications [16,29,38]. However, such a success is vulnerable and expensive in that it has been limited to supervised learning methods with abundant la- *Corresponding author.beled data. Some publicly available datasets exist certainly, which include a great mass of real-world images and ac- quire their labels via crowdsourcing generally. For example, ImageNet-1K [9] is of 1.28M images; MetaShift [26] has 2.56M natural images. Nevertheless, data collection and annotation for all tasks of domains of interest are imprac- tical since many of them require exhaustive manual efforts and valuable domain expertise, e.g., self-driving and medi- cal diagnosis. What’s worse, the label given by humans has no guarantee to be correct, resulting in unpredictable label noise. Besides, the poor-controlled data collection process produces a lot of nuisances, e.g., training and test data aren’t independent identically distributed (IID) and even have du- plicate images. All of these shortcomings could prevent the validation of typical insights and exposure to new findings. To remedy them, one can resort to synthetic data gen- eration via 3D rendering [10], where an arbitrary number of images can be produced with diverse values of imaging factors randomly chosen in a reasonable range, i.e., domain randomization [48]; such a dataset creation pipeline is thus very lucrative, where data with labels come for free. For image classification, Peng et al. [32] propose the first large- scale synthetic-to-real benchmark for visual domain adap- tation [30], VisDA-2017; it includes 152K synthetic images and55K natural ones. Ros et al. [37] produce 9K synthetic cityscape images for cross-domain semantic segmentation. Hinterstoisser et al. [19] densely render a set of 64retail ob- jects for retail detection. All these datasets are customized for specific tasks in cross-domain transfer. In this work, we push forward along this line extensively and profoundly. The deep models tend to find simple, unintended solu- tions and learn shortcut features less related to the seman- tics of particular object classes, due to systematic biases, as revealed in [14]. For example, a model basing its prediction on context would misclassify an airplane floating on water as a boat. The seminal work [14] emphasizes that short- cut opportunities are present in most data and rarely dis- appear by simply adding more data. Modifying the training data to block specific shortcuts may be a promising solution, e.g., making image variation factors consistently distributed 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. 15954 across all categories. To empirically verify the insight, we propose to compare the traditional fixed-dataset periodic training strategy with a new strategy of training with undu- plicated examples generated by 3D rendering, under the well-controlled, IID data setting. We run experiments on three representative network architectures of ResNet [18], ViT [12], and MLP-Mixer [49], which consistently show obvious advantages of the data-unrepeatable training (cf. Sec. 4.1). This also naturally validates the typical argu- ments of probably approximately correct (PAC) generaliza- tion [41] and variance-bias trade-off [11]. Thanks to the ideal IID data condition enabled by the well-controlled 3D rendering, we can also discover more reliable laws of var- ious data regimes and network architectures in generaliza- tion. Some interesting observations are as follows. •Do not learn shortcuts! The test results on synthetic data without background are good enough to show that the synthetically trained models do not learn shortcut solutions relying on context clues [14]. •A zero-sum game. For the data-unrepeatable train- ing, IID and OOD (Out-of-Distribution [34, 55]) gen- eralizations are some type of zero-sum game w.r.t. the strength of data augmentation. •Data augmentations do not help ViT much! In IID tests, ViT performs surprisingly poorly whatever the data augmentation is and even the triple number of training epochs does not improve much. •There is always a bottleneck from synthetic data to OOD/real data . Here, increasing data size and model capacity brings no more benefits, and domain adapta- tion [56] to bridge the distribution gap is indispensable except for evolving the image generation pipeline to synthesize more realistic images. Furthermore, we comprehensively assess image varia- tion factors, e.g., object scale, material texture, illumina- tion, camera viewpoint, and background in a 3D scene. We then find that to generalize well, deep neural networks must learn to ignore non-semantic variability , which may appear in the test. To this end, sufficient images with different val- ues of one imaging factor should be generated to learn a robust, unbiased model, proving the necessity of sample di- versity for generalization [42, 48, 55]. We also observe that different factors and even their different values have uneven importance to IID generalization , implying that the under- explored weighted rendering [3] is worth studying. Bare supervised learning on synthetic data results in poor performance in OOD/real tests, and pre-training and then domain adaptation can improve. Domain adaptation (DA) [56] is a hot research area, which aims to make predictions for unlabeled instances in the target domain by transferringknowledge from the labeled source domain. To our knowl- edge, there is little research on pre-training for DA [24] (with real data). We thus use the popular simulation-to-real classification adaptation [32] as a downstream task, study the transferability of synthetic data pre-trained models by comparing with those pre-trained on real data like ImageNet and MetaShift. We report results for several representative DA methods [7,13,40,45,47] on the commonly used back- bone, and our experiments yield some surprising findings. •The importance of pre-training for DA. DA fails without pre-training (cf. Sec. 4.3.1). •Effects of different pre-training schemes. Different DA methods exhibit different relative advantages un- der different pre-training data. The reliability of exist- ing DA method evaluation criteria is unguaranteed. •Synthetic data pre-training vs. real data pre- training. Synthetic data pre-training is better than pre- training on real data in our study. •Implications for pre-training data setting. Big Syn- thesis Small Real is worth researching. Pre-train with target classes first under limited computing resources. •The improved generalization of DA models. Real data pre-training with extra non-target classes, fine- grained target subclasses, or our synthesized data added for target classes helps DA. Last but not least, we introduce a new, large-scale synthetic-to-real benchmark for classification adaptation (S2RDA), which has two challenging tasks S2RDA-49 and S2RDA-MS-39. S2RDA contains more categories, more re- alistically synthesized source domain data coming for free, and more complicated target domain data collected from di- verse real-world sources, setting a more practical and chal- lenging benchmark for future DA research.

Tu_A_Bag-of-Prototypes_Representation_for_Dataset-Level_Applications_CVPR_2023

Abstract This work investigates dataset vectorization for two dataset-level tasks: assessing training set suitability and test set difficulty. The former measures how suitable a train- ing set is for a target domain, while the latter studies how challenging a test set is for a learned model. Central to the two tasks is measuring the underlying relationship be- tween datasets. This needs a desirable dataset vectoriza- tion scheme, which should preserve as much discriminative dataset information as possible so that the distance between the resulting dataset vectors can reflect dataset-to-dataset similarity. To this end, we propose a bag-of-prototypes (BoP) dataset representation that extends the image-level bag consisting of patch descriptors to dataset-level bag con- sisting of semantic prototypes. Specifically, we develop a codebook consisting of Kprototypes clustered from a ref- erence dataset. Given a dataset to be encoded, we quan- tize each of its image features to a certain prototype in the codebook and obtain a K-dimensional histogram. Without assuming access to dataset labels, the BoP representation provides rich characterization of the dataset semantic dis- tribution. Furthermore, BoP representations cooperate well with Jensen-Shannon divergence for measuring dataset-to- dataset similarity. Although very simple, BoP consistently shows its advantage over existing representations on a se- ries of benchmarks for two dataset-level tasks.

1. Introduction Datasets are fundamental in machine learning research, forming the basis of model training and testing [18, 51, 52, 61]. While large-scale datasets bring opportunities in al- gorithm design, there lack proper tools to analyze and make the best use of them [6,51,56]. Therefore, as opposed to tra- ditional algorithm-centric research where improving mod- els is of primary interest, the community has seen a grow- ing interest in understanding and analyzing the data used for developing models [51, 56]. Recent examples of such goal include data synthesis [29], data sculpting [25,51], and data valuation [6,32,56]. These tasks typically focus on individ- ual sample of a dataset. In this work, we aim to understandnature of datasets from a dataset-level perspective. This work considers two dataset-level tasks: suitability in training and difficulty in testing. First , training set suit- ability denotes whether a training set is suitable for training models for a target dataset. In real-world applications, we are often provided with multiple training sets from various data distributions ( e.g., universities and hospitals). Due to distribution shift, their trained models have different perfor- mance on the target dataset. Then, it is of high practical value to select the most suitable training set for the target dataset. Second , test set difficulty means how challenging a test set is for a learned model. In practice, test sets are usu- ally unlabeled and often come from different distributions than that of the training set. Measuring the test set difficulty for a learned model helps us understand the model reliabil- ity, thereby ensuring safe model deployment. The core of the two dataset-level tasks is to measure the relationship between datasets. For example, a training set is more suitable for learning a model if it is more similar to the target dataset. To this end, we propose a vectoriza- tion scheme to represent a dataset. Then, the relationship between a pair of datasets can be simply reflected by the distance between their representations. Yet, it is challeng- ing to encode a dataset as a representative vector, because (i)a dataset has a different cardinality (number of images) and(ii)each image has its own semantic content ( e.g., cate- gory). It is thus critical to find an effective way to aggregate all image features to uncover dataset semantic distributions. In the literature, some researchers use the first few mo- ments of distributions such as feature mean and co-variance to represent datasets [20, 62, 74, 75, 82]. While being com- putational friendly, these methods do not offer sufficiently strong descriptive ability of a dataset, such as class distri- butions, and thus have limited effectiveness in assessing at- tributes related to semantics. There are also some methods learn task-specific dataset representations [1, 63]. For ex- ample, given a dataset with labels and a task loss function, Task2Vec [1] computes an embedding based on estimates of the Fisher information matrix associated with a probe network’s parameters. While these task-specific represen- tations are able to predict task similarities, they are not suit- able for characterizing dataset properties of interest. They This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2881 require training a network on the specific task [1] or on mul- tiple datasets [63], so they are not effective in assessing the training set suitability. Additionally, they require image la- bels for the specific task, so they cannot be used to measure the difficulty of unlabeled test sets. In this work, we propose a simple and effective bag-of- prototypes (BoP) dataset representation. Its computation starts with partitioning the image feature space into seman- tic regions through clustering, where the region centers, or prototypes, form a codebook. Given a new dataset, we quantize its features to their corresponding prototypes and compute an assignment histogram, which, after normaliza- tion, gives the BoP representation. The dimensionality of BoP equals the codebook size, which is usually a few hun- dred and is considered memory-efficient. Meanwhile, the histogram computed on the prototypes is descriptive of the dataset semantic distribution. Apart from being low dimensional and semantically rich, BoP has a few other advantages. First , while recent works in task-specific dataset representation usually require full image annotations and additional learning procedure [1,63], the computation of BoP does not rely on any. It is relatively efficient and allows for unsupervised assessment of dataset attributes. Second , BoP supports dataset-to-dataset similar- ity measurement through Jensen-Shannon divergence. We show in our experiment that this similarity is superior to commonly used metrics such as Fr ´echet distance [27] and maximum mean discrepancy [33] in two dataset-level tasks.

Tumanyan_Plug-and-Play_Diffusion_Features_for_Text-Driven_Image-to-Image_Translation_CVPR_2023

Abstract Large-scale text-to-image generative models have been a revolutionary breakthrough in the evolution of genera- tive AI, synthesizing diverse images with highly complex visual concepts. However, a pivotal challenge in leverag- ing such models for real-world content creation is provid- ing users with control over the generated content. In this paper, we present a new framework that takes text-to- im- age synthesis to the realm of image-to-image translation – given a guidance image and a target text prompt as input, our method harnesses the power of a pre-trained text-to- image diffusion model to generate a new image that com- plies with the target text, while preserving the semantic lay- out of the guidance image. Specifically, we observe and empirically demonstrate that fine-grained control over the generated structure can be achieved by manipulating spa- tial features and their self-attention inside the model. This results in a simple and effective approach, where features extracted from the guidance image are directly injected into the generation process of the translated image, requiring no training or fine-tuning. We demonstrate high-quality results on versatile text-guided image translation tasks, including translating sketches, rough drawings and animations into realistic images, changing the class and appearance of ob- ∗Equal contribution.jects in a given image, and modifying global qualities such as lighting and color.

1. Introduction With the rise of text-to-image foundation models – billion-parameter models trained on a massive amount of text-image data, it seems that we can translate our imagina- tion into high-quality images through text [13, 35,37,41]. While such foundation models unlock a new world of cre- ative processes in content creation, their power and expres- sivity come at the expense of user controllability, which is largely restricted to guiding the generation solely through an input text. In this paper, we focus on attaining con- trol over the generated structure and semantic layout of the scene – an imperative component in various real-world con- tent creation tasks, ranging from visual branding and mar- keting to digital art. That is, our goal is to take text-to-image generation to the realm of text-guided Image-to-Image (I2I) translation, where an input image guides the layout (e.g., the structure of the horse in Fig. 1), and the text guides the per- ceived semantics and appearance of the scene (e.g., “robot horse” in Fig. 1). A possible approach for achieving control of the gen- erated layout is to design text-to-image foundation models that explicitly incorporate additional guiding signals, such as user-provided masks [13, 29,35]. For example, recently 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. 1921 Make-A-Scene [13] trained a text-to-image model that is also conditioned on a label segmentation mask, defining the layout and the categories of objects in the scene. However, such an approach requires an extensive compute as well as large-scale text-guidance-image training tuples, and can be applied at test-time to these specific types of inputs. In this paper, we are interested in a unified framework that can be applied to versatile I2I translation tasks, where the struc- ture guidance signal ranges from artistic drawings to photo- realistic images (see Fig. 1). Our method does not require any training or fine-tuning, but rather leverages a pre-trained andfixed text-to-image diffusion model [37]. We pose the fundamental question of how structure in- formation is internally encoded in such a model. We dive into the intermediate spatial features that are formed during the generation process, empirically analyze them, and de- vise a new framework that enables fine-grained control over the generated structure by applying simple manipulations to spatial features inside the model. Specifically, spatial fea- tures and their self-attentions are extracted from the guid- ance image, and are directly injected into the text-guided generation process of the target image. We demonstrate that our approach is not only applicable in cases where the guid- ance image is generated from text, but also for real-world images that are inverted into the model. To summarize, we make the following key contributions: (i) We provide new empirical insights about internal spatial features formed during the diffusion process. (ii) We introduce an effective framework that leverages the power of pre-trained and fixed guided diffusion, allowing to perform high-quality text-guided I2I translation without any training or fine-tuning. (iii) We show, both quantitatively and qualitatively that our method outperforms existing state-of-the-art baselines, achieving significantly better balance between preserving the guidance layout and deviating from its appearance.

Tokmakov_Breaking_the_Object_in_Video_Object_Segmentation_CVPR_2023

Abstract The appearance of an object can be fleeting when it transforms. As eggs are broken or paper is torn, their color, shape and texture can change dramatically, preserving vir- tually nothing of the original except for the identity itself. Yet, this important phenomenon is largely absent from ex- isting video object segmentation (VOS) benchmarks. In this work, we close the gap by collecting a new dataset for Video Object Segmentation under Transformations (VOST). It consists of more than 700 high-resolution videos, cap- tured in diverse environments, which are 20 seconds long on average and densely labeled with instance masks. We adopt a careful, multi-step approach to ensure that these videos focus on complex object transformations, capturing their full temporal extent. We then extensively evaluate state- of-the-art VOS methods and make a number of important discoveries. In particular, we show that existing methods struggle when applied to this novel task and that their main limitation lies in over-reliance on static appearance cues. This motivates us to propose a few modifications for the top- performing baseline that improve its capabilities by better modeling spatio-temporal information. More broadly, our work highlights the need for further research on learning more robust video object representations. Nothing is lost or created, all things are merely transformed. Antoine Lavoisier

1. Introduction Spatio-temporal cues are central in segmenting and tracking objects in humans, with static appearance play- ing only a supporting role [23, 27, 43]. In the most ex- treme scenarios, we can even localize and track objects de- fined by coherent motion alone, with no unique appearance whatsoever [20]. Among other benefits, this appearance- last approach increases robustness to sensory noise and enables object permanence reasoning [41]. By contrast, modern computer vision models for video object segmenta- tion [3, 11, 44, 64] operate in an appearance-first paradigm. Figure 1. Video frames from the DA VIS’17 dataset [42] (above), and our proposed VOST (below). While existing VOS datasets feature many challenges, such as deformations and pose change, the overall appearance of objects varies little. Our work focuses on object transformations, where appearance is no longer a reliable cue and more advanced spatio-temporal modeling is required. Indeed, the most successful approaches effectively store patches with associated instance labels and retrieve the clos- est patches to segment the target frame [11, 38, 44, 64]. What are the reasons for this stark disparity? While some are algorithmic (e.g., object recognition models being first developed for static images), a key reason lies in the datasets we use. See for instance the “Breakdance” sequence from the validation set of DA VIS’17 [42] in Figure 1: while the dancer’s body experiences significant deformations and pose changes, the overall appearance of the person remains constant, making it an extremely strong cue. However, this example – representative of many VOS datasets – covers only a narrow slice of the life of an object. In addition to translations, rotations, and minor deforma- tions, objects can transform. Bananas can be peeled, paper can be cut, clay can be molded into bricks, etc. These trans- formations can dramatically change the color, texture, and shape of an object, preserving virtually nothing of the orig- 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. 22836 Figure 2. Representative samples from VOST with annotations at three different time steps (see video for full results). Colours indicate instance ids, with grey representing ignored regions. VOST captures a wide variety of transformations in diverse environments and provides pixel-perfect labels even for the most challenging sequences. inal except for the identity itself (see Figure 1, bottom and Figure 2). As we show in this paper, tracking object identity through these changes is relatively easy for humans (e.g. la- belers), but very challenging for VOS models. In this work, we set out to fill this gap and study the problem of segment- ing objects as they undergo complex transformations. We begin by collecting a dataset that focuses on these scenarios in Section 3. We capitalize on the recent large- scale, ego-centric video collections [13, 21], which contain thousands of examples of human-object interactions with activity labels. We carefully filter these clips to only in- clude major object transformations using a combination of linguistic cues (change of state verbs [19, 29]) and man- ual inspection. The resulting dataset, which we call VOST (Video Object Segmentation under Transformations), con- tains 713 clips, covering 51 transformations over 155 object categories with an average video length of 21.2 seconds. We then densely label these videos with more than 175,000 masks, using an unambiguous principle inspired by spatio- temporal continuity: if a region is marked as an object in the first frame of a video, all the parts that originate from it maintain the same identity (see Figure 2). Equipped with this unique dataset, we analyze state- of-the-art VOS algorithms in Section 4. We strive to in- clude a representative set of baselines that illustrates the majority of the types of approaches to the problem in the literature, including classical, first frame matching meth- ods [61], local mask-propagation objectives [26], alter- native, object-level architectures [3], and the mainstream memory-based models [11,63–65]. Firstly, we observe that existing methods are indeed ill-equipped for segmenting ob- jects through complex transformations, as illustrated by the large (2.3-12.5 times) gap in performance between VOST and DA VIS’17 (see Table 2). A closer analysis of the results reveals the following discoveries: (1) performance of the methods is inversely proportional to their reliance on static appearance cues; (2) progress on VOST can be achieved byimproving the spatio-temporal modeling capacity of exist- ing architectures; (3) the problem is not easily solvable by training existing methods on more data. We conclude in Section 5 by summarizing the main chal- lenges associated with modeling object transformations. We hope that this work will motivate further exploration into more robust video object representations. Our dataset, source code, and models are available at vostdataset.org.

Sun_MISC210K_A_Large-Scale_Dataset_for_Multi-Instance_Semantic_Correspondence_CVPR_2023

Abstract Semantic correspondence have built up a new way for object recognition. However current single-object match- ing schema can be hard for discovering commonalities for a category and far from the real-world recognition tasks. To fill this gap, we design the multi-instance semantic corre- spondence task which aims at constructing the correspon- dence between multiple objects in an image pair. To sup- port this task, we build a multi-instance semantic corre- spondence (MISC) dataset from COCO Detection 2017 task called MISC210K. We construct our dataset as three steps: (1) category selection and data cleaning; (2) keypoint de- sign based on 3D models and object description rules; (3) human-machine collaborative annotation. Following these steps, we select 34 classes of objects with 4,812 challenging images annotated via a well designed semi-automatic work- flow, and finally acquire 218,179 image pairs with instance masks and instance-level keypoint pairs annotated. We de- sign a dual-path collaborative learning pipeline to train instance-level co-segmentation task and fine-grained level correspondence task together. Benchmark evaluation and further ablation results with detailed analysis are provided with three future directions proposed. Our project is avail- able on https://github.com/YXSUNMADMAX/MISC210K.

1. Introduction Building dense visual correspondences is a sub-task of image matching, which aims at finding semantic associ- ations of salient parts and feature points of objects or scenes [4,6,34,49]. This task has established a new way for understanding commonalities among objects in a more fine- grained manner and has been widely used for various com- puter vision tasks, including few shot learning [11, 21, 48], ∗: Contribution Equally †: Corresponding Authors Figure 1. An instructive overview for our multi-instance semantic correspondence task. In this task, models are required to determine the corresponding relationships among multiple instances, where instance masks are introduced for grouping matching key-points. multi-object tracking [24], and image editing [10, 15, 32]. To learn general semantic correspondence, several popu- lar datasets, such as Caltech-101 [14], FG3DCar [37], PF- WILLOW [8], PF-PASCAL [8], and SPair-71k [27], have been proposed by researchers to train machine learning models. These datasets were designed to capture large intra- class variations in color, scale, orientation, illumination and non-rigid deformation. However, although these datasets provide rich annotations, they are still far from real-world applications because each object category is only allowed to have at most one instance in each image. Moreover, for most object recognition tasks and applications, multiple ob- jects of the same category often appear at the same time. Existing datasets only focus on one-to-one matching with- out considering multi-instance scenes, and thus cannot be used as simulations of real-world applications. In this paper, we aim to reduce the gap between one- to-one matching and many-to-many matching by building a new multi-instance semantic correspondence dataset. Fol- lowing PF-PASCAL [8] and SPair-71k [27], we label key- points on objects to construct the dataset. There are sev- eral key challenges during data labeling. First, how to choose the collection of raw images that contain multiple objects in natural scenes? Second, how to choose candi- 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. 7121 date object categories that have rich keypoints with dis- criminative semantics? Third, how to ensure label quality while maintaining high annotation efficiency? Last, how to design an evaluation protocol for multi-instance seman- tic correspondence? While addressing the above issues, we construct a new multi-instance semantic correspondence dataset, called MISC210K, which collects 34 different ob- ject classes from COCO 2017 detection challenge [20] and contains 218,179 image pairs with large variations in view- point, scale, occlusion, and truncation. Compared with pop- ular PF-PASCAL [8] and SPair-71k [27], MISC210K has much more annotated instances, covers a boarder range of object categories and presents a large number of many-to- many matching cases. Besides, we also design a new pro- tocol to evaluate many-to-many semantic matching algo- rithms. All these characteristics make MISC210K appeal- ing to the relevant research community. We summarize main characteristics of MISC210K as follows. First, MISC210K provides annotations for many- to-many matching. Unlike previous datasets [8, 27] which only exploit one-to-one matching, we find out all semantic correspondences among multiple objects (up to 4) across image pairs as shown in Figure 1. Second, MISC210K has more complicated annotations. The number of keypoints in SPair-71k varies from 3 to 30 across categories. In con- trast, we well design more keypoints to highlight object contours, skeletal joints, and other distinctive feature points that can characterize objects in detail. Third, MISC210K has a larger scale in comparison to existing datasets. It contains 218,179 image pairs across 34 object categories, which is three times larger than the previous largest dataset, SPair-71k. Fourth, intra-class variations in MISC210K are more challenging. In addition to variations considered in Spair-71k [26], we also introduce more challenging varia- tions, such as mutual occlusion of multiple objects and per- spective distortions in complex scenes. To investigate whether MISC210K can help learn gen- eral correspondences across multiple object instances, we evaluate previous state-of-the-art methods, MMNet [49], CATs [4] on MSIC210K. We also propose a dual-path multi-task learning pipeline to solve the complicated multi- instance semantic correspondence problem. For both tasks of correspondence and instance co-segmentation, we de- signed multi-instance PCK (mPCK) and mIOU (instance) from works [8,25,49]. According to the results, we identify new challenges in this task: (1) extracting discriminative features plays a precursory role to find out commonalities across multiple objects; (2) the uncertainty in the number of matching keypoints makes the matching process more difficult; (3) multiple object instances bring occlusion, in- terlacing, and other challenging issues. These observations indicate that multi-instance semantic correspondence is a challenging problem deserving further investigation.This paper is organized as follows. We first describe the MISC210K dataset, its collection process, and statistics. Then we introduce a generic framework for multi-instance semantic correspondence, which enables neural networks to associate salient feature points of object instances across different images. The proposed dual-path collaborative learning (DPCL) pipeline outperforms the transfer of pre- vious one-to-one semantic correspondence algorithms. We further analyze the characteristics of MISC210K and dis- cuss key issues in multi-instance semantic correspondence.

Sun_Co-Speech_Gesture_Synthesis_by_Reinforcement_Learning_With_Contrastive_Pre-Trained_Rewards_CVPR_2023

Abstract There is a growing demand of automatically synthesizing co-speech gestures for virtual characters. However, it re- mains a challenge due to the complex relationship between input speeches and target gestures. Most existing works fo- cus on predicting the next gesture that fits the data best, however, such methods are myopic and lack the ability to plan for future gestures. In this paper, we propose a novel reinforcement learning (RL) framework called RACER to generate sequences of gestures that maximize the overall satisfactory. RACER employs a vector quantized varia- tional autoencoder to learn compact representations of ges- tures and a GPT-based policy architecture to generate co- herent sequence of gestures autoregressively. In particular, we propose a contrastive pre-training approach to calculate the rewards, which integrates contextual information into action evaluation and successfully captures the complex re- lationships between multi-modal speech-gesture data. Ex- perimental results show that our method significantly out- performs existing baselines in terms of both objective met- rics and subjective human judgements. Demos can be found athttps://github.com/RLracer/RACER.git .

1. Introduction Gesturing is important for human speakers to improve their expressiveness. It conveys the necessary non-verbal information to the audience, giving the speech a more emo- tional touch so that to enhance persuasiveness and credibil- ity. Similarly, in virtual world, high-quality 3D gesture ani- mations can make a talking character more vividly. For ex- ample, in human-computer interaction scenarios, vivid ges- tures performed by virtual characters can help listeners to concentrate and improve the intimacy between humans and *Equal contribution. †Corresponding author.characters [ 36]. Attracted by these merits, there has been a growing demand of automatically synthesizing high-quality co-speech gestures in computer animation. However, automatically synthesizing co-speech gestures remains a challenge due to the complicated relationship be- tween speech audios and gestures. On one hand, a speaker may play different gestures when speaking the same words due to different mental and physical states. On the other hand, a speaker may also play similar gestures when speak- ing different words. Therefore, co-speech gesture synthe- sizing is inherently a “many-to-many” problem [ 21]. More- over, in order to ensure the overall fluency and consistency, we must take into consideration both the contextual infor- mation and the subsequent effect upon playing a gesture [23,38]. Therefore, gesture synthesizing is rather a sequen- tial decision making problem than a simple matching be- tween speeches and gestures. Compared with traditional rule-based approaches [ 35], data-driven gesture synthesis approaches [ 10,39] has shown many advantages, including the low development cost and the ability to generalize. Most existing data-driven ap- proaches consider gesture synthesis as a classification task [6,26] or a regression task [ 17] in a deterministic way i.e. the same speech or text input always maps to the same ges- ture output. However, these models rely on the assumption that there exists a unique ground-truth label for each in- put sequence of speech, which contradicts to the “many-to- many” nature of the problem. As a consequence, they sac- rifice diversity and semantics of gestures and tend to learn some averaged gestures for any input speeches. Some other works adopt adversarial learning framework, where a dis- criminator is trained to distinguish between generated ges- tures and the recorded gestures in dataset [ 10,38]. Although they improve the generalizability of the gesture generator to some extent, they still fail to explore the essential relation- ship between speeches and gestures. In order to address the above challenges, we propose This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2331 a novel Reinforcement leArning framework with Contra- sitive prE-trained Rewards (RACER) for generating high- quality co-speech gestures. RACER is trained in an of- fline manner but can be used to generate the next gesture for a speaking character in real time. RACER consists of the following three components. Firstly, in order to ex- tract meaningful gestures from the infinite action space, RACER adopts a vector quantized variational autoencoder (VQ-V AE) [ 32] to learn compact gesture representations, which significantly reduces the action space. Secondly, we construct the Q-value network using a GPT-based [ 28] model, which has natural advantages of generating coher- ence sequence of gestures. Thirdly, inspired by the con- trastive language-image pre-training (CLIP) [ 27] and recent advances of contrastive learning [ 30], we propose a con- trastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to explore sophisticated relations between speeches and gestures. Note that these rewards evaluate the quality of gestures as a sequence. By contrast, in conventional supervised learning frameworks, the focus is on predicting only the next gesture, ignoring the quality of generated sequence as a whole. To sum up, the core contributions of this paper are as follows: (1)We formally model the co-speech gesture synthesis problem as a Markov decision process and propose a novel RL based approach called RACER to learn the optimal gesture synthesis policy. RACER can be trained in an offline manner and used to synthesis co- speech gestures in real time. (2)We introduce VQ-V AE to encode and quantize the motion segments to a codebook, which significantly reduces the action space and facilitates the RL phase. (3)We propose a contrastive speech-gesture pre-training method to compute the rewards, which guide the RL agent to discover deeper relations from multi-modal speech-gesture data. (4)Extensive experimental results show that RACER outperforms the existing baselines in terms of both objective metrics and subjective human judgements. This demonstrates the superiority and the potential of RL in co-speech gesture synthesis tasks.

Touvron_Co-Training_2L_Submodels_for_Visual_Recognition_CVPR_2023

Abstract We introduce submodel co-training, a regularization method related to co-training, self-distillation and stochas- tic depth. Given a neural network to be trained, for each sample we implicitly instantiate two altered networks, “sub- models”, with stochastic depth: we activate only a subset of the layers. Each network serves as a soft teacher to the other, by providing a loss that complements the regular loss provided by the one-hot label. Our approach, dubbed “co- sub”, uses a single set of weights, and does not involve a pre-trained external model or temporal averaging. Experimentally, we show that submodel co-training is effective to train backbones for recognition tasks such as image classification and semantic segmentation. Our ap- proach is compatible with multiple architectures, including RegNet, ViT, PiT, XCiT, Swin and ConvNext. Our training strategy improves their results in comparable settings. For instance, a ViT-B pretrained with cosub on ImageNet-21k obtains 87.4% top-1 acc. @448 on ImageNet-val.

1. Introduction Although the fundamental ideas of deep trainable neural networks have been around for decades, only recently have barriers been removed to allow breakthroughs in success- fully training deep neural architectures in practice. Many of these barriers are related to non-convex optimization in one way or another, which is central to the success of modern neural networks. The optimization challenges have been addressed from multiple angles in the literature. First, mod- ern architectures are designed to facilitate the optimization of very deep networks. An exceptionally successful design principle is using residual connections [ 24,25]. Although this does not change the expressiveness of the functions that the network can implement, the improved gradient flow al- leviates, to some extent, the difficulties of optimizing very deep networks. Another key element to the optimization is the importance of data, revealed by the step-change in vi- sual recognition performance resulting from the ImageNet dataset [ 11], and the popularization of transfer learning with pre-training on large datasets [ 39,58].θ<latexit sha1_base64="UakMuYN+OmatTz2zliypy+NMzBA=">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</latexit>T<latexit sha1_base64="eKHKbIM1GmwN+GJF4oitKCUY1GQ=">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</latexit>T<latexit sha1_base64="eKHKbIM1GmwN+GJF4oitKCUY1GQ=">AAACPXicdVBNTxRBEO1B0XVFBTkak44LCafJzIosRxIvHteEBeLOhNT01EJn+2PSXYNuJvMvuMJf8Xf4A7gZr17tXdZERF/SyatXVanXr6iU9JQk36KVBw9XHz3uPOk+XXv2/MX6xssjb2sncCSssu6kAI9KGhyRJIUnlUPQhcLjYvp+3j++QOelNYc0qzDXcGbkRAqgIH3KNNC5AMUPT9d7SbybJoP+gN8naZws0GNLDE83otdZaUWt0ZBQ4P04TSrKG3AkhcK2m9UeKxBTOMNxoAY0+rxZWG75dlBKPrEuPEN8of650YD2fqaLMDm36P/uzcV/9cY1TfbzRpqqJjTi9tCkVpwsn/+fl9KhIDULBISTwSsX5+BAUEjpzpXC2ilB4dtuNzP4WVitwZRN5qnU0tS+Had5qEJ6WNgvTRLvvWubrazSvJdutW0I9Hdq/P/kqB+nb+P+x93ewc4y2g57xd6wHZayATtgH9iQjZhghl2yK3YdfY1uou/Rj9vRlWi5s8nuIPr5C2MJrdw=</latexit>target2 target1 1−λλ submodel2submodel1 θ1 <latexit sha1_base64="vDhJRxCARDrEDJh8J1+qCQooVZs=">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</latexit> θ2 <latexit sha1_base64="pBZz6yBzS2GzIG2wxfqacBFCUHs=">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</latexit> augmentationlabel “tree” Figure 1. Co-training of submodels (cosub): for each image, two sub- models are sampled by randomly dropping layers of the full model. The training signal for each submodel mixes the cross-entropy loss from the image label with a self-distillation loss obtained from the other submodel. However, even when (pre-)trained with millions of im- ages, recent deep networks with millions if not billions of parameters, are still heavily overparameterized. Tradi- tional regularization like weight decay, dropout [ 46], or la- bel smoothing [ 47] are limited in their ability to address this issue. Data-augmentation strategies, including those mixing different images like Mixup [ 61] and CutMix [ 60], have proven to provide a complementary data-driven form of regularization. More recently, multiple works propose to resort to self-supervised pre-training. These approaches rely on a proxy objective that generally provides more su- pervision signal than the one available from labels, like in recent (masked) auto-encoders [ 5,16,22], which were popu- lar in the early deep learning literature [ 7,19,27]. Similarly, contrastive approaches [ 23] or self-distillation [ 9] provide a richer supervision less prone to supervision collapse [ 12]. Overall, self-supervised learning makes it possible to learn larger models with less data, possibly reducing the need of a pre-training stage [ 15]. 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. 11701 Distillation is a complementary approach to improve op- timization. Distillation techniques were originally devel- oped to transfer knowledge from a teacher model to a stu- dent model [ 4,28], allowing the student to improve over learning from the data directly. In contrast to traditional distillation, co-distillation does not require pre-training a (strong) teacher. Instead, a pool of models supervise each other. Practically, it faces several limitations, including the difficulty of jointly training more than two students for com- plexity reasons, as it involves duplicating the weights. In this paper, we propose a practical way to enable co- training for a very large number of students. We consider a single target model to be trained, and we instantiate two submodels on-the-fly , simply by layerwise dropout [ 20,31]. This gives us two neural networks through which we can backpropagate to the shared parameters of the target model. In addition to the regular training loss, each submodel serves as a teacher to the other, which provides an addi- tional supervision signal ensuring the consistency across the submodels. Our approach is illustrated in Figure 1: the pa- rameterλcontrols the importance of the co-training loss compared to the label loss, and our experiments show that it significantly increases the final model accuracy. This co-training across different submodels, which we refer to as cosub , can be regarded as a massive co-training between2Lmodels that share a common set of parameters, whereLis the number of layers in the target architecture. The target model can be interpreted as the expectation of all models. With a layer drop-rate set to 0.5, for instance for a ViT-H model, all submodels are equiprobable, and then it amounts to averaging the weights of 22×32models. Our contributions can be summarized as follows: • We introduce a novel training approach for deep neu- ral networks: we co-train submodels. This signifi- cantly improves the training of most models, establish- ing the new state of the art in multiple cases. For in- stance, after pre-training ViT-B on Imagenet-21k and fine-tuning it at resolution 448, we obtain 87.4% top-1 accuracy on Imagenet-val. • We provide an efficient implementation to subsample models on the fly. It is a simple yet effective variation of stochastic depth [ 31] to drop residual blocks. • We provide multiple analyses and ablations. Notice- ably, we show that our submodels are effective models by themselves even with significant trimming, similar to LayerDrop [ 20] in natural language processing. • We validate our approach on multiple architectures (like ViT, ResNet, RegNet, PiT, XCiT, Swin, Con- vNext), both for image classification –trained from scratch or with transfer–, and semantic segmentation. • We will share models/code for reproducibility in the DeiT repository .

Stergiou_The_Wisdom_of_Crowds_Temporal_Progressive_Attention_for_Early_Action_CVPR_2023

Abstract Early action prediction deals with inferring the ongoing action from partially-observed videos, typically at the out- set of the video. We propose a bottleneck-based attention model that captures the evolution of the action, through pro- gressive sampling over fine-to-coarse scales. Our proposed Temporal Progressive (TemPr) model is composed of mul- tiple attention towers, one for each scale. The predicted action label is based on the collective agreement consider- ing confidences of these towers. Extensive experiments over four video datasets showcase state-of-the-art performance on the task of Early Action Prediction across a range of en- coder architectures. We demonstrate the effectiveness and consistency of TemPr through detailed ablations.†

1. Introduction Early action prediction (EAP) is the task of inferring the action label corresponding to a given video, from only partially observing the start of that video. Interest in EAP has increased in recent years due to both the ever-growing number of videos recorded and the requirement of pro- cessing them with minimal latency. Motivated by the ad- vances in action recognition [6, 57], where the entire video is used to recognize the action label, recent EAP meth- ods [3,15,34,45,60] distill the knowledge from these recog- nition models to learn from the observed segments. Despite promising results, the information that can be extracted from partial and full videos is inevitably different. We in- stead focus on modeling the observed partial video better. Several neurophysiological studies [11, 29] have sug- gested that humans understand actions in a predictive and not reactive manner. This has resulted in the direct match- ing hypothesis [18, 46] where, actions are believed to be perceived through common patterns. Encountering any of these patterns prompts the expectation of specific action(s), even before the action is completed. Although the early pre- *Work carried out while A. Stergiou was at University of Bristol †Code is available at: https://tinyurl.com/temprog Figure 1. Early action prediction with TemPr involves the use of multiple scales for extracting features over partially observed videos. Encoded spatio-temporal features are attended by distinct transformer towers ( T) at each scale. We visualize two scales, where the fine scale Tipredicts ‘hold plate’, and the coarse scale Ti+1predicts ‘hold sponge’. Informative cues from both scales are combined for early prediction of the action ‘wash plate’. diction of actions is an inherent part of human cognition, the task remains challenging for computational modeling. Motivated by the direct matching hypothesis, we propose a Temporally Progressive (TemPr) approach to modeling partially observed videos. Inspired by multi-scale repre- sentations in images [7, 69] and video [27, 62], we repre- sent the observed video by a set of sub-sequences of tem- porally increasing lengths as in Figure 1, which we refer to as scales. TemPr uses distinct transformer towers over each video scale. These utilize a shared latent-bottleneck for cross-attention [28, 37], followed by a stack of self- attention blocks to concurrently encode and aggregate the input. From tower outputs, a shared classifier produces la- bel predictions for each scale. Labels are aggregated based on their collective similarity and individual confidences. In summary, our contributions are as follows: (i) We 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. 14709 propose a progressive fine-to-coarse temporal sampling ap- proach for EAP. (ii) We use transformer towers over sam- pled scales to capture discriminative representations and adaptively aggregate tower predictions, based on their con- fidence and collective agreement. (iii) We evaluate the effectiveness of our approach over four video datasets: UCF-101 [53], EPIC-KITCHENS [8], NTU-RGB [51] and Something-Something (sub-21 & v2) [21], consistently out- performing prior work.

Su_Towards_All-in-One_Pre-Training_via_Maximizing_Multi-Modal_Mutual_Information_CVPR_2023

Abstract To effectively exploit the potential of large-scale mod- els, various pre-training strategies supported by massive data from different sources are proposed, including su- pervised pre-training, weakly-supervised pre-training, and self-supervised pre-training. It has been proved that com- bining multiple pre-training strategies and data from var- ious modalities/sources can greatly boost the training of large-scale models. However, current works adopt a multi- stage pre-training system, where the complex pipeline may increase the uncertainty and instability of the pre-training. It is thus desirable that these strategies can be integrated in a single-stage manner. In this paper, we first pro- pose a general multi-modal mutual information formula as a unified optimization target and demonstrate that all mainstream approaches are special cases of our frame- work. Under this unified perspective, we propose an all-in- one single-stage pre-training approach, named Maximizing Multi-modal Mutual Information Pre-training ( M3I Pre- training ). Our approach achieves better performance than previous pre-training methods on various vision bench- marks, including ImageNet classification, COCO object de- tection, LVIS long-tailed object detection, and ADE20k se- mantic segmentation. Notably, we successfully pre-train a billion-level parameter image backbone and achieve state- of-the-art performance on various benchmarks under pub- lic data setting. Code shall be released at https:// github.com/OpenGVLab/M3I-Pretraining .

1. Introduction In recent years, large-scale pre-trained models [5,13,27, 30, 37, 55, 65, 91] have swept a variety of computer vision ∗Equal contribution.†This work is done when Weijie Su and Chenxin Tao are interns at Shanghai Artificial Intelligence Laboratory. Corre- sponding author. Self-supervised Pre -training (intra view) M3I Pre -training (ours)Supervised Pre -training input𝑥 target𝑦image backbone 𝑓𝜃 category embedding 𝑓𝜙GAP+Linear 𝑓𝜓 softmax cross - entropy loss Weakly -supervised Pre -training Self-supervised Pre -training (inter view)predicted feature Ƹ𝑧𝑦 Doginpu tfeature 𝑧𝑥 targe tfeature 𝑧𝑦𝐦𝐚𝐱𝑰(𝒛𝒙;𝒛𝒚) input𝑥 target𝑦image backbone 𝑓𝜃 text backbone + GAP 𝑓𝜙GAP 𝑓𝜓 softmax cross - entropy losspredicted feature Ƹ𝑧𝑦 A dog by theseainpu tfeature 𝑧𝑥 targe tfeature 𝑧𝑦𝐦𝐚𝐱𝑰(𝒛𝒙;𝒛𝒚) input𝑥 target𝑦image backbone 𝑓𝜃 imag ebackbone / Identity + (GAP) 𝑓𝜙Transformer/ MLP 𝑓𝜓 L2-normlosspredicted feature Ƹ𝑧𝑦 inpu tfeature 𝑧𝑥 targe tfeature 𝑧𝑦𝐦𝐚𝐱𝑰(𝒛𝒙;𝒛𝒚) input𝑥 target𝑦image backbone 𝑓𝜃 imag ebackbone / Identity + (GAP) 𝑓𝜙Transformer/ MLP 𝑓𝜓 L2-normlosspredicted feature Ƹ𝑧𝑦 inpu tfeature 𝑧𝑥 targe tfeature 𝑧𝑦𝐦𝐚𝐱𝑰(𝒛𝒙;𝒛𝒚) input𝑥 target𝑦image backbone 𝑓𝜃 image/text backbone+(GAP) 𝑓𝜙Transformer/ GAP+MLP 𝑓𝜓 L2-norm/ softmax CE losspredicted feature Ƹ𝑧𝑦 inpu tfeature 𝑧𝑥 targe tfeature 𝑧𝑦𝐦𝐚𝐱𝑰(𝒛𝒙;𝒛𝒚) Figure 1. Comparison between different pre-training paradigms and M3I Pre-training. Existing pre-training methods are all opti- mizing the mutual information between the input and target repre- sentations, which can be integrated by M3I Pre-training. “GAP” refers to global average pooling. tasks with their strong performance. To adequately train large models with billions of parameters, researchers design various annotation-free self-training tasks and obtain suffi- ciently large amounts of data from various modalities and sources. In general, existing large-scale pre-training strate- gies are mainly divided into three types: supervised learn- 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. 15888 ing [20, 65] on pseudo-labeled data ( e.g., JFT-300M [85]), weakly supervised learning [37,55] on web crawling images text pairs ( e.g., LAION-400M [56]), and self-supervised learning [5, 13, 27, 30, 91] on unlabeled images. Supported by massive data, all these strategies have their own advan- tages and have been proven to be effective for large models of different tasks. In pursuit of stronger representations of large models, some recent approaches [47, 78, 81] combine the advantages of these strategies by directly using differ- ent proxy tasks at different stages, significantly pushing the performance boundaries of various vision tasks. Nevertheless, the pipeline of these multi-stage pre- training approaches is complex and fragile, which may lead to uncertainty and catastrophic forgetting issues. Specif- ically, the final performance is only available after com- pleting the entire multi-stage pre-training pipeline. Due to the lack of effective training monitors in the intermedi- ate stages, it is difficult to locate the problematic training stage when the final performance is poor. To eliminate this dilemma, it is urgent to develop a single-stage pre-training framework that can take advantage of various supervision signals. It is natural to raise the following question: Is it possible to design an all-in-one pre-training method to have all the desired representational properties? To this end, we first point out that different single- stage pre-training methods share a unified design princi- ple through a generic pre-training theoretical framework. We further extend this framework to a multi-input multi- target setting so that different pre-training methods can be integrated systematically. In this way, we propose a novel single-stage pre-training method, termed M3I Pre-training, that all desired representational properties are combined in a unified framework and trained together in a single stage. Specifically, we first introduce a generic pre-training the- oretical framework that can be instantiated to cover existing mainstream pre-training methods. This framework aims to maximize the mutual information between input represen- tation and target representation, which can be further de- rived into a prediction term with a regularization term. (1) The prediction term reconstructs training targets from the network inputs, which is equivalent to existing well-known pre-training losses by choosing proper forms for the pre- dicted distribution. (2) The regularization term requires the distribution of the target to maintain high entropy to prevent collapse, which is usually implemented implicitly through negative samples or stop-gradient operation. As shown in Fig. 1, by adopting different forms of input-target paired data and their representations, our framework can include existing pre-training approaches and provide possible direc- tions to design an all-in-one pre-training method. To meet the requirement of large-scale pre-training with various data sources, we further extend our framework to the multi-input multi-target setting, with which we showthat multi-task pre-training methods are optimizing a lower bound of the mutual information. In addition, we mix two masked views from two different images as the in- put. The representation of one image is used to recon- struct the same view, while the other image is used to re- construct a different augmented view. Both representa- tions will predict their corresponding annotated category or paired texts. In this way, we propose a novel pre-training approach, called M3I Pre-training, which can effectively combine the merits of supervised/weakly-supervised/self- supervised pre-training and enables large-scale vision foun- dation models to benefit from multi-modal/source large- scale data. Our contributions can be summarized as follows: • We theoretically demonstrate all existing mainstream pre-training methods share a common optimization ob- jective, i.e.,maximizing the mutual information between input and target representation. We also show how to in- stantiate our framework as distinct pre-training methods. • We propose a novel single-stage pre-training approach called M3I Pre-training to gather the benefit of various pre-training supervision signals, via extending our mu- tual information pre-training framework to a multi-input multi-target setting. • Comprehensive experiments demonstrate the effective- ness of our approach. We successfully pre-train InternImage-H [79], a model with billion-level parame- ters, and set a new record on basic detection and segmen- tation tasks, i.e.,65.4 box AP on COCO test-dev [46], 62.9 mIoU on ADE20K [98].

Tseng_EDGE_Editable_Dance_Generation_From_Music_CVPR_2023

Abstract Dance is an important human art form, but creating new dances can be difficult and time-consuming. In this work, we introduce Editable Dance GEneration (EDGE), a state-of-the-art method for editable dance generation that is capable of creating realistic, physically-plausible dances while remaining faithful to the input music. EDGE uses a transformer-based diffusion model paired with Jukebox, a strong music feature extractor, and confers powerful editing capabilities well-suited to dance, including joint-wise con- ditioning, and in-betweening. We introduce a new metric for physical plausibility, and evaluate dance quality gener- ated by our method extensively through (1) multiple quanti- tative metrics on physical plausibility, beat alignment, and diversity benchmarks, and more importantly, (2) a large- scale user study, demonstrating a significant improvement over previous state-of-the-art methods. Qualitative samples from our model can be found at our website.

1. Introduction Dance is an important part of many cultures around the world: it is a form of expression, communication, and social interaction [29]. However, creating new dances or dance an- imations is uniquely difficult because dance movements are expressive and freeform, yet precisely structured by music. In practice, this requires tedious hand animation or motion capture solutions, which can be expensive and impractical.On the other hand, using computational methods to gener- ate dances automatically can alleviate the burden of the cre- ation process, leading to many applications: such methods can help animators create new dances or provide interac- tive characters in video games or virtual reality with real- istic and varied movements based on user-provided music. In addition, dance generation can provide insights into the relationship between music and movement, which is an im- portant area of research in neuroscience [2]. Previous work has made significant progress using ma- chine learning-based methods, but has achieved limited suc- cess in generating dances from music that satisfy user con- straints. Furthermore, the evaluation of generated dances is subjective and complex, and existing papers often use quan- titative metrics that we show to be flawed. In this work, we propose Editable Dance GEneration (EDGE), a state-of-the-art method for dance generation that creates realistic, physically-plausible dance motions based on input music. Our method uses a transformer-based dif- fusion model paired with Jukebox, a strong music feature extractor. This unique diffusion-based approach confers powerful editing capabilities well-suited to dance, includ- ing joint-wise conditioning and in-betweening. In addition to the advantages immediately conferred by the modeling choices, we observe flaws with previous metrics and pro- pose a new metric that captures the physical accuracy of ground contact behaviors without explicit physical model- ing. In summary, our contributions are the following: 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. 448 1. We introduce EDGE, a diffusion-based approach for dance generation that combines state-of-the-art perfor- mance with powerful editing capabilities and is able to generate arbitrarily long sequences. EDGE im- proves on previous hand-crafted audio feature extrac- tion strategies by leveraging music audio representa- tions from Jukebox [5], a pre-trained generative model for music that has previously demonstrated strong per- formance on music-specific prediction tasks [3, 7]. 2. We analyze the metrics proposed in previous works and show that they do not accurately represent human- evaluated quality as reported by a large user study. 3. We propose a new approach to eliminating foot-sliding physical implausibilities in generated motions using a novel Contact Consistency Loss, and introduce Phys- ical Foot Contact Score, a simple new acceleration- based quantitative metric for scoring physical plausi- bility of generated kinematic motions that requires no explicit physical modeling. This work is best enjoyed when accompanied by our demo samples. Please see the samples at our website.

Tang_DETR_With_Additional_Global_Aggregation_for_Cross-Domain_Weakly_Supervised_Object_CVPR_2023

Abstract This paper presents a DETR-based method for cross- domain weakly supervised object detection (CDWSOD), aiming at adapting the detector from source to target do- main through weak supervision. We think DETR has strong potential for CDWSOD due to an insight: the encoder and the decoder in DETR are both based on the attention mech- anism and are thus capable of aggregating semantics across the entire image. The aggregation results, i.e., image- level predictions, can naturally exploit the weak supervi- sion for domain alignment. Such motivated, we propose DETR with additional Global Aggregation (DETR-GA), a CDWSOD detector that simultaneously makes “instance- level + image-level” predictions and utilizes “strong + weak“ supervisions. The key point of DETR-GA is very simple: for the encoder / decoder, we respectively add mul- tiple class queries / a foreground query to aggregate the semantics into image-level predictions. Our query-based aggregation has two advantages. First, in the encoder, the weakly-supervised class queries are capable of roughly lo- cating the corresponding positions and excluding the dis- traction from non-relevant regions. Second, through our design, the object queries and the foreground query in the decoder share consensus on the class semantics, therefore making the strong and weak supervision mutually benefit each other for domain alignment. Extensive experiments on four popular cross-domain benchmarks show that DETR- GA significantly improves cross-domain detection accuracy (e.g., 29.0% →79.4% mAP on PASCAL VOC →Clipart all dataset) and advances the states of the art.

1. Introduction The cross-domain problem is a critical challenge for ob- ject detection in real-world applications. Concretely, there is usually a domain gap between the training and testing *Corresponding author: Si Liu object queryfeaturetokenclass queryforeground querypersondog correlatedtransformerencodertransformerdecoder Figure 1. To exploit the weak supervision, DETR-GA aggregates the semantic information across the entire image into image-level predictions. Specifically, DETR-GA adds multiple class queries / a foreground query into the transformer encoder / decoder, respec- tively. The foreground query is correlated with the object queries but has no position embedding. We visualize the attention score of some queries, e.g., “person“ and “dog”. Despite no position su- pervision, each class query / the foreground query attends to the class-specific / foreground regions for semantic aggregation. data. This domain gap significantly compromises the detec- tion accuracy when the detector trained on the source do- main is directly deployed on a novel target domain. To mit- igate the domain gap, existing domain adaptation methods can be categorized into supervised, unsupervised [7,10,39], and weakly supervised approaches [17, 21, 33, 58]. Among the three approaches, we are particularly interested in the weakly supervised one because it requires only image-level annotations and achieves a good trade-off between the adap- tation effect and the annotation cost. Therefore, this paper challenges the cross-domain weakly supervised object de- tection (CDWSOD), aiming at adapting the detector from the source to target domain through weak supervision. We think the DETR-style detector [3,27,67] has high po- tential for solving CDWSOD. In contrast to current CDW- 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. 11422 SOD methods dominated by pure convolutional neural net- work detectors (“CNN detectors”), this paper is the first to explore DETR-style detectors for CDWSOD, to the best of our knowledge. Our optimism for DETR is NOT due to its prevalence or competitive results in generic object detec- tion. In fact, we empirically find the DETR-style detector barely achieves any superiority against CNN detectors for direct cross-domain deployment (Section 4.4). Instead, our motivation is based on the insight, i.e., the DETR-style de- tector has superiority for combining the strong and weak supervision, which is critical for CDWSOD [17, 21, 58]. Generally, CDWSOD requires using weak ( i.e., image- level) supervision on target domain to transfer the knowl- edge from source domain. Therefore, it is essential to sup- plement the detector with image-level prediction capability. We argue that this essential can be well accommodated by two basic components in DETR, i.e., the encoder and the decoder. Both the encoder and the decoder are based on the attention mechanism and thus have strong capability to capture long-range dependencies. This long-range model- ing capability, by its nature, is favorable for aggregating se- mantic information to make image-level predictions. To fully exploit the weak supervision in CDWSOD, this paper proposes DETR with additional Global Aggregation (DETR-GA). DETR-GA adds attention-based global aggre- gation into DETR so as to make image-level predictions, while simultaneously preserving the original instance-level predictions. Basically, DETR uses multiple object queries in the decoder to probe local regions and gives instance- level predictions. Based on DETR, DETR-GA makes two simple and important changes: for the encoder / decoder, it respectively adds multiple class queries / a foreground query to aggregate semantic information across the entire image. The details are explained below: 1) The encoder makes image-level prediction through a novel class query mechanism . Specifically, the encoder adds multiple class queries into its input layer, with each query responsible for an individual class. Each class query probes the entire image to aggregate class-specific informa- tion and predicts whether the corresponding class exists in the image. During training, we use the image-level multi- class label to supervise the class query predictions. Despite NO position supervision, we show these class queries are capable to roughly locate the corresponding po- sition (Fig. 1) and thus exclude the distraction from non- relevant regions. Therefore, our class query mechanism achieves better image-level aggregation effect than the av- erage pooling strategy that is commonly adopted in pure- CNN CDWSOD methods. Empirically, we find this simple component alone brings significant improvement for CDW- SOD, e.g., +20.8 mAP on PASCAL VOC →Clipart test. 2) The decoder gives image-level and instance-level pre- dictions simultaneously through correlated object and fore-ground queries . To this end, we simply remove the position embedding from an object query and use the remained con- tent embedding as the foreground query. The insight for this design is: in a object query, the position embedding en- courages focus on local region [27,31,32], while the content embedding is prone to global responses to all potential fore- ground regions. Therefore, when we remove the position embedding, an object query discards the position bias and becomes a foreground query with global responses (as visu- alized in Fig. 1). Except this difference ( i.e., with or without position embedding), the object queries and the foreground query share all the other elements in the decoder, e.g., the self-attention layer, cross-attention layer. Such correlation encourages them to share consensus on the class semantic and thus benefits the domain alignment along all the classes. Overall, DETR-GA utilizes the weak supervision on the encoder and decoder to transfer the detection capability from source to target domain. Experimental results show that DETR-GA improves cross-domain detection accuracy by a large margin. Our main contributions can be summa- rized as follows: •As the first work to explore DETR-style detector for CDWSOD, this paper reveals that DETR has strong poten- tial for weakly-supervised domain adaptation because its attention mechanism can fully exploit image-level supervi- sion by aggregating semantics across the entire image. •We propose DETR-GA, a CDWSOD detector that si- multaneously makes “instance-level + image-level” predic- tions and can utilize both “strong + weak” supervision. The key point of DETR-GA is the newly-added class queries / foreground query in the encoder / decoder, which promotes global aggregation for image-level prediction. •Extensive experiments on four popular cross-domain benchmarks show that DETR-GA significantly improves CDWSOD accuracy and advances the states of the art. For example, on PASCAL VOC →Clipart all, DETR-GA im- proves the baseline from 29.0% to 79.4% .

Tejankar_Defending_Against_Patch-Based_Backdoor_Attacks_on_Self-Supervised_Learning_CVPR_2023

Abstract Recently, self-supervised learning (SSL) was shown to be vulnerable to patch-based data poisoning backdoor at- tacks. It was shown that an adversary can poison a small part of the unlabeled data so that when a victim trains an SSL model on it, the final model will have a back- door that the adversary can exploit. This work aims to defend self-supervised learning against such attacks. We use a three-step defense pipeline, where we first train a model on the poisoned data. In the second step, our pro- posed defense algorithm (PatchSearch) uses the trained model to search the training data for poisoned samples and removes them from the training set. In the third step, a final model is trained on the cleaned-up training set. Our results show that PatchSearch is an effective defense. As an example, it improves a model’s accuracy on im- ages containing the trigger from 38.2% to 63.7% which is very close to the clean model’s accuracy, 64.6%. More- over, we show that PatchSearch outperforms baselines and state-of-the-art defense approaches including those using additional clean, trusted data. Our code is available at https://github.com/UCDvision/PatchSearch

1. Introduction Self-supervised learning (SSL) promises to free deep learning models from the constraints of human supervision. It allows models to be trained on cheap and abundantly available unlabeled, uncurated data [22]. A model trained this way can then be used as a general feature extractor for various downstream tasks with a small amount of labeled data. However, recent works [8, 44] have shown that un- curated data collection pipelines are vulnerable to data poi- soning backdoor attacks. Data poisoning backdoor attacks on self-supervised learning (SSL) [44] work as follows. An attacker intro- duces “backdoors” in a model by simply injecting a few carefully crafted samples called “poisons” in the unlabeled *Corresponding author <[email protected]> . Work done while interning at Meta AI.training dataset. Poisoning is done by pasting an attacker chosen “trigger” patch on a few images of a “target” cat- egory. A model pre-trained with such a poisoned data be- haves similar to a non-poisoned/clean model in all cases ex- cept when it is shown an image with a trigger. Then, the model will cause a downstream classifier trained on top of it to mis-classify the image as the target category. This types of attacks are sneaky and hard to detect since the trigger is like an attacker’s secret key that unlocks a failure mode of the model. Our goal in this work is to defend SSL models against such attacks. While there have been many works for defending su- pervised learning against backdoor attacks [34], many of them directly rely upon the availability of labels. There- fore, such methods are hard to adopt in SSL where there are no labels. However, a few supervised defenses [5, 28] can be applied to SSL with some modifications. One such de- fense, proposed in [28], shows the effectiveness of applying a strong augmentation like CutMix [55]. Hence, we adopt i-CutMix [33], CutMix modified for SSL, as a baseline. We show that i-CutMix is indeed an effective defense that addi- tionally improves the overall performance of SSL models. Another defense that does not require labels was pro- posed in [44]. While this defense successfully mitigates the attack, its success is dependent on a significant amount of clean, trusted data. However, access to such data may not be practical in many cases. Even if available, the trusted data may have its own set of biases depending on its sources which might bias the defended model. Hence, our goal is to design a defense that does not need any trusted data. In this paper, we propose PatchSearch, which aims at identifying poisoned samples in the training set without ac- cess to trusted data or image labels. One way to identify a poisoned sample is to check if it contains a patch that be- haves similar to the trigger. A characteristic of a trigger is that it occupies a relatively small area in an image ( ≈5%) but heavily influences a model’s output. Hence, we can use an input explanation method like Grad-CAM [45] to high- light the trigger. Note that this idea has been explored in supervised defenses [14, 16]. However, Grad-CAM relies on a supervised classifier which is unavailable in our case. 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. 12239 𝑔 𝑐!𝑘-means𝑋 𝑐"𝑐#𝑐$𝑐" candidate trigger 𝑡%c. calculate poison score of 𝒙𝒊 𝑐!𝑐#𝑐$poison score = 2flip test setoriginalassignmentsflip test set +candidate triggernewassignments originalflippedflipped𝑐"𝑐"𝑐$𝑋&⊕𝑡%𝑋&d. iterative searchrandomly sample 𝒔images from 𝒄𝒊calculatepoisonscoresofsscore 𝒄𝒊as max of all scores in itprune 𝒓fractionfrom 𝑪withleastscorespick 𝒄𝒊from𝑪increment𝒊from 𝟏→|𝑪|rank all images and pick top-𝑘if 𝑪empty𝑪clusters Grad-CAMextractmax 𝑤 ×𝑤𝑥%𝑔 b. get candidate trigger 𝒕𝒊 from 𝒙𝒊assigncluster 𝑋 𝑋(predict 0 top-𝑘 poisonsℎ predict 1ℎ𝑋⊕e. poison classifiera. assign clustersFigure 1. Illustration of PatchSearch. SSL has been shown to group visually similar images in [4, 47], and [44] showed that poisoned images are close to each other in the representation space. Hence, the very first step (a) is to cluster the training dataset. We use clustering to locate the trigger and to efficiently search for poisoned samples. The second step locates the candidate trigger in an image with Grad-CAM (b) and assigns a poison score to it (c). The third step (d) searches for highly poisonous clusters and only scores images in them. This step outputs a few highly poisonous images which are used in the fourth and final step (e) to train a more accurate poison classifier. This is a key problem that we solve with k-means cluster- ing. An SSL model is first trained on the poisoned data and its representations are used to cluster the training set. SSL has been shown to group visually similar images in [4, 47], and [44] showed that poisoned images are close to each other in the representation space. Hence, we expect poi- soned images to be grouped together and the trigger to be the cause of this grouping. Therefore, cluster centers pro- duced by k-means can be used as classifier weights in place of a supervised classifier in Grad-CAM. Finally, once we have located a salient patch, we can quantify how much it behaves like a trigger by pasting it on a few random images, and counting how many of them get assigned to the cluster of the patch (Figure 1 (c)). A similar idea has been explored in [14], but unlike ours, it requires access to trusted data, it is a supervised defense, and it operates during test time. One can use the above process of assigning poison scores to rank the entire training set and then treat the top ranked images as poisonous. However, our experiments show that in some cases, this poison detection system has very low precision at high recalls. Further, processing all images is redundant since only a few samples are actually poisonous. For instance, there can be as few as 650 poisons in a dataset of 127K samples ( ≈0.5%) in our experiments. Hence, we design an iterative search process that focuses on findinghighly poisonous clusters and only scores images in them (Figure 1 (d)). The results of iterative search are a few but highly poisonous samples. These are then used in the next step to build a poison classifier which can identify poisons more accurately. This results in a poison detection system with about 55% precision and about 98% recall in average. In summary, we propose PatchSearch, a novel de- fense algorithm that defends self-supervised learning mod- els against patch based data poisoning backdoor attacks by efficiently identifying and filtering out poisoned samples. We also propose to apply i-CutMix [33] augmentation as a simple but effective defense. Our results indicate that Patch- Search is better than both the trusted data based defense from [44] and i-CutMix. Further, we show that i-CutMix and PatchSearch are complementary to each other and com- bining them results in a model with an improved overall performance while significantly mitigating the attack.

Tan_Learning_on_Gradients_Generalized_Artifacts_Representation_for_GAN-Generated_Images_Detection_CVPR_2023

Abstract Recently, there has been a significant advancement in image generation technology, known as GAN. It can easily generate realistic fake images, leading to an increased risk of abuse. However, most image detectors suffer from sharp performance drops in unseen domains. The key of fake im- age detection is to develop a generalized representation to describe the artifacts produced by generation models. In this work, we introduce a novel detection framework, named Learning on Gradients (LGrad), designed for identifying GAN-generated images, with the aim of constructing a gen- eralized detector with cross-model and cross-data. Specif- ically, a pretrained CNN model is employed as a trans- formation model to convert images into gradients. Sub- sequently, we leverage these gradients to present the gen- eralized artifacts, which are fed into the classifier to as- certain the authenticity of the images. In our framework, we turn the data-dependent problem into a transformation- model-dependent problem. To the best of our knowledge, this is the first study to utilize gradients as the representa- tion of artifacts in GAN-generated images. Extensive ex- periments demonstrate the effectiveness and robustness of gradients as generalized artifact representations. Our de- tector achieves a new state-of-the-art performance with a remarkable gain of 11.4%. The code is released at https: //github.com/chuangchuangtan/LGrad .

1. Introduction Over the past years, remarkable progress has been made in deep generative models, i.e.generative adversarial net- works (GAN) [14], its variations [3, 21–23, 36], and V AE [25]. The generated media is highly realistic and indistin- *Corresponding author StyleGAN CelebaHQ (a) Images (b)Gradients (c)Grad-R (d)Grad-G (e)Grad-B StyleGAN2 ProGANFigure 1. Visualization of gradients of real images and GAN- generated images extracted from a pre-trained model. To fully understand the gradients, heatmaps for R, G, and B channels also are shown, where red is high. In the gradients, the content of im- ages is filtered out, and only the discriminative pixels are retained for the pre-trained model’s target task. We utilize the gradients as the generalized artifacts representation to develop a novel detec- tion framework. guishable from real to human eyes. Although it has the po- tential for many novel applications [8], it also brings new risks due to the abuse of fake information. The misuse of DeepFakes has been confirmed that some people swap the faces of women onto pornographic videos [8]. In addi- tion, some individuals, even non-experts, can malevolently manipulate or create fake images or videos for political or economic purposes, leading to serious social problems [17]. Thus, it is extremely necessary to develop forgery detection techniques to help people determine the credibility 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. 12105 media [17, 40]. Various detectors [11–13,16,18,19,30,42,45] have been developed to detect GAN-generated images. Some stud- ies [16, 30, 45] focus on human face images, while oth- ers [11,18,19,42] handle various categories of images. They mainly depend on local regions artifacts [4, 44], blending boundary [26], global textures [30], and frequency-level ar- tifacts [13, 18, 19, 45]. However, those methods heavily rely on the training settings, resulting in failure detection of images from unseen categories or GAN models. The test images in the actual scene are always from unknown sources [17], rendering it challenging to develop gener- alized detectors. There are some works [19, 42] exploit- ing pre-processing, data augmentation, and reducing the ef- fects of frequency-level artifacts to develop a robust detec- tor. Nevertheless, there still needs to be a more generalized representation of the clue produced by generation models, which is critical for robust fake image detection. To tackle this problem, we propose a novel and simple detection framework, referred to as Learning on Gradients (LGrad). A new generalized feature, Gradients, is devel- oped to serve as a representation of the artifacts produced by GAN models. We believe that the gradients of a trained CNN model can highlight the important pixels in the tar- get task, thereby serving as a valuable cue for detecting fake images. As shown in Figure 1, we adopt a pre-trained discriminator of ProGAN [21] to extract gradients of im- ages produced by Celeba-HQ [21], ProGAN [21], Style- GAN [22], StyleGAN2 [23]. In these gradients, the content of images is filtered out, and only the discriminative pixels that are relevant to the pre-trained model’s target task are re- tained. Therefore, the gradients are more dependent on the pre-trained model rather than on training sources, thereby enhancing the detector’s performance with unseen data. In our framework, a pretrained model, called transformation model , is employed to convert images to gradients. These gradients serve as the generalized artifacts and are fed into the classifier to obtain a robust detector. Since the transfor- mation model is indeterminate in our framework, targeted anti-detection cannot be effectively launched. To validate the performance of our LGrad, we only use images generated by ProGAN to train the detector and eval- uate it with various sources, including cross-category, cross- model, and cross-model & category. Numerous experi- ments prove the effectiveness and robustness of gradients as generalized artifacts. Our detector achieves a new state- of-the-art performance in known and unseen settings. Our paper makes the following contributions: • We develop a new detection framework, Learning on Gradient (LGrad), to detect GAN-generated images. Our detector achieves a new state-of-the-art perfor- mance.• We introduce a new generalized artifact representation, Gradients, for GAN-generated image detection. Fur- thermore, we are the first to use gradients as the repre- sentation of artifacts. • Our framework turns the data-driven problem into a transformation-model-driven problem. The robustness of the detector is improved with the introduction of the transformation model. • We prove the great potential of the discriminator of GAN in detecting GAN-generated images.

Tang_HumanBench_Towards_General_Human-Centric_Perception_With_Projector_Assisted_Pretraining_CVPR_2023

Abstract Human-centric perceptions include a variety of vision tasks, which have widespread industrial applications, in- cluding surveillance, autonomous driving, and the meta- verse. It is desirable to have a general pretrain model for versatile human-centric downstream tasks. This pa- per forges ahead along this path from the aspects of both benchmark and pretraining methods. Specifically, we pro- pose a HumanBench based on existing datasets to com- prehensively evaluate on the common ground the gener- alization abilities of different pretraining methods on 19 datasets from 6 diverse downstream tasks, including person ReID, pose estimation, human parsing, pedestrian attribute recognition, pedestrian detection, and crowd counting. To learn both coarse-grained and fine-grained knowledge in human bodies, we further propose a Projector AssisTed Hierarchical pretraining method ( PATH ) to learn diverse knowledge at different granularity levels. Comprehensive evaluations on HumanBench show that our PATH achieves new state-of-the-art results on 17 downstream datasets and on-par results on the other 2 datasets. The code will be publicly at https://github.com/OpenGVLab/HumanBench.

1. Introduction Human-centric perception has been a long-standing pur- suit for computer vision and machine learning communi- ties. It encompasses massive research tasks and applica- tions including person ReID in surveillance [16, 17, 60, 96, 112], human parsing and pose estimation in the meta- verse [47, 48, 61, 79, 90, 92], and pedestrian detection in au- tonomous driving [8, 51, 87]. Although significant progress has been made, most existing human-centric studies and pipelines are task-specific for better performances, leading *Equal contribution. This work was done in SenseTime. †Corresponding author.to huge costs in representation/network design, pretraining, parameter-tuning, and annotations. To promote real-world deployment, we ask: whether a general human-centric pre- training model can be developed that can benefit diverse human-centric tasks and be efficiently adapted to down- stream tasks? Intuitively, we argue that pretraining such general human-centric models is possible for two reasons. First, there are obvious correlations among different human- centric tasks. For example, both human parsing and pose es- timation predict the fine-grained parts of human bodies [29, 49] with differences in annotation granularities. Thus, the annotations in one human-centric task may benefit other human-centric tasks when trained together. Second, recent achievements in foundation models [5, 11, 38, 65, 66, 80] have shown that large-scale deep neural networks ( e.g., transformers [13]) have the flexibility to handle diverse in- put modalities and the capacity to deal with different tasks. For example, Uni-Percevier [115] and BEITv3 [85] are ap- plicable to multiple vision and language tasks. Despite the opportunities of processing multiple human- centric tasks with one pretraining model, there are two ob- stacles for developing general human-centric pretraining models. First, although there are many benchmarks for ev- ery single human-centric task, there is still no benchmark to fairly and comprehensively compare various pretraining methods on a common ground for a broad range of human- centric tasks, data distributions, and application scenarios. Second, different from most existing general foundation models trained by unified global vision-language consis- tencies, pretraining human-centric models are required to learn both global ( e.g., person ReID and pedestrian detec- tion) and fine-grained semantic features ( e.g., pose estima- tion and human parsing) of human bodies from diverse an- notation granularity simultaneously. In this paper, we first build a benchmark, called Hu- manBench , based on existing datasets to enable pretrain- 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. 21970 MAECLIPSoTAOurs(a) Diversity of ImagesDay (b) Comprehensiveness of Evaluation Scene imagesPerson-centric images NightCrowdOutdoorIndoor Detection Parsing Pose Attribute reIDCounting(c) High Performance by our pretraining method Figure 1. (a-b) Overview of our proposed HumanBench. HumanBench includes diverse images, including scene images and person-centric images. Our HumanBench also has comprehensive evaluation. Specifically, it evaluates pretraining models on 6 tasks, including pedestrian detection, human parsing, pose estimation, pedestrian attribute recognition, person ReID, and crowd counting. (c) High performances are achieved by our pretraining method on HumanBench. We report 1-heavy occluded MR−2and 1-EPE for Caltech and H3.6pose. ing and evaluating human-centric representations that can be generalized to various downstream tasks. HumanBench has two appealing properties. (1) Diversity. The images in our HumanBench include diverse image properties, rang- ing from person-centric cropped images to scene images with crowd pedestrians, ranging from indoor scenes to out- door scenes (Fig. 1(a)), and from surveillance to metaverse. (2) Comprehensiveness. Humanbench covers comprehen- sive image-based human-centric tasks in both pretraining datasets and downstream tasks (Fig. 1(b)). For pretrain- ing, we include 11 million images from 37 datasets across five representative human-centric tasks, i.e.,person ReID, pose estimation, human parsing, pedestrian attribute recog- nition, and pedestrian detection. For evaluation, Human- Bench evaluates the generalization abilities on 12 pretrain- ing datasets, 6 unseen datasets of pretraining tasks, and 2 datasets out of pretraining tasks, ranging from global pre- diction, i.e., ReID, to local prediction, i.e., human pars- ing and pose estimation. Results on our HumanBench (Fig. 1(c)) lead to two interesting findings. First, compared with datasets with natural images for general pretrained models, HumanBench is more effective for human-centric perception tasks. Second, as human-centric pretraining re- quires to learn features of diverse granularity, supervised pretraining methods with proper designs can learn from di- verse annotations in HumanBench and perform better than the existing unsupervised pretraining methods, for which details will be shown in Sec. 5.3. Based on HumanBench, we further investigate how to learn a better human-centric supervised pretraining model from diverse datasets with various annotations. How- ever, naive multitask pretraining may easily suffer from the task conflicts [53, 97] or overfitting to pretrained annota- tions [67, 107], losing the desirable generalization ability of pretraining. Inspired by [86], which suggests adding anMLP projector before the task head can significantly en- hance the generalization ability of supervised pretraining, we propose Projector AssisTedHierarchical Pre-training (PATH ), a projector assisted pretraining method with hi- erarchical weight sharing to tackle the task conflicts of su- pervised pretraining from diverse annotations. Specifically, the weights of backbones are shared among all datasets, and the weights of projectors are shared only for datasets of the same tasks, while the weights of the heads are shared only for a single dataset – forming a hierarchical weight-sharing structure. During the pretraining stage, we insert the task- specific projectors before dataset heads but discard them when evaluating models on downstream tasks. With the hi- erarchical weight-sharing strategy, our pretraining method enforces the backbone to learn the shared knowledge pool, the projector to attend to the task-specific knowledge, and the head to focus on the dataset with specific annotation and data distribution. In summary, our contributions are two folds: (1) we build HumanBench, a large-scale dataset for human-centric pretraining including diverse images and comprehensive evaluations. (2) To tackle the diversity of input images and annotations of various human-centric datasets, we pro- pose PATH, a projector-assisted hierarchical weight-sharing method for pretraining the general human-centric represen- tations. We achieve state-of-the-art results by PATH on 15 datasets throughout 6 downstream tasks (Fig. 1(c)), on- par results on 2 datasets, and slightly lower results on 2 datasets on HumanBench when using ViT-Base. Experi- ments with ViT-Large backbone show that our method can further achieve considerable gains over ViT-Base, achiev- ing another 2 new state-of-the-art results and showing the promising scalability of our method. We hope our work can shed light on future research on pretraining human-centric representations, such as unified structures. 21971 Table 1. Statistics of Pretraining Datasets Task Number of datasets Number of Images Person ReID 7 5,446,419 Pose estimation 11 3,739,291 Human parsing 7 1,419,910 Pedestrian Attribute 6 242,880 Pedestrian Detection 6 170,687 In total 37 11,019,187

Song_Efficient_Hierarchical_Entropy_Model_for_Learned_Point_Cloud_Compression_CVPR_2023

Abstract Learning an accurate entropy model is a fundamental way to remove the redundancy in point cloud compression. Recently, the octree-based auto-regressive entropy model which adopts the self-attention mechanism to explore de- pendencies in a large-scale context is proved to be promis- ing. However, heavy global attention computations and auto-regressive contexts are inefficient for practical appli- cations. To improve the efficiency of the attention model, we propose a hierarchical attention structure that has a lin- ear complexity to the context scale and maintains the global receptive field. Furthermore, we present a grouped context structure to address the serial decoding issue caused by the auto-regression while preserving the compression perfor- mance. Experiments demonstrate that the proposed entropy model achieves superior rate-distortion performance and significant decoding latency reduction compared with the state-of-the-art large-scale auto-regressive entropy model.

1. Introduction Point cloud is a fundamental data structure to represent 3D scenes. It has been widely applied in 3D vision systems such as autonomous driving and immersive applications. The large-scale point cloud typically contains millions of points [36]. It is challenging to store and transmit such massive data. Hence, efficient point cloud compression that reduces memory footprints and transmission bandwidth is necessary to develop practical point cloud applications. Recently, deep learning methods have promoted the de- velopment of point cloud compression [4, 9, 10, 16, 31, 36, 40,43]. It is a common pipeline to learn an octree-based en- tropy model to estimate octree node symbol ( i.e., occupancy symbol) distributions. The point cloud is first organized as an octree, and occupancy symbols are then encoded into the bitstream losslessly by an entropy coder ( e.g., arithmetic *Corresponding author. 0.1 1 10 100 1000 Decoding Latency (s)-10-5051015202530354045Bitrate Reduction over G-PCC (%)EHEM Light EHEM OctAttentionSparsePCGC OctSqueezeG-PCCFigure 1. Bitrate and decoding speed in the log scale for lossless compression on 16-bit quantized SemanticKITTI. The proposed method EHEM yields state-of-the-art performance with a compa- rable decoding latency to the efficient traditional method G-PCC. coder [47]). An accurate entropy model is required since it reduces the cross entropy between the estimated distribution and ground truth, which is corresponding to actual bitrates. Various attempts have been made to improve the accuracy by designing different context structures [4,10,16,37]. The key of these advances is to increase the context capacity and introduce references from high-resolution octree represen- tations. For example, the context in OctAttention [10] in- cludes hundreds of previously decoded siblings ( i.e., nodes at the same octree level). The large-scale context incorpo- rates more references for the entropy coder, and the high- resolution context preserves detailed features of the point cloud. Both of them contribute to building informative con- texts and effective entropy models. However, large-scale context requires heavy computa- tions to model dependencies among numerous references. The previous work [10] uses the global self-attention mech- anism to model long-range dependencies within the con- text, and its complexity is quadratic to the length 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. 14368 large-scale context. Furthermore, considering the efficiency issue, it is infeasible to build a deeper entropy model or extend the context scale to enhance the modeling capabil- ity based on global attention. Another concern is the serial decoding process caused by the inclusion of previously de- coded siblings. This auto-regressive context structure incurs a practically unacceptable decoding latency ( e.g., around 700 seconds per frame). To address these issues, we build an entropy model with an efficient hierarchical attention model and a parallel- friendly grouped context structure. The hierarchical atten- tion model partitions the context into local windows and computes attention within these windows independently. Therefore, the complexity is linear to the context scale, which allows the further extension of network depth and context capacity to improve performance. Since the recep- tive field of the localized attention is limited, we adopt a multi-scale network structure to query features across dif- ferent windows. The context is progressively downsam- pled by merging neighboring nodes to generate a new to- ken. Then, cross-window dependencies can be captured by incorporating these new tokens in the same window. The grouped context divides the occupancy symbol sequence into two groups. Each group is conditioned on ancestral features and previously decoded groups, and hence nodes in the same group can be coded in parallel. Furthermore, in contrast to the previous auto-regressive context that only ex- ploits causal parts of the ancestral context [10], the grouped context allows to make use of a complete ancestral context. The proposed efficient hierarchical entropy model called EHEM is evaluated on SemanticKITTI [3] and Ford [32] benchmarks. It surpasses state-of-the-art methods in terms of both compression performance and efficiency. Contribu- tions of this work can be summarized from the following perspectives: • We propose a hierarchical attention model, which yields improved compression performance by extend- ing model and context while keeping the efficiency. • We design a grouped context structure that enables par- allel decoding. It adapts the entropy model using high- resolution references to practical applications. • The proposed method achieves state-of-the-art RD per- formance with a practically applicable coding speed.

Tu_Learning_With_Noisy_Labels_via_Self-Supervised_Adversarial_Noisy_Masking_CVPR_2023

Abstract Collecting large-scale datasets is crucial for training deep models, annotating the data, however, inevitably yields noisy labels, which poses challenges to deep learning algo- rithms. Previous efforts tend to mitigate this problem via identifying and removing noisy samples or correcting their labels according to the statistical properties (e.g., loss val- ues) among training samples. In this paper, we aim to tackle this problem from a new perspective, delving into the deep feature maps, we empirically find that models trained with clean and mislabeled samples manifest distinguishable acti- vation feature distributions. From this observation, a novel robust training approach termed adversarial noisy mask- ing is proposed. The idea is to regularize deep features with a label quality guided masking scheme, which adap- tively modulates the input data and label simultaneously, preventing the model to overfit noisy samples. Further, an auxiliary task is designed to reconstruct input data, it nat- urally provides noise-free self-supervised signals to rein- force the generalization ability of models. The proposed method is simple yet effective, it is tested on synthetic and real-world noisy datasets, where significant improvements are obtained over previous methods. Code is available at https://github.com/yuanpengtu/SANM .

1. Introduction Deep learning has achieved remarkable success, relying on large-scale datasets with human-annotated accurate la- bels. However, collecting such high-quality labels is time- consuming and expensive. As an alternative, inexpensive strategies are usually used for generating labels for large- scale samples, such as web crawling, leveraging search en- gines, or using machine-generated annotations. All these * Equal contribution. †Corresponding author. (a)Clean-trained Mis-predicted(b) Noisy-trained Mis-predicted(c) Noisy-trained Correctly-predicted Figure 1. Activation maps of mis-predicted (a-b) and correctly- predicted (c) samples when training PreAct ResNet-18 with clean (i.e., clean-trained) and noisy (i.e., noisy-trained) data on CIFAR- 10 (1st and 2nd row) and Clothing1M [38] (3rd row). alternative methods inevitably yield numerous noisy sam- ples. However, previous research [2] has revealed that deep networks can easily overfit to noisy labels and suffer from dramatic degradation in the generalization performance. Towards this problem, numerous Learning with Noisy labels (LNL) approaches have been proposed. Sample se- lection methods [3, 8, 34, 41] are the most straightforward methods, which attempt to select clean samples based on certain criterion (e.g., loss value) and then reweight or dis- card the noisy instances to reduce the interference of mis- labeled training samples. However, these methods fail to leverage the potential information of the discarded samples. Similarly, label correction based methods [32, 40, 48] try to correct labels, which often impose assumptions on the existence of a small correctly-labeled validation subset or directly utilize predictions of deep models. However, such assumptions can not be always full-filled in real-world noisy datasets [27], leading to limited application scenarios. Be- sides, the predictions of deep model tend to fluctuate when training with mislabeled samples, making the label correc- tion process unstable and sub-optimal [36]. On the contrary, regularization based methods [2, 6, 15, 20, 26, 29, 43, 47, 50] aim to alleviate the side effect of label noise by prevent- ing the deep models from overfitting to all training sam- 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. 16186 ples, which is flexible and can work collaboratively with other LNL methods. Both explicit and implicit regulariza- tion techniques were proposed, the former generally cali- brates the parameter update of networks by modifying the expected training loss, i,e., dropout. The latter focuses on improving the generalization by utilizing the stochasticity, i,e., data augmentation strategy and stochastic gradient de- scent [29]. However, most of these methods are designed for general fully-supervised learning tasks with correctly- labeled samples, and poor generalization ability could be obtained when the label noise is severe [29]. In this paper, we present a novel regularization based method specialized for LNL problem. Our method is in- spired by an observation that some differences exist in activation maps between the predictions of deep models trained with clean and noisy labels. As shown in Fig. 1 (a), activation maps of mis-predicted samples by the model trained with clean labels are always focused on their fore- ground areas. By contrast, when the model is trained with noisy labels, it tends to generates results focused on the meaningless background area for the mis-predicted sam- ples (Fig. 1 (b)). And even for cases that the prediction is correct, i.e., Fig. 1 (c), the high-activated region drifts to irrelevant (i.e., edge) areas of the object, rather than the main regions of the object. This indicates that model trained with mislabeled samples is likely to overfit to some corner parts of the object from noisy data or remember the less- informative regions (i.e., background) . Therefore, from this observation, we propose a novel Self-supervised Adversar- ial Noisy Masking (SANM) method for LNL. The idea is to explicitly regularize the model to generate more diverse activation maps through masking the certain region (e.g., max-activated) of the input image and thus alleviates the confirmation bias [8]. Moreover, to investigate how the masking strategies af- fect the performance of training deep models with noisy labels, we conduct a simple experiment, where differ- ent adversarial masking strategies (i.e., masking the max- activated region of input images with fixed, random, and noise-aware mask ratio) were applied to the popular LNL framework DivideMix [15]. As shown in Fig. 2, we find that the performance gain is much more sensitive to the masking strategy. A fixed or random mask ratio obtain only marginal gain, but when dealing with the correctly-labeled and misla- beled samples with different mask ratios, a significant per- formance gain is achieved. This indicates that how to design an effective adversarial masking method in the case of LNL is non-trivial and not well addressed. Towards the problem above, a label quality guided adap- tive masking strategy is proposed, which modulates the im- age label and the mask ratio of image masking simultane- ously. The label is updated via first leveraging the soft- distributed model prediction and then reducing the proba-bility of max-activated class with the noise-aware masking ratio, while at the same time lifting the probability of other classes. Intuitively, a sample with a high probability of be- ing mislabeled will possess a larger masking ratio, leading to a more strictly regularized input image and modulated label. Therefore, the negative impact of noisy labels can be greatly reduced. As for the correctly-labeled samples, the masking ratio is relatively small, which plays the role of general regularization strategy and improves the model gen- eralization ability by preventing the model from overfitting training samples. Further, we specially customized an aux- iliary decode branch to reconstruct the original input image, which provides noise-free self-supervised information for learning a robust deep model. The proposed SANM is flex- ible and can further boost existing LNL methods. It is tested on synthetic and real-world large-scale noisy datasets, and elaborately ablative experiments were conducted to verify each design component of the proposed method. In a nut- shell, the key contributions of this work are: •We propose a novel self-supervised adversarial noisy masking method named SANM to explicitly impose reg- ularization for LNL problem, preventing the model from overfitting to some corner parts of the object or less- informative regions from noisy data; •A label quality guided masking strategy is proposed to differently adjust the process for clean and noisy sam- ples according to the label quality estimation. This strategy modulates the image label and the ratio of image masking simultaneously; •A self-supervised mask reconstruction auxiliary task is designed to reconstruct the original images based on the features of masked ones, which aims at enhancing general- ization by providing noise-free supervision signals.

Su_Language_Adaptive_Weight_Generation_for_Multi-Task_Visual_Grounding_CVPR_2023

Abstract Although the impressive performance in visual ground- ing, the prevailing approaches usually exploit the visual backbone in a passive way, i.e., the visual backbone ex- tracts features with fixed weights without expression-related hints. The passive perception may lead to mismatches (e.g., redundant and missing), limiting further performance im- provement. Ideally, the visual backbone should actively extract visual features since the expressions already pro- vide the blueprint of desired visual features. The active perception can take expressions as priors to extract rel- evant visual features, which can effectively alleviate the mismatches. Inspired by this, we propose an active per- ception VisualGrounding framework based on Language Adaptive Weights, called VG-LAW. The visual backbone serves as an expression-specific feature extractor through dynamic weights generated for various expressions. Ben- efiting from the specific and relevant visual features ex- tracted from the language-aware visual backbone, VG-LAW does not require additional modules for cross-modal inter- action. Along with a neat multi-task head, VG-LAW can be competent in referring expression comprehension and seg- mentation jointly. Extensive experiments on four represen- tative datasets, i.e., RefCOCO, RefCOCO+, RefCOCOg, and ReferItGame, validate the effectiveness of the proposed framework and demonstrate state-of-the-art performance.

1. Introduction Visual grounding (such as referring expression compre- hension [4, 23, 42, 45, 46, 48, 50], referring expression seg- mentation [6, 14, 17, 23, 32, 33, 44], and phrase grounding [4, 23, 50]) aims to detect or segment the specific object *corresponding author. Linguistic Backbone Visual Backbone F ( I ; W , A ) Visual Backbone F ( I ; W , A ) Linguistic Backbone Visual Backbone F ( I ; W , A ) Cross - modal Interaction Task - specific Head Cross - modal Interaction Task - specific Head Multi - task Head ( a ) ( b ) ( c ) “ white cake with cherries ” weights stage 1 stage 2 stage 3 stage 4 layer N - 1 layer N layer 1 layer 2 Linguistic Backbone “ white cake with cherries ” “ white cake with cherries ” A : architecture I : image W : weights stage 4 stage 3 Stage 2 stage 1 Figure 1. The comparison of visual grounding frameworks. (a) The visual and linguistic backbone independently extracts fea- tures, which are fused through cross-modal interaction. (b) Ad- ditional designed modules are inserted into the visual backbone to modulate visual features using linguistic features. (c) VG-LAW can generate language-adaptive weights for the visual backbone and directly output referred objects through our designed multi- task head without additional cross-modal interaction modules. based on a given natural language description. Compared to general object detection [38] or instance segmentation [11], which can only locate objects within a predefined and fixed category set, visual grounding is more flexible and purpose- ful. Free-formed language descriptions can specify specific visual properties of the target object, such as categories, at- tributes, relationships with other objects, relative/absolute positions, and etc. Due to the similarity with detection tasks, previous vi- sual grounding approaches [23, 33, 46, 50] usually follow the general object detection frameworks [1,11,37], and pay 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. 10857 (a) (b) (c) "white bird" (d) "right bird"Figure 2. Attention visualization of the visual backbone with dif- ferent weights. (a) input image, (b) visual backbone with fixed weights, (c) and (d) visual backbone with weights generated for “white bird” and “right bird”, respectively. attention to the design of cross-modal interaction modules. Despite achieving impressive performance, the visual back- bone is not well explored. Concretely, the visual backbone passively extracts visual features with fixed architecture and weights, regardless of the referring expressions, as illus- trated in Fig. 1 (a). Such passive feature extraction may lead to mismatches between the extracted visual features and those required for various referring expressions, such as missing or redundant features. Taking Fig. 2 as an exam- ple, the fixed visual backbone has an inherent preference for the image, as shown in Fig. 2 (b), which may be irrelevant to the referring expression “white bird”. Ideally, the visual backbone should take full advantage of expressions, as the expressions can provide information and tendencies about the desired visual features. Several methods have noticed this phenomenon and pro- posed corresponding solutions, such as QRNet [45], and LA VT [44]. Both methods achieve the expression-aware visual feature extraction by inserting carefully designed interaction modules (such as QD-ATT [45], and PWAN [44]) into the visual backbone, as illustrated in Fig. 1 (b). Concretely, visual features are first extracted and then ad- justed using QD-ATT (channel and spatial attention) or PWAM (transformer-based pixel-word attention) in QR- Net and LA VT at the end of each stage, respectively. Al- though performance improvement with adjusted visual fea- tures, the extract-then-adjust paradigm inevitably contains a large number of feature-extraction components with fixed weights, e.g., the components belonging to the original visual backbone in QRNet and LA VT. Considering that the architecture and weights jointly determine the func- tion of the visual backbone, this paper adopts a simpler and fine-grained scheme that modifies the function of the visual backbone with language-adaptive weights, as illus- trated in Fig. 1 (c). Different from the extract-then-adjust paradigm used by QRNet and LA VT, the visual backbone equipped with language-adaptive weights can directly ex- tract expression-relevant visual features without additional feature-adjustment modules. In this paper, we propose an active perception Visual Grounding framework based on Language Adaptive Weights, called VG-LAW. It can dynamically adjust the behavior of the visual backbone by injecting the informa-tion of referring expressions into the weights. Specifi- cally, VG-LAW first obtains the specific language-adaptive weights for the visual backbone through two successive processes of linguistic feature aggregation and weight gen- eration. Then, the language-aware visual backbone can extract expression-relevant visual features without manu- ally modifying the visual backbone architecture. Since the extracted visual features are highly expression-relevant, cross-modal interaction modules are not required for fur- ther cross-modal fusion, and the entire network architecture is more streamlined. Furthermore, based on the expression- relevant features, we propose a lightweight but neat multi- task prediction head for jointly referring expression com- prehension (REC) and referring expression segmentation (RES) tasks. Extensive experiments on RefCOCO [47], RefCOCO+ [47], RefCOCOg [36], and ReferItGame [19] datasets demonstrate the effectiveness of our method, which achieves state-of-the-art performance. The main contributions can be summarized as follows: • We propose an active perception visual ground- ing framework based on the language adaptive weights, called VG-LAW, which can actively extract expression-relevant visual features without manually modifying the visual backbone architecture. • Benefiting from the active perception of visual feature extraction, we can directly utilize our proposed neat but efficient multi-task head for REC and RES tasks jointly without carefully designed cross-modal inter- action modules. • Extensive experiments demonstrate the effectiveness of our framework, which achieves state-of-the-art per- formance on four widely used datasets, i.e.,RefCOCO, RefCOCO+, RefCOCOg, and ReferItGame.

Tarasiou_ViTs_for_SITS_Vision_Transformers_for_Satellite_Image_Time_Series_CVPR_2023

Abstract In this paper we introduce the Temporo-Spatial Vision Transformer (TSViT), a fully-attentional model for general Satellite Image Time Series (SITS) processing based on the Vision Transformer (ViT). TSViT splits a SITS record into non-overlapping patches in space and time which are tokenized and subsequently processed by a factorized temporo-spatial encoder. We argue, that in contrast to nat- ural images, a temporal-then-spatial factorization is more intuitive for SITS processing and present experimental evi- dence for this claim. Additionally, we enhance the model’s discriminative power by introducing two novel mechanisms for acquisition-time-specific temporal positional encodings and multiple learnable class tokens. The effect of all novel design choices is evaluated through an extensive ablation study. Our proposed architecture achieves state- of-the-art performance, surpassing previous approaches by a significant margin in three publicly available SITS semantic segmentation and classification datasets. All model, training and evaluation codes can be found at https://github.com/michaeltrs/DeepSatModels .

1. Introduction The monitoring of the Earth surface man-made impacts or activities is essential to enable the design of effective in- terventions to increase welfare and resilience of societies. One example is the sector of agriculture in which monitor- ing of crop development can help design optimum strategies aimed at improving the welfare of farmers and resilience of the food production system. The second of United Nations Sustainable Development Goals (SDG) of Ending Hunger relies on increasing the crop productivity and revenues of farmers in poor and developing countries [35] - approxi- mately 2.5 billion people’s livelihoods depend mainly on producing crops [10]. Achieving SDG 2 goals requires to be able to accurately monitor yields and the evolution of culti- vated areas in order to measure the progress towards achiev- ing several goals, as well as to evaluate the effectiveness of different policies or interventions. In the European Union Figure 1. Model and performance overview. (top) TSViT archi- tecture. A more detailed schematic is presented in Fig.4. (bottom) TSViT performance compared with previous arts (Table 2). (EU) the Sentinel for Common Agricultural Policy program (Sen4CAP) [2] focuses on developing tools and analytics to support the verification of direct payments to farmers with underlying environmental conditionalities such as the adop- tion of environmentally-friendly [50] and crop diversifica- tion [51] practices based on real-time monitoring by the European Space Agency’s (ESA) Sentinel high-resolution satellite constellation [1] to complement on site verifica- tion. Recently, the volume and diversity of space-borne Earth Observation (EO) data [63] and post-processing tools [18, 61, 70] has increased exponentially. This wealth of resources, in combination with important developments in machine learning for computer vision [20, 28, 53], provides an important opportunity for the development of tools for the automated monitoring of crop development. 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. 10418 Towards more accurate automatic crop type recognition, we introduce TSViT, the first fully-attentional1architecture for general SITS processing. An overview of the proposed architecture can be seen in Fig.1 (top). Our novel design introduces some inductive biases that make TSViT particu- larly suitable for the target domain: • Satellite imagery for monitoring land surface variabil- ity boast a high revisit time leading to long temporal sequences. To reduce the amount of computation we factorize input dimensions into their temporal and spa- tial components, providing intuition (section 3.4) and experimental evidence (section 4.2) about why the or- der of factorization matters. • TSViT uses a Transformer backbone [64] following the recently proposed ViT framework [13]. As a result, every TSViT layer has a global receptive field in time or space, in contrast to previously proposed convolu- tional and recurrent architectures [14, 24, 40, 45, 49]. • To make our approach more suitable for SITS mod- elling we propose a tokenization scheme for the in- put image timeseries and propose acquisition-time- specific temporal position encodings in order to extract date-aware features and to account for irregularities in SITS acquisition times (section 3.6). • We make modifications to the ViT framework (sec- tion 3.2) to enhance its capacity to gather class-specific evidence which we argue suits the problem at hand and design two custom decoder heads to accommodate both global and dense predictions (section 3.5). Our provided intuitions are tested through extensive abla- tion studies on design parameters presented in section 4.2. Overall, our architecture achieves state-of-the-art perfor- mance in three publicly available datasets for classification and semantic segmentation presented in Table 2 and Fig.1.

Sohn_Visual_Prompt_Tuning_for_Generative_Transfer_Learning_CVPR_2023

Abstract Learning generative image models from various domains efficiently needs transferring knowledge from an image syn- thesis model trained on a large dataset. We present a recipe for learning vision transformers by generative knowledge transfer. We base our framework on generative vision trans- formers representing an image as a sequence of visual to- kens with the autoregressive or non-autoregressive trans- formers. To adapt to a new domain, we employ prompt tun- ing, which prepends learnable tokens called prompts to the image token sequence and introduces a new prompt design for our task. We study on a variety of visual domains with varying amounts of training images. We show the effective- ness of knowledge transfer and a significantly better image generation quality.1

1. Introduction Image synthesis has witnessed tremendous progress re- cently with the advancement of deep generative models [ 2, 1https://github.com/google-research/generative_ transfer12,20,67,69]. An ideal image synthesis system generates diverse, plausible, and novel scenes capturing the appear- ance of objects and depicting their interactions. The success of image synthesis does heavily rely on the availability of a large amount of diverse training data [ 73]. Transfer learning, a cornerstone invention in deep learn- ing, has proven indispensable in an array of computer vision tasks, including classification [ 35], object detection [ 18,19], image segmentation [ 23,24],etc. However, transfer learn- ing is not widely used for image synthesis. While recent efforts have shown success in transferring knowledge from pre-trained Generative Adversarial Network (GAN) mod- els [46,60,71,76], their demonstrations are limited to nar- row visual domains, e.g., faces or cars [ 46,76], as in Fig. 1, or requiring a non-trivial amount of training data [ 60,71] to transfer to out-of-distribution domains. In this work, we approach transfer learning for image synthesis using generative vision transformers, an emerg- ing class of image synthesis models, such as DALL ·E[53], Taming Transformer [ 15], MaskGIT [ 7], CogView [ 13], N¨UWA [ 75], Parti [ 79], among others, which excel in im- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19840 age synthesis tasks. We closely follow the recipe of trans- fer learning for image classification [ 35], in which a source model is first trained on a large dataset ( e.g., ImageNet) and then transferred to a diverse collection of downstream tasks. Except, in our setting, the input and output are reversed and the model generates images from a class label. We present a transfer learning framework using prompt tuning [38,40]. While the technique has been used for trans- fer learning of discriminative models for vision tasks [ 1,29], we appear to be the first to adopt prompt tuning for trans- fer learning of image synthesis . To this end, we propose a parameter-efficient design of a prompt token generator that admits condition variables ( e.g., class), a key for control- lable image synthesis neglected in prompt tuning for dis- criminative transfer [ 29,38]. We also introduce a marquee header prompt that engineers learned prompts to enhance generation diversity while retaining the generation quality. We conduct a large-scale study to understand the me- chanics of transfer learning for generative vision transform- ers. Two types of generative transformers – AutoRegressive (AR) andNon-AutoRegressive (NAR) – are examined. AR transformers ( e.g., DALL ·E[53], Taming Transformer [ 15], Parti [ 79]) generate image tokens sequentially with an autoregressive lang

Tiwary_ORCa_Glossy_Objects_As_Radiance-Field_Cameras_CVPR_2023

Abstract Reflections on glossy objects contain valuable and hidden information about the surrounding environment. By con- verting these objects into cameras, we can unlock exciting applications, including imaging beyond the camera’s field- of-view and from seemingly impossible vantage points, e.g. from reflections on the human eye. However, this task is challenging because reflections depend jointly on object ge- ometry, material properties, the 3D environment, and the observer’s viewing direction. Our approach converts glossy objects with unknown geometry into radiance-field cameras to image the world from the object’s perspective. Our key insight is to convert the object surface into a virtual sensor that captures cast reflections as a 2D projection of the 5D environment radiance field visible to and surrounding the object. We show that recovering the environment radiance fields enables depth and radiance estimation from the ob- ject to its surroundings in addition to beyond field-of-view *Equal contributionnovel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene, but not the observer. Moreover, using the radiance field we can image around occluders caused by close-by objects in the scene. Our method is trained end-to-end on multi-view images of the object and jointly estimates object geometry, diffuse radiance, and the 5D environment radiance field. For more information, visit our website.

1. Introduction Imagine that you’re driving down a city street that is packed with lines of parked cars on both sides. Inspection of the cars’ glass windshields, glossy paint, and plastic re- veal sharp, but faint and distorted views of the surroundings that might be otherwise hidden from you. Humans can infer depth and semantic cues about the occluded areas in the en- vironment by processing reflections visible on reflective ob- jects, internally decomposing the object geometry and radi- ance from the specular radiance being reflected onto it. Our aim is to decompose the object from its reflections to “see” the world from the object’s perspective, effectively turning 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. 20773 Figure 2. ORCa Overview . We jointly estimate the object’s geometry and diffuse along with the environment radiance field estimation through a three-step approach. First, we model the object as a neural implicit surface (a). We model the reflections as probing the environment on virtual viewpoints (b) estimated analytically from surface properties. We model the environment as a radiance field queried on these viewpoints (c). Both neural implicit surface and environment radiance field are trained jointly on multi-view images of the object using a photometric loss. the object into a camera that images its environment. How- ever, reflections pose a long-standing challenge in computer vision as the reflections are a 2D projection of an unknown 3D environment that is distorted based on the shape of the reflector. To capture the 3D world from the object’s perspective, we model the object’s surface as a virtual sensor that cap- tures the 2D projection of a 5D environment radiance field surrounding the object. This environment radiance field consists largely of areas only visible to the observer through the object’s reflections. Our use of environment radiance fields not only enables depth and radiance estimation from the object to its surroundings but also enables beyond field- of-view novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene but not the observer. Unlike conventional ap- proaches that model the environment as a 2D map, our ap- proach models it as a 5D field without assuming the scene is infinitely far away. Moreover, by sampling the 5D radiance field, instead of a 2D map, we can capture depth and images around occluders, such as close-by objects in the scene, as shown in Fig. 3. These applications cannot be done from a 2D environment map. We aim to decompose reflections on the object’s surface, from its surface and exploit those reflections to construct a radiance field surrounding the object, therefore captur- ing the 3D world in the process. This is a challenging task because the reflections are extremely sensitive to local ob- ject geometry, viewing direction and inter-reflections due to the object’s surface. To capture this radiance field, we convert glossy objects with unknown geometry and texture into radiance-field cameras. Specifically, we exploit neural rendering to estimate the local surface of the object viewed Figure 3. Advantages of 5D environment radiance field . Mod- eling reflections on object surfaces (a) as a 5D env. radiance field enables beyond field-of-view novel-view synthesis, including ren- dering of the environment from translated virtual camera views (b). Depth (c) and environment radiance of translated and parallax views can further enable imaging behind occluders, for example revealing the tails behind the primary Pokemon occluders (d). from each pixel of the real camera. We then convert this local surface into a virtual pixel that captures radiance from the environment. This virtual pixel captures the environ- ment radiance as shown in Fig 5. We estimate the outgo- 20774 ing frustum from the virtual pixel as a cone that samples the scene. By sampling the scene from many virtual pix- els on the object surface, we construct an environment ra- diance field that can be queried independently of the object surface, enabling beyond field-of-view novel-view synthesis from previously unsampled viewpoints. Our approach jointly estimates object geometry, diffuse radiance, and the environment radiance field from multi- view images of glossy objects with unknown geometry and diffuse texture in three steps. First, we use neural signed distance functions (SDF) and an MLP to model the glossy object’s geometry as a neural implicit surface and diffuse radiance, respectively, similar to PANDORA [10]. Then, for every pixel on the observer’s camera, we estimate the virtual pixels on the object’s surface based on the estimated local geometry from the neural SDF. We analytically com- pute the parameters of the virtual cone through the virtual pixel. Lastly, we use the cone formulation in MipNeRF [5] to cast virtual cones from the virtual camera to recover the environment radiance . To summarize, we make the following contributions: • We present a method to convert implicit surfaces into virtual sensors that can image their surroundings using virtual cones. (Sec. 3.3) • We jointly estimate object geometry, diffuse radiance, and estimate the 5D environment radiance field sur- rounding the object. (Fig. 7 & 9) • We show that the environment radiance field can be queried to perform beyond-field-of-view novel view- point synthesis, i.e render views only visible to the ob- ject in the scene (Section 3.4) Scope. We only model glossy objects with low rough- ness as such specular reflections tend to have a high signal- to-noise ratio, therefore, are a sharper estimate of the en- vironment radiance field. However, we note that the vir- tual cone computation can be extended to model the cone radius as a function of surface roughness. Deblurring ap- proaches can further improve the resolution of estimated environment. In addition, we approximate the local curva- ture using mean curvature, which fails for objects with vary- ing radius of curvature along the tangent space. We explain how our virtual cone curvature estimation can be extended to handle general shape operators in the supplementary ma- terial. Lastly, similar to other multi-view approaches, our approach relies on a sufficient virtual baseline between vir- tual viewpoints to recover the environment radiance field.

Urooj_Learning_Situation_Hyper-Graphs_for_Video_Question_Answering_CVPR_2023

Abstract Answering questions about complex situations in videos requires not only capturing the presence of actors, objects, and their relations but also the evolution of these relation- ships over time. A situation hyper-graph is a representa- tion that describes situations as scene sub-graphs for video frames and hyper-edges for connected sub-graphs and has been proposed to capture all such information in a compact structured form. In this work, we propose an architecture for Video Question Answering (VQA) that enables answering questions related to video content by predicting situation hyper-graphs, coined Situation Hyper-Graph based Video Question Answering (SHG-VQA). To this end, we train a situation hyper-graph decoder to implicitly identify graph representations with actions and object/human-object rela- tionships from the input video clip. and to use cross-attention between the predicted situation hyper-graphs and the ques- tion embedding to predict the correct answer. The proposed method is trained in an end-to-end manner and optimized by a VQA loss with the cross-entropy function and a Hungarian matching loss for the situation graph prediction. The effec- tiveness of the proposed architecture is extensively evaluated on two challenging benchmarks: AGQA and STAR. Our results show that learning the underlying situation hyper- graphs helps the system to significantly improve its perfor- mance for novel challenges of video question-answering tasks1.

1. Introduction Video question answering in real-world scenarios is a challenging task as it requires focusing on several factors including the perception of the current scene, language un- derstanding, situated reasoning, and future prediction. Visual perception in the reasoning task requires capturing various aspects of visual understanding, e.g., detecting a diverse set 1Code will be available at https://github.com/aurooj/SHG- VQA Figure 1. The situation hyper-graph for a video is composed of situations with entities and their relationships (shown as subgraphs in the pink box). These situations may evolve over time. Temporal actions act as hyper-edges connecting these situations into one situ- ation hyper-graph. Learning situation graphs, as well as temporal actions, is vital for reasoning-based video question answering. of entities, recognizing their interactions, as well as under- standing the changing dynamics between these entities over time. Similarly, linguistic understanding has its challenges as some question or answer concepts may not be present in the input text or video. Visual question answering, as well as its extension over time, video question answering, have both benefited from representing knowledge in graph structures, e.g., scene graphs [20, 35], spatio-temporal graphs [4, 55], and knowl- edge graphs [36, 45]. Another approach in this direction is the re-introduction of the concept of “ situation cognition ” embodied in “ situation hyper-graphs ” [47]. This adds the computation of actions to the graphs that capture the interac- tion between entities. In this case, situations are represented by hyper-graphs that join atomic entities and relations (e.g., agents, objects, and relationships) with their actions (Fig. 1). This is an ambitious task for existing systems as it is impractical to encapsulate all possible interactions in the real-world context. Recent work [23] shows that transformers are capable 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. 14879 of learning graphs without adapting graph-specific details in the architectures achieving competitive or even better performance than sophisticated graph-specific models. Our work supports this idea by implicitly learning the underlying hyper-graphs of a video. Thus, it requires no graph compu- tation for inference and uses decoder’s output directly for cross attention module. More precisely, we propose to learn situation hyper-graphs, namely framewise actor-object and object-object relations as well as their respective actions, from the input video directly without the need for explicit object detection or other required prior knowledge. While the actions capture events across transitions over multiple frames, such as Drinking from a bottle , the relationship en- coding actually considers all possible combinations of static, single frame actor-object, and object-object relationships as unique classes, e.g., in the form of person – hold – bottle orbottle – stands on – table , thus serving as an object and relation classifier. Leveraging this setup allows us to stream- line the spatio-temporal graph learning as a set prediction task for predicting relationship predicates and actions in a Situation hyper-graph Decoder block. To train the Situation Graph Decoder, we use a bipartite matching loss between the predicted set and ground truth hyper-graph tokens. The output of the situation graph decoder is a set of action and relationship tokens, which are then combined with the em- bedding of the associated question to derive the final answer. An overview of the proposed architecture is given in Fig. 2. Note that, compared to other works targeting video scene graph generation, e.g., those listed in [63], we are less fo- cused on learning the best possible scene graph, but rather on learning the representation of the scene which best sup- ports the question answering task. Thus, while capturing the essence of a scene, as well as the transition from one scene to the other, we are not only optimizing the scene graph accuracy but also considering the VQA loss. We evaluate the proposed method on two challenging video question answering benchmarks: a) STAR [47], fea- turing four different question types, interaction, sequence, prediction, and feasibility based on a subset of the real- world Charades dataset [44]; and b) Action Genome QA (AGQA) [11] dataset which tests vision focused reasoning skills based on novel compositions, novel reasoning steps, and indirect references. Compared to other VQA datasets, these datasets provide dense ground truth hyper-graph infor- mation for each video, which allows us to learn the respective embedding. Our results show that the proposed hyper-graph encoding significantly improves VQA performance as it has the ability to infer correct answers from spatio-temporal graphs from the input video. Our ablations further reveal that achieving high-quality graphs can be critical for VQA performance. Our contributions to this paper are as follows: •We introduce a novel architecture that enables the com-putation of situation hyper-graphs from video data to solve the complex reasoning task of video question- answering; •We propose a situation hyper-graph decoder module to decode the atomic actions and object/actor-object relationships and model the hyper-graph learning as a transformer-based set prediction task and use a set pre- diction loss function to predict actions and relationships between entities in the input video; •We use the resulting high-level embedding information as sole visual information for the reasoning and show that this is sufficient for an effective VQA system.

Somasundaram_Role_of_Transients_in_Two-Bounce_Non-Line-of-Sight_Imaging_CVPR_2023

Abstract The goal of non-line-of-sight (NLOS) imaging is to image objects occluded from the camera’s field of view using mul- tiply scattered light. Recent works have demonstrated the feasibility of two-bounce (2B) NLOS imaging by scanning a laser and measuring cast shadows of occluded objects in scenes with two relay surfaces. In this work, we study the role of time-of-flight (ToF) measurements, i.e. transients, in 2B-NLOS under multiplexed illumination. Specifically, we study how ToF information can reduce the number of mea- surements and spatial resolution needed for shape recon- struction. We present our findings with respect to trade- offs in (1) temporal resolution, (2) spatial resolution, and (3) number of image captures by studying SNR and recov- erability as functions of system parameters. This leads to a formal definition of the mathematical constraints for 2B lidar. We believe that our work lays an analytical ground- work for design of future NLOS imaging systems, especially as ToF sensors become increasingly ubiquitous.

1. Introduction Non-line-of-sight (NLOS) imaging aims to reconstruct objects occluded from direct line of sight and has the po- tential to be transformative in numerous applications across autonomous driving, search and rescue, and non-invasive medical imaging [24]. The key approach is to measure light that has undergone multiple surface scattering events and computationally invert these measurements to estimate hid- den geometries. Recent work used two-bounce light, as shown in Fig. 2a, to reconstruct high quality shapes behind occluders [15]. The key idea is that two-bounce light cap- tures information about the shadows of the occluded object. By scanning the laser source at different points lon a relay surface and measuring multiple shadow images, it is possi- ble to reconstruct the hidden object by computing the visual hull [20] of the measured shadows. Benefits of Two-Bounce Two-bounce (2B) light can be captured using two relay surfaces on opposite sides of the hidden scene (Fig. 1a). This can occur in a variety of real- world settings such as tunnels, hallways, streets, and clut- 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. 9192 𝒔!𝒔" Cast shadowsOccluderCameraLaserDiffuse Surface𝒍#𝒍$𝒍"Occluded Object𝒔!𝒔" Mixed shadowsOccluded ObjectOccluderSPADLaserDiffuse Surface𝒍#𝒍$𝒍" 𝒔!%" (a) Two-Bounce NLOS Imaging (b) Two-Bounce Transient NLOS Figure 2. Two-Bounce NLOS Imaging. (a) Two-bounce NLOS imaging performs 3D reconstruction by using information con- tained in the shadows of the hidden object when illuminated by one virtual source lat a time. (b) In this work, we study the bene- fits of using transient information for two-bounce NLOS imaging under the presence of multiplexed illumination. tered environments like forests (Fig. 1b). For example, imagine an autonomous vehicle in a tunnel being able to see ahead of the car in front of it by using two-bounce sig- nals from the sides of the tunnel. Two-bounce light also helps alleviate the signal-to-noise ratio (SNR) issue in 3B- NLOS imaging, since 3B signals experience significant sig- nal attenuation at each surface scatter event. Furthermore, 2B light captures information about the cast shadows of the hidden object instead of reflectance, as in the 3B case. As a result, the SNR of 2B signals is independent of the object’s albedo and can therefore robustly image darker objects. Contributions In this work, we propose using two- bounce transients for NLOS imaging to enable the use of multiplexed illumination. 3B-NLOS transient imaging sys- tems have been analyzed previously, but we believe we are the first to analyze the use of multiplexed illumination and transient measurements for 2B-NLOS (Fig. 2b). Our key in- sight is that multiplexed illumination enables measurement of occluded objects with fewer image captures, and 2B tran- sients enable demultiplexing shadows, as shown in Fig. 3. The robust SNR and promise of few-shot capture afforded by 2B transients makes the idea an important direction for few-shot NLOS imaging in general environments. We sum- marize our contributions below. •Model : We present new insights into the algebraic for- mulations of two-bounce lidar for NLOS imaging. •Analysis : We analyze tradeoffs with spatial resolution, temporal resolution, and number of image captures by using quantitative metrics such as SNR and recover- ability, as well as qualitative validation from real and simulated results. Although we envision the ideas introduced here to inspire future works in single-shot NLOS imaging, we do notclaim this as a contribution in this paper.Scope of This Work All data used for this work is from simulation and a single-pixel SPAD sensor. We do not phys- ically realize few-shot results in this paper due to limited availability of high-resolution SPAD arrays presently. In- stead, we emulate SPAD array measurements by scanning a single-pixel SPAD across the field of view and and emu- late multiplexed illumination by scanning a laser spot, ac- quiring individual transient images, and summing the im- ages in post-processing. These measurements, however, are more conducive to our goal of performing analysis that ex- plores the landscape of possibilities in NLOS imaging with advances in ToF sensors. We believe that the ideas and anal- ysis introduced in this work will be increasingly relevant as SPAD array technology matures [19, 26, 44].

Smith_CODA-Prompt_COntinual_Decomposed_Attention-Based_Prompting_for_Rehearsal-Free_Continual_Learning_CVPR_2023

Abstract Computer vision models suffer from a phenomenon known as catastrophic forgetting when learning novel con- cepts from continuously shifting training data. Typical solu- tions for this continual learning problem require extensive rehearsal of previously seen data, which increases memory costs and may violate data privacy. Recently, the emer- gence of large-scale pre-trained vision transformer mod- els has enabled prompting approaches as an alternative to data-rehearsal. These approaches rely on a key-query mechanism to generate prompts and have been found to be highly resistant to catastrophic forgetting in the well- established rehearsal-free continual learning setting. How- ever, the key mechanism of these methods is not trained end-to-end with the task sequence. Our experiments show that this leads to a reduction in their plasticity, hence sac- rificing new task accuracy, and inability to benefit from ex- panded parameter capacity. We instead propose to learn a set of prompt components which are assembled with input- conditioned weights to produce input-conditioned prompts, resulting in a novel attention-based end-to-end key-query scheme. Our experiments show that we outperform the cur- rent SOTA method DualPrompt on established benchmarks by as much as 4.5% in average final accuracy. We also outperform the state of art by as much as 4.4% accuracy on a continual learning benchmark which contains both class-incremental and domain-incremental task shifts, cor- responding to many practical settings. Our code is avail- able at https://github.com/GT-RIPL/CODA-Prompt

1. Introduction For a computer vision model to succeed in the real world, it must overcome brittle assumptions that the concepts it will encounter after deployment will match those learned a-priori during training. The real world is indeed dynamic and contains continuously emerging objects and categories. Once deployed, models will encounter a range of differ- ences from their training sets, requiring us to continuously 𝚺Query valueprompt key component weightAttentionα1 α2 α3 αMk1 k2 k3 kMv1 vMv2 v3 c1 cMc2 c3× promptModel Model … ………Prior Work Our Workgradient flow from model Querylearned in separate optimization without model gradientsseparate optimizationFigure 1. Prior work [65] learns a pool of key-value pairs to select learnable prompts which are inserted into several layers of a pre- trained ViT (the prompting parameters are unique to each layer). Our work introduces a decomposed prompt which consists of learnable prompt components that assemble to produce attention- conditioned prompts. Unlike prior work, ours is optimized in an end-to-end fashion (denoted with thick, green lines). update them to avoid stale and decaying performance. One way to update a model is to collect additional train- ing data, combine this new training data with the old train- ing data, and then retrain the model from scratch. While this will guarantee high performance, it is not practical for large-scale applications which may require lengthy training times and aggressive replacement timelines, such as models for self-driving cars or social media platforms. This could incur high financial [26] and environmental [33] costs if the replacement is frequent. We could instead update the model by only training on the new training data, but this leads to a phenomenon known as catastrophic forgetting [46] where the model overwrites existing knowledge when learning the new data, a problem known as continual learning . The most effective approaches proposed by the continual 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. 11909 learning community involve saving [50] or generating [55] a subset of past training data and mixing it with future task data, a strategy referred to as rehearsal . Yet many impor- tant applications are unable to store this data because they work with private user data that cannot be stored long term. In this paper, we instead consider the highly-impactful and well-established setting of rehearsal-free continual learning [38, 56, 57, 65, 66]1, limiting our scope to strate- gies for continual learning which do not store training data. While rehearsal-free approaches have classically under-performed rehearsal-based approaches in this chal- lenging setting by wide-margins [57], recent prompt-based approaches [65, 66], leveraging pre-trained Vision trans- formers (ViT), have had tremendous success and can even outperform state-of-the-art (SOTA) rehearsal-based meth- ods. These prompting approaches boast a strong protection against catastrophic forgetting by learning a small pool of insert-able model embeddings (prompts) rather than mod- ifying vision encoder parameters directly. One drawback of these approaches, however, is that they cannot be opti- mized in an end-to-end fashion as they use a key and query to select a prompt index from the pool of prompts, and thus rely on a second, localized optimization to learn the keys because the model gradient cannot backpropagate through the key/query index selection. Furthermore, these meth- ods reduce forgetting by sacrificing new task accuracy (i.e., they lack sufficient plasticity ). We show that expanding the prompt size does not increase the plasticity, motivating us to grow learning capacity from a new perspective. We replace the prompt pool with a decomposed prompt that consists of a weighted sum of learnable prompt com- ponents (Figure 1). This decomposition enables higher prompting capacity by expanding in a new dimension (the number of components). Furthermore, this inherently en- courages knowledge re-use, as future task prompts will in- clude contributions from prior task components. We also introduce a novel attention-based component-weighting scheme, which allows for our entire method to be optimized in an end-to-end fashion unlike the existing works, increas- ing our plasticity to better learn future tasks. We boast sig- nificant performance gains on existing rehearsal-free con- tinual learning benchmarks w.r.t. SOTA prompting base- lines in addition to a challenging dual-shift benchmark con- taining both incremental concept shifts and covariate do- main shifts, highlighting our method’s impact and gener- ality. Our method improves results even under equivalent parameter counts, but importantly can scale performance by increasing capacity, unlike prior methods which quickly plateau. In summary, we make the following contributions: 1We focus on continual learning over a single, expanding classification head called class-incremental continual learning . This is different from the multi-task continual learning setting, known as task-incremental continual learning, where we learn separate classification heads for each task and the task label is provided during inference. [25, 61]1. We introduce a decomposed attention-based prompt- ing for rehearsal-free continual learning, characterized by an expanded learning capacity compared to exist- ing continual prompting approaches. Importantly, our approach can be optimized in an end-to-end fashion, unlike the prior SOTA. 2. We establish a new SOTA for rehearsal-free continual learning on the well-established ImageNet-R [21, 65] and CIFAR-100 [32] benchmarks, beating the previous SOTA method DualPrompt [65] by as much as 4.5%. 3. We evaluate on a challenging benchmark with dual distribution shifts (semantic and covariate) using the ImageNet-R dataset, and again outperform the state of art, highlighting the real-world impact and generality of our approach.

Tian_GeoMAE_Masked_Geometric_Target_Prediction_for_Self-Supervised_Point_Cloud_Pre-Training_CVPR_2023

Abstract This paper tries to address a fundamental question in point cloud self-supervised learning: what is a good sig- nal we should leverage to learn features from point clouds without annotations? To answer that, we introduce a point cloud representation learning framework, based on geo- metric feature reconstruction. In contrast to recent papers that directly adopt masked autoencoder (MAE) and only predict original coordinates or occupancy from masked point clouds, our method revisits differences between im- ages and point clouds and identifies three self-supervised learning objectives peculiar to point clouds, namely cen- troid prediction, normal estimation, and curvature predic- tion. Combined, these three objectives yield an nontrivial self-supervised learning task and mutually facilitate mod- els to better reason fine-grained geometry of point clouds. Our pipeline is conceptually simple and it consists of two major steps: first, it randomly masks out groups of points, followed by a Transformer-based point cloud encoder; sec- ond, a lightweight Transformer decoder predicts centroid, normal, and curvature for points in each voxel. We transfer the pre-trained Transformer encoder to a downstream pe- ception model. On the nuScene Datset, our model achieves 3.38 mAP improvment for object detection, 2.1 mIoU gain for segmentation, and 1.7 AMOTA gain for multi-object tracking. We also conduct experiments on the Waymo Open Dataset and achieve significant performance improvements over baselines as well.1

1. Introduction While object detection and segmentation from LiDAR point clouds have achieved significant progress, these mod- els usually demand a large amount of 3D annotations that are hard to acquire. To alleviate this issue, recent works ex- plore learning representations from unlabeled point clouds, *Corresponding to: [email protected] 1Our code is available at https://github.com/Tsinghua- MARS-Lab/GeoMAE . Pre-trained with geometric featuresPre-trained with point valuesPrediction TargetPerformance + 0.9+ 3.4 ScratchMasked Token Prediction Pixel values: brightness Point values: coordinates Geometric features P Figure 1. Pixel value regression has been proved effective in masked autoencoder pre-training for images. We find this practice ineffective in point cloud pre-training and propose a set of geome- try aware prediction targets. such as contrastive learning [20, 46, 53], and mask model- ing [32, 51]. Similar to image-based representation learn- ing settings, these representations are transferred to down- stream tasks for weight initialization. However, the exist- ing self-supervised pretext tasks do not bring adequate im- provements to the downstream tasks as expected. Contrastive learning based methods typically encode dif- ferent ‘views’ (potentially with data augmentation) of point clouds into feature space. They bring features of the same point cloud closer and make features of different point clouds ‘repel’ each other. Other recent works use masked modeling to learn point cloud features through self re- construction [32, 51]. That is, randomly sparsified point clouds are encoded by point cloud feature extractors, fol- lowed by a reconstruction module to predict original point clouds. These methods, when applied to point clouds, ig- nore the fundamental difference of point clouds from im- ages – point clouds provide scene geometry while images provide brightness. As shown in Figure 1, this modality disparity hampers direct use of methods developed in the image domain for point cloud domain, and thus calls for novel self-supervised objectives dedicated to point clouds. Inspired by modeling and computational techniques in geometry processing, we introduce a self-supervised learn- ing framework dedicated to point clouds. Most importantly, we design a series of prediction targets which describe the fine-grained geometric features of the point clouds. These 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. 13570 geometric feature prediction tasks jointly drive models to recognize different shapes and areas of scenes. Concretely, our method starts with a point cloud voxelizer, followed by a feature encoder to transform each voxel into a feature token. These feature tokens are randomly dropped based on a pre-defined mask ratio. Similar to the original MAE work [18], visible tokens are encoded by a Transformer en- coder. Then a Transformer decoder reconstructs the fea- tures of the original voxelized point clouds. Finally, our model predicts point statistics and surface properties in par- allel branches. We conduct experiments on a diverse set of large- scale point cloud datasets including nuScenes [4] and Waymo [39]. Our setting consists of a self-supervised pre- training stage and a downstream task stage (3D detection, 3D tracking, segmentation), where they share the same point cloud backbone. Our results show that even with- out additional unlabeled point clouds, self-supervised pre- training with objectives proposed by this paper can signif- icantly boost the performance of 3D object detection. To summarize, our contributions are: • We introduce geometry aware self-supervised objec- tives for point clouds pre-training. Our method lever- ages fine-grained point statistics and surface properties to enable effective representation learning. • With our novel learning objectives, we achieve state- of-the-art performance compared to previous 3D self- supervised learning methods on a variety of down- stream tasks including 3D object detection, 3D/BEV segmentation, and 3D multi-object tracking. • We conduct comprehensive ablation studies to under- stand the effectiveness of each module and learning objective in our approach.

Tao_Siamese_Image_Modeling_for_Self-Supervised_Vision_Representation_Learning_CVPR_2023

Abstract Self-supervised learning (SSL) has delivered superior performance on a variety of downstream vision tasks. Two main-stream SSL frameworks have been proposed, i.e., In- stance Discrimination (ID) and Masked Image Modeling (MIM). ID pulls together representations from different views of the same image, while avoiding feature collapse. It lacks spatial sensitivity, which requires modeling the lo- cal structure within each image. On the other hand, MIM reconstructs the original content given a masked image. It instead does not have good semantic alignment, which requires projecting semantically similar views into nearby representations. To address this dilemma, we observe that (1) semantic alignment can be achieved by matching dif- ferent image views with strong augmentations; (2) spatial sensitivity can benefit from predicting dense representations with masked images. Driven by these analysis, we propose Siamese Image Modeling (SiameseIM), which predicts the dense representations of an augmented view, based on an- other masked view from the same image but with different augmentations. SiameseIM uses a Siamese network with two branches. The online branch encodes the first view, and predicts the second view’s representation according to the relative positions between these two views. The target branch produces the target by encoding the second view. SiameseIM can surpass both ID and MIM on a wide range of downstream tasks, including ImageNet finetuning and linear probing, COCO and LVIS detection, and ADE20k se- mantic segmentation. The improvement is more significant in few-shot, long-tail and robustness-concerned scenarios. Code shall be released. ∗Equal contribution.†This work is done when Chenxin Tao and Weijie Su are interns at Shanghai Artificial Intelligence Laboratory. BCorresponding author.

1. Introduction Self-supervised learning (SSL) has been pursued in the vision domain for a long time [32]. It enables us to pre- train models without human-annotated labels, which makes it possible to exploit huge amounts of unlabeled data. SSL has provided competitive results against supervised pre- training baselines in various downstream tasks, including image classification [4, 23], object detection [35] and se- mantic segmentation [24]. To effectively train models in the SSL manner, re- searchers design the so-called “pretext tasks” to generate supervision signals. One of the most typical frameworks isInstance Discrimination (ID) , whose core idea is to pull together representations of different augmented views from the same image, and avoid representational collapse. Differ- ent variants of ID have been proposed, including contrastive learning [9, 24], asymmetric networks [12, 21], and feature decorrelation [5, 57]. A recent work [43] has shown the in- trinsic consistency among these methods via their similar gradient structures. For ID methods, the representations of each image are well separated, thus inducing good linear separability. However, as shown in [35], for transfer learn- ing on detection tasks with Vision Transformers [19], ID is not superior to supervised pre-training, and even lags be- hind random initialization given enough training time. Recently, another SSL framework has gradually at- tracted more attention, namely Masked Image Modeling (MIM) [4,23]. MIM methods train the model to reconstruct the original content from a masked image. Such practice can help to learn the rich local structures within an image, leading to excellent performance in dense prediction tasks such as object detection [35]. Nevertheless, MIM does not have good linear separability as ID, and usually performs poorly under the few-shot classification settings [1]. Both ID and MIM methods have their own strengths and This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2132 ImageNet COCO ADE20k LVIS Robustness FT LIN FT 1% APbAPmmIoU APb rare APm rare avg score∗ MoCo-v3 (ID method) 83.0 76.7 63.4 47.9 42.7 47.3 25.5 25.8 43.4 MAE (MIM method) 83.6 68.0 51.1 51.6 45.9 48.1 29.3 29.1 41.8 SiameseIM (ours) 84.1 78.0 65.1 52.1 46.2 51.1 30.9 30.1 47.9 Improve w.r.t. MoCo-v3 +1.1 +1.3 +1.7 +4.2 +3.5 +3.8 +5.4 +4.3 +4.5 Improve w.r.t. MAE +0.5 +10.0 +14.0 +0.5 +0.3 +3.0 +1.6 +1.0 +6.1 Table 1. SiameseIM surpasses MoCo-v3 (ID method) and MAE (MIM method) on a wide range of downstream tasks.∗The robustness average score is calculated by averaging top-1 acc of IN-A, IN-R, IN-S, and 1-mCE of IN-C. For detailed results, please refer to Section 4.2. Masked Image ModelingInstance Discrimination Siamese Image Modeling (ours) spatial augmentationscolor augmentationsmask augmentationsOnline Model … predictionOnline Model prediction Target Model target Online Model Target ModelDense LossGlobal Loss Dense Loss… target … prediction… targetspatial augmentationsmask augmentationsspatial augmentationscolor augmentations Figure 1. Comparisons among ID, MIM and SiameseIM. Matching different augmented views can help to learn semantic alignment, which is adopted by ID and SiameseIM. Predicting dense representations from masked images is beneficial to obtain spatial sensitivity, which is adopted by MIM and SiameseIM. weaknesses. We argue that this dilemma is caused by ne- glecting the representation requirements of either seman- tic alignment or spatial sensitivity. Specifically, MIM op- erates within each image independently, regardless of the inter-image relationship. The representations of semanti- cally similar images are not well aligned, which further re- sults in poor linear probing and few-shot learning perfor- mances of MIM. On the other hand, ID only uses a global representation for the whole image, and thus fails to model the intra-image structure. The spatial sensitivity of features is therefore missing, and ID methods usually produce infe- rior results on dense prediction. To overcome this dilemma, we observe the key factors for semantic alignment and spatial sensitivity: (1) semantic alignment requires that images with similar semantics are projected into nearby representations. This can be achieved by matching different augmented views from the same im- age. Strong augmentations are also beneficial because they provide more invariance to the model; (2) spatial sensitiv- ity needs modeling the local structures within an image. Predicting dense representations from masked images thus helps, because it models the conditional distribution of im- age content within each image. These observations motivate us to predict the dense representations of an image from amasked view with different augmentations. To this end, we propose Siamese Image Modeling (SiameseIM), which reconstructs the dense representations of an augmented view, based on another masked view from the same image but with different augmentations (see Fig. 1). It adopts a Siamese network with an online and a target branch. The online branch consists of an encoder that maps the first masked view into latent representations, and a decoder that reconstructs the representations of the sec- ond view according to the relative positions between these two views. The target branch only contains a momentum encoder that encodes the second view into the prediction target. The encoder is made up of a backbone and a projec- tor. After the pre-training, we only use the online backbone for downstream tasks. As shown in Tab. 1, SiameseIM is able to surpass both MIM and ID methods over a wide range of evaluation tasks, including full-data fine-tuning, few-shot learning and linear probing on ImageNet [16], object detection on COCO [36] and LVIS [22], semantic segmentation on ADE20k [58], as well as several robustness benchmarks [26–28, 47]. By gathering semantic alignment and spatial sensitivity in one model, SiameseIM can deliver superior results for all tasks. We also note that such improvements are more obvious on 2133 ADE20k( ∼3 points) and LVIS( ∼1.6 point for rare classes) datasets. These long-tailed datasets demands semantic alignment and spatial sensitivity at the same time, and SiameseIM thus can deliver superior performance on them. Our contributions can be summarized as follows: • As a new form of SSL, SiameseIM is proposed to explore the possibilities of self-supervised pre-training. It dis- plays for the first time that that using only a single dense loss is enough to learn semantic alignment and spatial sensitivity well at the same time; • Compared with MIM methods, SiameseIM shows that re- constructing another view helps to obtain good semantic alignment. This also suggests that MIM framework can be used to reconstruct other targets with proper guidance, which opens a possible direction for MIM pretraining; • Compared with ID methods, SiameseIM shows that dense supervision can be applied by matching the dense correspondence between two views strictly through their relative positions. We demonstrate dense supervision can bring a considerable improvement of spatial sensitivity; • SiameseIM is able to surpass both MIM and ID meth- ods over a wide range of tasks. SiameseIM obtains more improvements in few-shot, long-tail and robustness- concerned scenarios.

Su_Physics-Driven_Diffusion_Models_for_Impact_Sound_Synthesis_From_Videos_CVPR_2023

Abstract Modeling sounds emitted from physical object interac- tions is critical for immersive perceptual experiences in real and virtual worlds. Traditional methods of impact sound synthesis use physics simulation to obtain a set of physics parameters that could represent and synthesize the sound. However, they require fine details of both the object geome- tries and impact locations, which are rarely available in the real world and can not be applied to synthesize impact sounds from common videos. On the other hand, existing video-driven deep learning-based approaches could only capture the weak correspondence between visual content and impact sounds since they lack of physics knowledge. In this work, we propose a physics-driven diffusion model that can synthesize high-fidelity impact sound for a silent video clip. In addition to the video content, we propose to use ad- ditional physics priors to guide the impact sound synthesis procedure. The physics priors include both physics parame- ters that are directly estimated from noisy real-world impact sound examples without sophisticated setup and learned residual parameters that interpret the sound environment via neural networks. We further implement a novel diffusion model with specific training and inference strategies to com- bine physics priors and visual information for impact sound synthesis. Experimental results show that our model outper- forms several existing systems in generating realistic impact sounds. Lastly, the physics-based representations are fully interpretable and transparent, thus allowing us to perform sound editing flexibly. We encourage the readers visit our project page1to watch demo videos with the audio turned on to experience the result.

1. Introduction Automatic sound effect production has become demand- ing for virtual reality, video games, animation, and movies. Traditional movie production heavily relies on talented Foley artists to record many sound samples in advance and man- *Work done while interning at MIT-IBM Watson AI Lab 1https://sukun1045.github.io/video-physics-sound- diffusion/ Figure 1. The physics-driven diffusion model takes physics priors and video input as conditions to synthesize high-fidelity impact sound. Please also see the supplementary video and materials with sample results. ually perform laborious editing to fit the recorded sounds to visual content. Though we could obtain a satisfactory sound experience at the cinema, it is labor-intensive and challenging to scale up the sound effects generation of vari- ous complex physical interactions. Recently, much progress has been made in automatic sound synthesis, which can be divided into two main cate- gories. The first category is physics-based modal synthesis methods [37, 38, 50], which are often used for simulating sounds triggered by various types of object interactions. Al- though the synthesized sounds can reflect the differences between various interactions and the geometry property of the objects, such approaches require a sophisticated designed environment to perform physics simulation and compute a set of physics parameters for sound synthesis. It is, therefore, impractical to scale up for a complicated scene because of the time-consuming parameter selection procedure. On the other hand, due to the availability of a significant amount of impact sound videos in the wild, training deep learning models for impact sound synthesis turns out to be a promis- ing direction. Indeed, several works have shown promising results in various audio-visual applications [63]. Unfortu- nately, most existing video-driven neural sound synthesis methods [7, 62] apply end-to-end black box model training and lack of physics knowledge which plays a significant role in modeling impact sound because a minor change in the impact location could exert a significant difference in the sound generation process. As a result, these methods are prone to learning an average or smooth audio representa- tion that contains artifacts, which usually leads to generating 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. 9749 unfaithful sound. In this work, we aim to address the problem of auto- matic impact sound synthesis from video input. The main challenge for the learning-based approach is the weak cor- respondence between visual and audio domains since the impact sounds are sensitive to the undergoing physics. With- out further physics knowledge, generating high-fidelity im- pact sounds from videos alone is insufficient. Motivated by physics-based sound synthesis methods using a set of physics mode parameters to represent and re-synthesize im- pact sounds, we design a physics prior that could contain sufficient physics information to serve as a conditional sig- nal to guide the deep generative model synthesizes impact sounds from videos. However, since we could not perform physics simulation on raw video data to acquire precise physics parameters, we explored estimating and predicting physics priors from sounds in videos. We found that such physics priors significantly improve the quality of synthe- sized impact sounds. For deep generative models, recent successes in image generation such as DALL-E 2 and Ima- gen [44] show that Denoising Diffusion Probabilistic Models (DDPM) outperform GANs in terms of fidelity and diversity, and its training process is usually with less instability and mode collapse issues. While the idea of the denoising pro- cess is naturally fitted with sound signals, it is unclear how video input and physics priors could jointly condition the DDPM and synthesize impact sounds. To address all these challenges, we propose a novel sys- tem for impact sound synthesis from videos. The system in- cludes two main stages. In the first stage, we encode physics knowledge of the sound using physics priors, including es- timated physical parameters using signal processing tech- niques and learned residual parameters interpreting the sound environment via neural networks. In the second stage, we formulate and design a DDPM model conditioned on visual input and physics priors to generate a spectrogram of impact sounds. Since the physics priors are extracted from the audio samples, they become unavailable at the inference stage. To solve this problem, we propose a novel inference pipeline to use test video features to query a physics latent feature from the training set as guidance to synthesize impact sounds on unseen videos. Since the video input is unseen, we can still generate novel impact sounds from the diffusion model even if we reuse the training set’s physics knowledge. In summary, our main contributions to this work are: •We propose novel physics priors to provide physics knowledge to impact sound synthesis, including estimated physics parameters from raw audio and learned residual parameters approximating the sound environment. •We design a physics-driven diffusion model with different training and inference pipeline for impact sound synthesis from videos. To the best of our knowledge, we are the first work to synthesize impact sounds from videos using thediffusion model. •Our approach outperforms existing methods on both quan- titative and qualitative metrics for impact sound synthe- sis. The transparent and interpretable properties of physics priors unlock the possibility of interesting sound editing applications such as controllable impact sound synthesis.

Tang_You_Need_Multiple_Exiting_Dynamic_Early_Exiting_for_Accelerating_Unified_CVPR_2023

Abstract Large-scale Transformer models bring significant im- provements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full computation of the large- scale model, not all of inputs need the same amount of com- putation to conduct, potentially leading to computation re- source waste. To handle this challenge, early exiting is pro- posed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The exist- ing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. How- ever, such strategies cannot be applied to encoder in the widely-used unified architecture with both encoder and de- coder due to difficulty of output confidence estimation in the encoder layers. It is suboptimal in term of saving compu- tation power to ignore the early exiting in encoder compo- nent. To address this issue, we propose a novel early exiting strategy for unified vision language models, which allows to dynamically skip the layers in encoder and decoder si- multaneously in term of input layer-wise similarities with multiple times of early exiting, namely MuE . By decompos- ing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing perfor- mance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce expected inference time by up to 50% and 40% while main- taining 99% and 96% performance respectively.

1. Introduction Recent advances in multi-modal Transformer-based large-scale models [25, 28, 48, 57] bring improve- ments across various vision language tasks. Among the Transformer-based models, the unified sequence-to- sequence architecture [38, 46] has attracted much attention due to its potential to become a universal computation engine to diverse tasks. Although large-scale models have achieved unattainable performance, their expensive computational cost usually hinders their applications in real-time scenarios. While the scaling effect suggests that the performance of the model benefits from its increased size, not every in- put requires the same amount of computation to achieve similar performance. Such an observation is particularly valid in visual language tasks, where inputs from different modalities may require different amounts of computation. Early exiting is a popular method to reduce computational costs by adaptively skipping layers on top of the input while preserving the general knowledge of the large-scale model. Existing studies aim to deal with early exiting decisions for encoder-only models [51, 52] or decoder components in encoder-decoder architectures [8], but cannot induce early exiting decisions for both components at the same time. Considering that single-component strategies may be sub- optimal in terms of saving computation cost, in this paper, we investigate how to perform early exiting for both en- coder and decoder components in a sequence-to-sequence architecture to elucidate a new way to further improve in- ference efficiency. Given the varied complexity of the inputs, it is natural to consider skipping some layers of the encoder as well as the decoder. Current decision mechanisms use classi- fiers to predict the output confidence of intermediate repre- sentations and stop computation if the confidence reaches 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. 10781 -10% -20% -30% -40%-50% 5060708090Accuracy MuE(Ours) PABEE DeeBERT -10% -20% -30% -40%-50% 20507090110130150CIDEr MuE(Ours) DeeCap PABEE DeeBERTFigure 1. The performance of different early exiting methods on SNLI-VE [49] and MS COCO [1] with certain expected inference time reduction rates. predefined threshold. However, extending this to unified sequence-to-sequence model is non-trivial due to two main challenges: (1) there are dependencies in making decisions for exiting decisions in the encoder and decoder, and (2) it is difficult to apply confidence classifiers to skip the encoder layer before going through the decoder layer for task output. To address the above challenges and to enable early exit- ing of encoder and decoder in sequence-to-sequence frame- work, we propose a novel early exiting strategies based on layer-wise input similarity, which is different from existing works based on task confidence [42]. More specifically, the model is encouraged to skip following layers in both en- coder and decoder when the layer-wise similarity reaches a certain threshold. This method is inspired by the saturation observa- tion [11] which shows that the hidden-state of each Trans- former layer arrives at a saturation status as going into deep layers. For the vision-language tasks, we find that the sim-ilar observation regarding saturation is also valid as shown in Figure 3. This observation lands the foundation that we could make the exiting decision based on the intermediate layer-wise input similarities without going through the fol- lowing layers. Besides, since the computation needed for input in different modalities usually varies, we propose a modality decomposition mechanism, which could further enable early fusion large-scale multi-modal models to break tie between modalities and enjoy the flexible exiting deci- sion over modalities. To encourage the early exiting behav- ior with a minimal performance loss, we design a layer-wise task loss, which enforce the each layer to output informative features for final task. Figure 1 shows the results on SNLI- VE dataset [49] and MS COCO [1] in term of expected time reduction rate and task performance. We compare MuE with several State-of-the-art early existing methods and ob- serve that MuE is able to largely reduce inference time with a minimal performance drop compared to other SoTA meth- ods. Our main contributions are summarized as follows: • To the best of our knowledge, this is a pioneering work to extend early exiting choices to both encoder and de- coder of sequence-to-sequence architecture. To this end, we propose a novel early exiting strategy based on layer-wise input similarity with the valid assump- tion on saturation states in vision language models. • Given the different characteristics of the modalities and tasks, we decompose the modalities encoder for early-fusion pre-trained models, bringing additional improvements in terms of inference efficiency. • We introduce layer-wise task loss, linking each layer in the decoder to the final task, effectively helping to maintain task performance when a significant time re- duction is required. • Extensive experiments show that our method can largely reduce the inference time of vision language models by up to 50% with minimal performance drop.

Tang_Contrastive_Grouping_With_Transformer_for_Referring_Image_Segmentation_CVPR_2023

Abstract Referring image segmentation aims to segment the tar- get referent in an image conditioning on a natural language expression. Existing one-stage methods employ per-pixel classification frameworks, which attempt straightforwardly to align vision and language at the pixel level, thus failing to capture critical object-level information. In this paper, we propose a mask classification framework, Contrastive Grouping with Transformer network (CGFormer), which explicitly captures object-level information via token-based querying and grouping strategy. Specifically, CGFormer first introduces learnable query tokens to represent objects and then alternately queries linguistic features and groups visual features into the query tokens for object-aware cross- modal reasoning. In addition, CGFormer achieves cross- level interaction by jointly updating the query tokens and decoding masks in every two consecutive layers. Finally, CGFormer cooperates contrastive learning to the grouping strategy to identify the token and its mask corresponding to the referent. Experimental results demonstrate that CG- Former outperforms state-of-the-art methods in both seg- mentation and generalization settings consistently and sig- nificantly. Code is available at https://github.com/ Toneyaya/CGFormer .

1. Introduction Referring Image Segmentation (RIS) aims to segment the target referent in an image given a natural language ex- pression [12, 17, 64]. It attracts increasing attention in the research community and is expected to show its potential in real applications, such as human-robot interaction via natural language [50] and image editing [3]. Compared to classical image segmentation that classifies pixels or masks into a closed set of fixed categories, RIS requires locat- ing referent at pixel level according to the free-form nat- ural languages with open-world vocabularies. It faces chal- †Sibei Yang is the corresponding author. (a) CRIS/ReSTR (b) VLT (c) Our CGFormerKey Reshape Key …… …… Mask Head Visual Features Linguistic Features Pull CloserInitial Tokens Query Vectors Push Away …… Encoder/DecoderVisual Encoder Language Encoder“couch with squares” Decoder CGFormer Visual Encoder Language Encoder “couch with squares” “couch with squares” Visual Encoder Language Encoder …… Binary Mask Binary Mask Figure 1. Comparison of Transformer-based RIS methods and our CGFormer. (a) CRIS [51] and ReSTR [25] fuse relevant linguistic features into visual features, while (b) VLT [12] generates query vectors to query visual features for segmentation. In contrast, (c) our CGFormer introduces learnable query tokens and explicitly groups visual features into tokens conditioning on language. Co- operating the grouping strategy with contrastive learning, it iden- tifies the token and its mask corresponding to the referent. lenges of comprehensively understanding vision and lan- guage modalities and aligning them at pixel level. Existing works mainly follow the segmentation frame- work of per-pixel classification [17, 34] integrated with multi-modal fusion to address the challenges. They intro- duce various fusion methods [19,20,26,32,46,63] to obtain vision-language feature maps and predict the segmentation results based on the feature maps. Recently, the improve- ment of vision-language fusion for RIS mainly lies in utiliz- ing the Transformer [12, 25, 51, 62]. LA VT [62] integrates fusion module into the Transformer-based visual encoder. CRIS [51] and ReSTR [25] fuse linguistic features into each feature of the visual feature maps, as shown in Figure 1a. In contrast, VLT [12] integrates the relevant visual features into language-conditional query vectors via transformer de- coder, as shown in Figure 1b. Although these methods have improved the segmenta- 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. 23570 tion accuracy, they still face several intrinsic limitations. First, the works [25,51,62] based on pixel-level fusion only model the pixel-level dependencies for each visual feature, which fails to capture the crucial object/region-level infor- mation. Therefore, they cannot accurately ground expres- sions that require efficient cross-modal reasoning on ob- jects. Second, although VLT [12]’s query vectors contain object-level information after querying, it directly weights and reshapes different tokens into one multi-modal feature map for decoding the final segmentation mask. Therefore, it loses the image’s crucial spatial priors (relative spatial ar- rangement among pixels) in the reshaping process. More importantly, it does not model the inherent differences be- tween query vectors, resulting in that even though differ- ent query vectors comprehend expressions in their own way, they still focus on similar regions, but fail to focus on dif- ferent regions and model their relations. In this paper, we aim to propose a simple and effective framework to address these limitations. Instead of using per-pixel classification framework, we adopt an end-to-end mask classification framework [1, 16] (see Figure 1c) to explicitly capture object-level information and decode seg- mentation masks for both the referent and other disturbing objects/stuffs. Therefore, we can simplify RIS task by find- ing the corresponding mask for the expression. Note that our framework differs from two-stage RIS methods [53,64] which require explicitly detecting the objects first and then predicting the mask in the detected bounding boxes. Specifically, we propose a Contrastive Grouping with Transformer (CGFormer) network consisting of the Group Transformer and Consecutive Decoder modules. The Group Transformer aims to capture object-level informa- tion and achieve object-aware cross-modal reasoning. The success of applying query tokens in object detection [1, 69] and instance segmentation [8, 16, 54] could be a potential solution. However, it is non-trivial to apply them to RIS. Without the annotation supervision of other mentioned ob- jects other than the referent, it is hard to make tokens pay attention to different objects and distinguish the token cor- responding to the referent from other tokens. Therefore, although we also specify query tokens as object-level in- formation representations, we explicitly group the visual feature map’s visual features into query tokens to ensure that different tokens focus on different visual regions with- out overlaps. Besides, we can further cooperate contrastive learning with the grouping strategy to make the referent to- ken attend to the referent-relevant information while forcing other tokens to focus on different objects and background regions, as shown in Figure 1c. In addition, we alternately query the linguistic features and group the visual features into the query tokens for cross-modal reasoning. Furthermore, integrating and utilizing multi-level fea- ture maps are crucial for accurate segmentation. Previousworks [32,40,63] fuse visual and linguistic features at mul- tiple levels in parallel and late integrate them via ConvL- STM [47] or FPNs [30]. However, their fusion modules are solely responsible for cross-modal alignment at each level, which fails to perform joint reasoning for multiple levels. Therefore, we propose a Consecutive Decoder that jointly updates query tokens and decodes masks in every two con- secutive layers to achieve cross-level reasoning. To evaluate the effectiveness of CGFormer, we conduct experiments on three standard benchmarks, i.e., RefCOCO series datasets [39,41,65]. In addition, unlike semantic seg- mentation, RIS is not limited by the close-set classification but to open-vocabulary alignment. It is necessary to evalu- ate the generalization ability of RIS models. Therefore, we introduce new subsets of training sets on the three datasets to ensure the categories of referents in the test set are not seen in the training stage, inspired by the zero-shot visual grounding [44] and open-set object detection [67]. In summary, our main contributions are as follows, • We propose a Group Transformer cooperated with con- trastive learning to achieve object-aware cross-modal reasoning by explicitly grouping visual features into different regions and modeling their dependencies con- ditioning on linguistic features. • We propose a Consecutive Decoder to achieve cross- level reasoning and segmentation by jointly perform- ing the cross-modal inference and mask decoding in every two consecutive layers in the decoder. • We are the first to introduce an end-to-end mask clas- sification framework, the Contrastive Grouping with Transformer (CGFormer), for referring image segmen- tation. Experimental results demonstrate that our CG- Former outperforms all state-of-the-art methods on all three benchmarks consistently. • We introduce new splits on datasets for evaluating generalization for referring image segmentation mod- els. CGFormer shows stronger generalizability com- pared to state-of-the-art methods thanks to object- aware cross-modal reasoning via contrastive learning.

Tosi_NeRF-Supervised_Deep_Stereo_CVPR_2023

Abstract We introduce a novel framework for training deep stereo networks effortlessly and without any ground-truth. By leveraging state-of-the-art neural rendering solutions, we generate stereo training data from image sequences col- lected with a single handheld camera. On top of them, a NeRF-supervised training procedure is carried out, from which we exploit rendered stereo triplets to compensate for occlusions and depth maps as proxy labels. This results in stereo networks capable of predicting sharp and detailed disparity maps. Experimental results show that models trained under this regime yield a 30-40% improvement over existing self-supervised methods on the challenging Middle- bury dataset, filling the gap to supervised models and, most times, outperforming them at zero-shot generalization.

1. Introduction Depth from stereo is one of the longest-standing research fields in computer vision [ 38]. It involves finding pixels correspondences across two rectified images to obtain the disparity – i.e., their difference in terms of horizontal coor- dinates – and then use it to triangulate depth. After years of studies with hand-crafted algorithms [ 56], deep learning radically changed the way of approaching the problem [ 74]. End-to-end deep networks [ 50] rapidly became the domi- nant solution for stereo, delivering outstanding results on benchmarks [ 42,55,58] given sufficient training data.This latter requirement is the key factor for their suc- cess, but it is also one of the greatest limitations. Annotated data is hard to source when dealing with depth estimation since additional sensors are required (e.g., LiDARs), and thus represents a thick entry barrier to the field. Over the years, two main trends have allowed to soften this problem: self-supervised learning paradigms [ 3,23,76] and the use of synthetic data [ 30,41,75]. Despite these advances, both approaches still have weaknesses to address. Self-supervised learning: despite the possibility to train on any unlabeled stereo pair collected by any user – and potentially opening to data democratization – the use of self-supervised losses is ineffective at dealing with ill-posed stereo settings (e.g. occlusions, non-Lambertian surfaces, etc.). Albeit recent approaches soften the occlusions prob- lem [ 3], predictions are far from being as sharp, detailed and accurate as those obtained through supervised training. Moreover, the self-supervised stereo literature [ 13,29,76] often focuses on well-defined domains (i.e., KITTI) and rarely exposes domain generalization capabilities [ 3]. Synthetic data: although training on densely annotated synthetic images can guide the networks towards sharp, de- tailed and accurate predictions, the domain-shift that occurs when testing on real data dampens the full potential of the trained model. A large body of recent literature addressing zero-shot generalization [ 2,16,30,35,36,87] proves how relevant the problem is. However, obtaining stereo pairs as realistic as possible requires significant effort, despite 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. 855 thetic depth labels being easily sourced through a graphics rendering pipeline. Indeed, modelling high-quality assets is crucial for mitigating the domain shift and requires ex- cellent graphics skills. While artists make various assets available, these are seldom open source and necessitate ad- ditional human labor to be organized into plausible scenes. In short, in a world where data is the new gold, obtain- ing flexible and scalable training samples to unleash the full potential of deep stereo networks still remains an open prob- lem. In this paper, we propose a novel paradigm to ad- dress this challenge. Given the recent advances in neural rendering [ 44,45], we exploit them as data factories : we collect sparse sets of images in-the-wild with a standard, single handheld camera. After that, we train a Neural Ra- diance Field (NeRF) model for each sequence and use it to render arbitrary, novel views of the same scene. Specifi- cally, we synthesize stereo pairs from arbitrary viewpoints by rendering a reference view corresponding to the real ac- quired image, and a target one on the right of it, displaced by means of a virtual arbitrary baseline. This allows us to generate countless samples to train any stereo network in a self-supervised manner by leveraging popular photomet- ric losses [ 23]. However, this na ¨ıve approach would inherit the limitations of self-supervised methods [ 13,29,76] at oc- clusions, which can be effectively addressed by rendering athird view for each pair, placed on the left of the source view specularly to the other target image. This allows to compensate for the missing supervision at occluded regions. Moreover, proxy-supervision in the form of rendered depth by NeRF completes our NeRF-Supervised training regime. With it, we can train deep stereo networks by conducting a low-effort collection campaign, and yet obtain state-of-the- art results without requiring any ground-truth label – or not even a real stereo camera! – as shown on top of Fig. 1. We believe that our approach is a significant step towards democratizing training data. In fact, we will demonstrate how the efforts of just the four authors were enough to col- lect sufficient data (roughly 270 scenes) to allow our NeRF- Supervised stereo networks to outperform models trained on synthetic datasets, such as [ 2,16,30,35,36,87], as well as existing self-supervised methods [ 3,79] in terms of zero- shot generalization, as depicted at the bottom of Fig. 1. We summarize our main contributions as: • A novel paradigm for collecting and generating stereo training data using neural rendering and a collection of user-collected image sequences. • A NeRF-Supervised training protocol that combines rendered image triplets and depth maps to address oc- clusions and enhance fine details. • State-of-the art, zero-shot generalization results on challenging stereo datasets [ 55], without exploiting any ground-truth or real stereo pair.

Tascon-Morales_Logical_Implications_for_Visual_Question_Answering_Consistency_CVPR_2023

Abstract Despite considerable recent progress in Visual Question Answering (VQA) models, inconsistent or contradictory an- swers continue to cast doubt on their true reasoning ca- pabilities. However, most proposed methods use indirect strategies or strong assumptions on pairs of questions and answers to enforce model consistency. Instead, we propose a novel strategy intended to improve model performance by directly reducing logical inconsistencies. To do this, we in- troduce a new consistency loss term that can be used by a wide range of the VQA models and which relies on know- ing the logical relation between pairs of questions and an- swers. While such information is typically not available in VQA datasets, we propose to infer these logical relations using a dedicated language model and use these in our pro- posed consistency loss function. We conduct extensive ex- periments on the VQA Introspect and DME datasets and show that our method brings improvements to state-of-the- art VQA models while being robust across different archi- tectures and settings.

1. Introduction Visual Questioning Answering (VQA) models have drawn recent interest in the computer vision community as they allow text queries to question image content. This has given way to a number of novel applications in the space of model reasoning [8, 29, 54, 56], medical diagno- sis [21,37,51,60] and counterfactual learning [1,2,11]. With the ability to combine language and image information in a common model, it is unsurprising to see a growing use of VQA methods. Despite this recent progress, however, a number of im- portant challenges remain when making VQAs more pro- ficient. For one, it remains extremely challenging to build VQA datasets that are void of bias. Yet this is critical to ensure subsequent models are not learning spurious cor- relations or shortcuts [49]. This is particularly daunting in applications where domain knowledge plays an impor- tant role ( e.g., medicine [15, 27, 33]). Alternatively, ensur- VQAIs it the middle of summer? Yes Yes Is there snow on the ground? VQAIs it the middle of summer? It is the middle of summer There is no snow on the groundNo Yes Is there snow on the ground? RelationConventional VQA Inconsistent ConsistentOur MethodFigure 1. Top: Conventional VQA models tend to produce incon- sistent answers as a consequence of not considering the relations between question and answer pairs. Bottom: Our method learns the logical relation between question and answer pairs to improve consistency. ing that responses of a VQA are coherent, or consistent , is paramount as well. That is, VQA models that answer differ- ently about similar content in a given image imply inconsis- tencies in how the model interprets the inputs. A number of recent methods have attempted to address this using logic- based approaches [19], rephrashing [44], question gener- ation [18, 40, 41] and regularizing using consistency con- straints [47]. In this work, we follow this line of research and look to yield more reliable VQA models. We wish to ensure that VQA models are consistent in an- swering questions about images. This implies that if multi- ple questions are asked about the same image, the model’s answers should not contradict themselves. For instance, if one question about the image in Fig. 1 asks “Is there snow on the ground?”, then the answer inferred should be consis- tent with that of the question “Is it the middle of summer?” As noted in [43], such question pairs involve reasoning and perception, and consequentially lead the authors to define inconsistency when the reasoning and perception questions are answered correctly and incorrectly, respectively. Along this line, [47] uses a similar definition of inconsistency 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. 6725 regularize a VQA model meant to answer medical diagno- sis questions that are hierarchical in nature. What is crit- ical in both cases, however, is that the consistency of the VQA model depends explicitly on its answers, as well as the question and true answer. This hinges on the assump- tion that perception questions are sufficient to answer rea- soning questions. Yet, for any question pair, this may not be the case. As such, the current definition of consistency (or inconsistency) has been highly limited and does not truly reflect how VQAs should behave. To address the need to have self-consistent VQA mod- els, we propose a novel training strategy that relies on logi- cal relations. To do so, we re-frame question-answer (QA) pairs as propositions and consider the relational construct between pairs of propositions. This construct allows us to properly categorize pairs of propositions in terms of their logical relations. From this, we introduce a novel loss function that explicitly leverages the logical relations be- tween pairs of questions and answers in order to enforce that VQA models be self-consistent. However, datasets typ- ically do not contain relational information about QA pairs, and collecting this would be extremely laborious and dif- ficult. To overcome this, we propose to train a dedicated language model that infers logical relations between propo- sitions. Our experiments show that we can effectively in- fer logical relations from propositions and use them in our loss function to train VQA models that improve state-of- the-art methods via consistency. We demonstrate this over two different VQA datasets, against different consistency methods, and with different VQA model architectures. Our code and data are available at https://github.com/ sergiotasconmorales/imp_vqa .

Sun_Hierarchical_Semantic_Contrast_for_Scene-Aware_Video_Anomaly_Detection_CVPR_2023

Abstract Increasing scene-awareness is a key challenge in video anomaly detection (VAD). In this work, we propose a hier- archical semantic contrast (HSC) method to learn a scene- aware VAD model from normal videos. We first incorporate foreground object and background scene features with high- level semantics by taking advantage of pre-trained video parsing models. Then, building upon the autoencoder- based reconstruction framework, we introduce both scene- level and object-level contrastive learning to enforce the en- coded latent features to be compact within the same seman- tic classes while being separable across different classes. This hierarchical semantic contrast strategy helps to deal with the diversity of normal patterns and also increases their discrimination ability. Moreover, for the sake of tack- ling rare normal activities, we design a skeleton-based mo- tion augmentation to increase samples and refine the model further. Extensive experiments on three public datasets and scene-dependent mixture datasets validate the effectiveness of our proposed method.

1. Introduction With the prevalence of surveillance cameras deployed in public places, video anomaly detection (V AD) has at- tracted considerable attention from both academia and in- dustry. It aims to automatically detect abnormal events so that the workload of human monitors can be greatly reduced. By now, numerous V AD methods have been developed under different supervision settings, including weakly supervised [13, 50, 55, 58, 64, 76], purely unsu- pervised [69, 72], and ones learning from normal videos only [20, 24, 33, 44, 45]. However, it is extremely diffi- cult or even impossible to collect sufficient and comprehen- sive abnormal data due to the rare occurrence of anomalies, whereas collecting abundant normal data is relatively easy. Therefore, the setting of learning from normal data is more *Corresponding author. SceneAppearance/Motion Hierarchical Semantic Contrast+++PushPullPushFigure 1. An illustration of hierarchical semantic contrast. The encoded scene-appearance/motion features are gathered together with respect to their semantic classes. Best viewed in color. practical and plays the dominant role in past studies. Although a majority of previous techniques learn their V AD models from normal data, this task has still not been well addressed due to the following reasons. First, some anomalies are scene-dependent [46, 51], implying that an appearance or motion may be anomalous in one scene but normal in other scenes. How to detect scene- dependent anomalies while preventing background bias ( i.e. learning the background noise rather than the essence of anomaly [31]) is a challenging problem. Second, normal patterns are diverse. How to enable a deep V AD model to represent the diverse normality well but not generalize to anomalous data is also a challenge [18, 44]. Last but not least, samples collected from different normal patterns are imbalanced because some normal activities may appear very sparsely [46]. How to deal with rare but normal activ- ities is challenging as well. Previous V AD methods mainly perform learning at frame-level [20, 47, 75] or in an object-centric [17, 24, 78] way. The former is prone to suffer from the background bias [31] while most of the latter methods are background- agnostic. There are some attempts to address the above- mentioned challenges in one or another aspect. For in- stance, a spatio-temporal context graph [51] and a hierar- chical scene normality-binding model [1] are constructed to discover scene-dependent anomalies. Memory-augmented autoencoders (AE) [18,44] are designed to represent diverse normal patterns while lessening the powerful capacity 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. 22846 AEs. An over-sampling strategy [32] is adopted but to solve the imbalance between normal and abnormal data. Con- trastively, in this work we address all of these challenges simultaneously and in distinct ways. The primary objective of our work is to handle scene- dependent anomalies. An intuition behind scene-dependent anomalies is that, if a type of object or activity is never observed in one scene in normal videos, then it should be viewed as an anomaly. It implies that we can first determine the scene type and then check if an object or activity has oc- curred in normal patterns of this scene. Based on this obser- vation, we propose a hierarchical semantic contrast method to learn a scene-aware V AD model. Taking advantage of pre-trained video parsing networks, we group the appear- ance and activity of objects and background scenes into se- mantic categories. Then, building upon the autoencoder- based reconstruction framework, we design both scene- level and object-level contrastive learning to enforce the en- coded latent features to gather together with respect to their semantic categories, as shown in Fig. 1. When a test video is input, we retrieve weighted normal features for reconstruc- tion and the clips of high errors are detected as anomalies. The contributions of this work are as follows: • We build a scene-aware reconstruction framework composed of scene-aware feature encoders and object- centric feature decoders for anomaly detection. The scene-aware encoders take background scenes into ac- count while the object-centric decoders are to reduce the background noise. • We propose hierarchical semantic contrastive learn- ing to regularize the encoded features in the latent spaces, making normal features more compact within the same semantic classes and separable between dif- ferent classes. Consequently, it helps to discriminate anomalies from normal patterns. • We design a skeleton-based augmentation method to generate both normal and abnormal samples based on our scene-aware V AD framework. The augmented samples enable us to additionally train a binary clas- sifier that helps to boost the performance further. • Experiments on three public datasets demonstrate promising results on scene-independent V AD. More- over, our method also shows a strong ability in detect- ing scene-dependent anomalies on self-built datasets.

Song_Multi-Mode_Online_Knowledge_Distillation_for_Self-Supervised_Visual_Representation_Learning_CVPR_2023

Abstract Self-supervised learning (SSL) has made remarkable progress in visual representation learning. Some studies combine SSL with knowledge distillation (SSL-KD) to boost the representation learning performance of small models. In this study, we propose a Multi-mode Online Knowledge Dis- tillation method (MOKD) to boost self-supervised visual rep- resentation learning. Different from existing SSL-KD meth- ods that transfer knowledge from a static pre-trained teacher to a student, in MOKD, two different models learn collab- oratively in a self-supervised manner. Specifically, MOKD consists of two distillation modes: self-distillation and cross- distillation modes. Among them, self-distillation performs self-supervised learning for each model independently, while cross-distillation realizes knowledge interaction between dif- ferent models. In cross-distillation, a cross-attention feature search strategy is proposed to enhance the semantic feature alignment between different models. As a result, the two models can absorb knowledge from each other to boost their representation learning performance. Extensive experimen- tal results on different backbones and datasets demonstrate that two heterogeneous models can benefit from MOKD and outperform their independently trained baseline. In addition, MOKD also outperforms existing SSL-KD methods for both the student and teacher models.

1. Introduction Due to the promising performance of unsupervised visual representation learning in many computer vision tasks, self- supervised learning (SSL) has attracted widespread attention from the computer vision community. SSL aims to learn general representations that can be transferred to downstream tasks by utilizing massive unlabeled data. Among various SSL methods, contrastive learning [8, 21] has shown significant progress in closing the performance *Corresponding author. 1f1f/g99 Student1 Teacher1 EMA Self- Distillation 2f/g992f Teacher2 Student2 EMA Cross- Distillation Self- Distillation Cross- Distillation Figure 1. Overview of the proposed Multi-mode Online Knowledge Distillation (MOKD). In MOKD, two different models are trained collaboratively through two types of knowledge distillation modes, i.e., a self-distillation mode and a cross-distillation mode. EMA denotes exponential-moving-average. gap with supervised methods in recent years. It aims at max- imizing the similarity between views from the same instance (positive pairs) while minimizing the similarity among views from different instances (negative pairs). MoCo [10, 21] and SimCLR [8, 9] use both positive and negative pairs for con- trast. They significantly improve the performance compared to previous methods [45, 51]. After that, many methods are proposed to solve the limitations in contrastive learning, such as the false negative problem [14, 20, 29, 32, 43], the limi- tation of large batch size [31, 54], and the problem of hard augmented samples [27, 49]. At the same time, other stud- ies [2,4,5,11,17,18,55] abandon the negative samples during contrastive learning. With relatively large models, such as ResNet50 [23] or larger, these methods achieve comparable performance on different tasks than their supervised coun- terparts. However, as revealed in previous studies [15, 16], they do not perform well on small models [26, 42] and have a large gap from their supervised counterparts. To address this challenge in contrastive learning, some 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. 11848 studies [1, 9, 15, 16, 37, 44, 58] propose to combine knowl- edge distillation [24] with contrastive learning (SSL-KD) to improve the performance of small models. These methods first train a larger model in a self-supervised manner and then distill the knowledge of the trained teacher model to a smaller student model. There is a limitation in these SSL- KD methods, i.e., knowledge is distilled to the student model from the static teacher model in a unidirectional way. The teacher model cannot absorb knowledge from the student model to boost its performance. In this study, we propose a Multi-mode Online Knowl- edge Distillation method (MOKD), as illustrated in Fig. 1, to boost the representation learning performance of two models simultaneously. Different from existing SSL-KD methods that transfer knowledge from a static pre-trained teacher to a student, in MOKD, two different models learn collabora- tively in a self-supervised manner. Specifically, MOKD con- sists of a self-distillation mode and a cross-distillation mode. Among them, self-distillation performs self-supervised learn- ing for each model independently, while cross-distillation realizes knowledge interaction between different models. In addition, a cross-attention feature search strategy is pro- posed in cross-distillation to enhance the semantic feature alignment between different models. Extensive experimental results on different backbones and datasets demonstrate that model pairs can both benefit from MOKD and outperform their independently trained baseline. For example, when trained with ResNet [23] and ViT [13], two models can ab- sorb knowledge from each other, and representations of the two models show the characteristics of each other. In addi- tion, MOKD also outperforms existing SSL-KD methods for both the student and teacher models. The contributions of this study are threefold: •We propose a novel self-supervised online knowledge distillation method, i.e., MOKD. •MOKD can boost the performance of two models simul- taneously, achieving state-of-the-art contrastive learn- ing performance on different models. •MOKD achieves state-of-the-art SSL-KD performance.

Tseng_Consistent_View_Synthesis_With_Pose-Guided_Diffusion_Models_CVPR_2023

Abstract Novel view synthesis from a single image has been a cornerstone problem for many Virtual Reality applicationsthat provide immersive experiences. However , most exist- ing techniques can only synthesize novel views within alimited range of camera motion or fail to generate consis-tent and high-quality novel views under significant cam- era movement. In this work, we propose a pose-guideddiffusion model to generate a consistent long-term videoof novel views from a single image. We design an atten- tion layer that uses epipolar lines as constraints to facili-tate the association between different viewpoints. Experi- mental results on synthetic and real-world datasets demon-strate the effectiveness of the proposed diffusion modelagainst state-of-the-art transformer-based and GAN-based approaches. More qualitative results are available athttps://poseguided-diffusion.github.io/ .

1. Introduction Offering immersive 3D experiences from daily photos has attracted considerable attention. It is a cornerstonetechnique for a wide range of applications such as 3Dphoto [ 18,49], 3D asset generation [ 35], and 3D scene nav- igation [ 4]. Notably, rapid progress has been made in ad- dressing the single-image view synthesis [40,50,61,69] is- sue. Given an arbitrarily narrow field-of-view image, theseframeworks can produce high-quality images from novelviewpoints. However, these methods are limited to view- 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. 16773 points that are within a small range of the camera motion. The long-term single-image view synthesis task is re- cently proposed to address the limitation of small cameramotion range. As demonstrated in Figure 1, the task at- tempts to generate a video from a single image and a se- quence of camera poses. Note that different from the single- image view synthesis problem, the viewpoints of the lastfew video frames produced under this setting may be faraway from the original viewpoint. Take the results shownin Figure 1, for instance, the cameras are moving into dif- ferent rooms that were not observed in the input images. Generating long-term view synthesis results from a sin- gle image is challenging for two main reasons. First, dueto the large range of the camera motion, e.g., moving intoa new room, a massive amount of new content needs to be hallucinated for the regions that are not observed in the in- put image. Second, the view synthesis results should beconsistent across viewpoints, particularly in the regions ob- served in the input viewpoint or previously hallucinated inthe other views. Both explicit- and implicit-based solutions are proposed to handle these issues. Explicit-based approaches [ 17,24, 25,40] use a “warp and refine” strategy. Specifically, the image is first warped from the input to novel viewpointsaccording to some 3D priors, i.e., monocular depth estima-tion [ 37,38]. Then a transformer or GAN-based generative model is designed to refine the warped image. However,the success of the explicit-based schemes hinges on the ac-curacy of the monocular depth estimation. To address thislimitation, Rombach et al .[42] designed a geometry-free transformer to implicitly learn the 3D correspondences be-tween the input and output viewpoints. Although reason-able new content is generated, the method fails to producecoherent results across viewpoints. The LoR [ 39] frame- work leverages the auto-regressive transformer to furtherimprove the consistency. Nevertheless, generating consis-tent, high-quality long-term view synthesis results remainschallenging. In this paper, we propose a framework based on dif- fusion models for consistent and realistic long-term novelview synthesis. Diffusion models [ 14,52,54] have achieved impressive performance on many content creation applica- tions, such as image-to-image translation [ 44] and text-to- image generation [ 2,36,45]. However, these methods only work on 2D images and lack 3D controllability. To thisend, we develop a pose-guided diffusion model with the epipolar attention layers. Specifically, in the UNet [ 43] network of the proposed diffusion model, we design theepipolar attention layer to associate the input view and out-put view features. According to the camera pose informa-tion, we estimate the epipolar line on the input view featuremap for each pixel on the output view feature map. Sincethese epipolar lines indicate the candidate correspondences,we use the lines as the constraint to compute the attention weight between the input and output views. We conduct extensive quantitative and qualitative stud- ies on real-world Realestate10K [ 76] and synthetic Mat- terport3D [ 7] datasets to evaluate the proposed approach. With the epipolar attention layer, our pose-guided diffusionmodel is capable of synthesizing long-term novel views that1) have realistic new content in unseen regions and 2) are consistent with the other viewpoints. We summarize thecontributions as follows: • We propose a pose-guided diffusion model for the long-term single-image view synthesis task. • We consider the epipolar line as the constraint and de- sign an epipolar attention to associate pixels in the im-ages at input and output views for the UNet network inthe diffusion model. • We validate that the proposed method synthesizes re- alistic and consistent long-term view synthesis results on the Realestate10K and Matterport3D datasets.

Tang_Graph_Transformer_GANs_for_Graph-Constrained_House_Generation_CVPR_2023

Abstract We present a novel graph Transformer generative adver- sarial network (GTGAN) to learn effective graph node re- lations in an end-to-end fashion for the challenging graph- constrained house generation task. The proposed graph- Transformer-based generator includes a novel graph Trans- former encoder that combines graph convolutions and self- attentions in a Transformer to model both local and global interactions across connected and non-connected graph nodes. Specifically, the proposed connected node atten- tion (CNA) and non-connected node attention (NNA) aim to capture the global relations across connected nodes and non-connected nodes in the input graph, respectively. The proposed graph modeling block (GMB) aims to exploit local vertex interactions based on a house layout topology. More- over, we propose a new node classification-based discrim- inator to preserve the high-level semantic and discrimina- tive node features for different house components. Finally, we propose a novel graph-based cycle-consistency loss that aims at maintaining the relative spatial relationships be- tween ground truth and predicted graphs. Experiments on two challenging graph-constrained house generation tasks (i.e., house layout and roof generation) with two public datasets demonstrate the effectiveness of GTGAN in terms of objective quantitative scores and subjective visual real- ism. New state-of-the-art results are established by large margins on both tasks.

1. Introduction This paper focuses on converting an input graph to a re- alistic house footprint, as depicted in Figure 1. Existing house generation methods such as [2, 16, 20, 28, 32, 45, 47], typically rely on building convolutional layers. However, convolutional architectures lack an understanding of long- range dependencies in the input graph since inherent in- ductive biases exist. Several Transformer architectures [3, 6, 11, 17, 18, 24, 43, 44, 46, 54, 55] based on the self- attention mechanism have recently been proposed to encodelong-range or global relations, thus learn highly expressive feature representations. On the other hand, graph convolu- tion networks are good at exploiting local and neighborhood vertex correlations based on a graph topology. Therefore, it stands to reason to combine graph convolution networks and Transformers to model local as well as global interac- tions for solving graph-constrained house generation. To this end, we propose a novel graph Transformer gen- erative adversarial network (GTGAN), which consists of two main novel components, i.e., a graph Transformer- based generator and a node classification-based discrimi- nator (see Figure 1). The proposed generator aims to gen- erate a realistic house from the input graph, which consists of three components, i.e., a convolutional message passing neural network (Conv-MPN), a graph Transformer encoder (GTE), and a generation head. Specifically, Conv-MPN first receives graph nodes as inputs and aims to extract discrim- inative node features. Next, the embedded nodes are fed to GTE, in which the long-range and global relation reasoning is performed by the connected node attention (CNA) and non-connected node attention (NNA) modules. Then, the output from both attention modules is fed to the proposed graph modeling block (GMB) to capture local and neigh- borhood relationships based on a house layout topology. Fi- nally, the output of GTE is fed to the generative head to pro- duce the corresponding house layout or roof. To the best of our knowledge, we are the first to use a graph Transformer to model local and global relations across graph nodes for solving graph-constrained house generation. In addition, the proposed discriminator aims to distin- guish real and fake house layouts, which ensures that our generated house layouts or roofs look realistic. At the same time, the discriminator classifies the generated rooms to their corresponding real labels, preserving the discrimina- tive and semantic features ( e.g., size and position) for differ- ent house components. To maintain the graph-level layout, we also propose a novel graph-based cycle-consistency loss to preserve the relative spatial relationships between ground truth and predicted graphs. Overall, our contributions are summarized as follows: This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2173 Figure 1. Overview of the proposed GTGAN on house layout generation. It consists of a novel graph Transformer-based generator Gand a novel node classification-based discriminator D. The generator takes graph nodes as input and aims to capture local and global relations across connected and non-connected nodes using the proposed graph modeling block and multi-head node attention, respectively. Note that we do not use position embeddings since our goal is to predict positional node information in the generated house layout. The discriminator Daims to distinguish real and generated layouts and simultaneously classify the generated house layouts to their corresponding room types. The graph-based cycle-consistency loss aligns the relative spatial relationships between ground truth and predicted nodes. The whole framework is trained in an end-to-end fashion so that all components benefit from each other. • We propose a novel Transformer-based network ( i.e., GT- GAN) for the challenging graph-constrained house gener- ation task. To the best of our knowledge, GTGAN is the first Transformer-based framework, enabling more effec- tive relation reasoning for composing house layouts and validating adjacency constraints. • We propose a novel graph Transformer generator that combines both graph convolutional networks and Trans- formers to explicitly model global and local correlations across both connected and non-connected nodes simul- taneously. We also propose a new node classification- based discriminator to preserve high-level semantic and discriminative features for different types of rooms. • We propose a novel graph-based cycle-consistency loss to guide the learning process toward accurate relative spatial distance of graph nodes. • Qualitative and quantitative experiments on two challeng- ing graph-constrained house generation tasks ( i.e., house layout generation and house roof generation) with two datasets demonstrate that GTGAN can generate better house structures than state-of-the-art methods, such as HouseGAN [28] and RoofGAN [32].

Spratley_Unicode_Analogies_An_Anti-Objectivist_Visual_Reasoning_Challenge_CVPR_2023

Abstract Analogical reasoning enables agents to extract relevant information from scenes, and efficiently navigate them in familiar ways. While progressive-matrix problems (PMPs) are becoming popular for the development and evaluation of analogical reasoning in computer vision, we argue that the dominant methodology in this area struggles to expose the lack of meaningful generalisation in solvers, and rein- forces an objectivist stance on perception – that objects can only be seen one way – which we believe to be counter- productive. In this paper, we introduce the Unicode Analo- gies challenge, consisting of polysemic, character-based PMPs to benchmark fluid conceptualisation ability in vision systems. Writing systems have evolved characters at mul- tiple levels of abstraction, from iconic through to symbolic representations, producing both visually interrelated yet ex- ceptionally diverse images when compared to those exhib- ited by existing PMP datasets. Our framework has been de- signed to challenge models by presenting tasks much harder to complete without robust feature extraction, while remain- ing largely solvable by human participants. We therefore argue that Unicode Analogies elegantly captures and tests for a facet of human visual reasoning that is severely lack- ing in current-generation AI.

1. Introduction Traditionally, statistical classification models have been designed to neatly cleave data into categories. Even in tasks such as visual scene decomposition, where data re- sists full description by any one label, there is an underly- ing objectivist assumption being made; the expectation of there being an objective number of distinguishable “things” present, themselves belonging to singular classes. Human visual perception makes a departure from this. The sym- bolic world to which we attend, with firm compositional rules for scenes and their objects, and with their parts and positions, is subsisted by a churning sea of ongoing concep- tualisation processes deeply fluid and contextual [15]. In recent years, there has been a proliferation of com- puter vision architectures built with object-centric inductive Figure 1. An example problem in UA, instantiating the Distribute- Three rule with the Closure concept. Five out of six context frames are provided (left), with four answer frames to choose from (right). The correct answer is emboldened. biases [21], many of which represent states-of-the-art on popular datasets [7, 35, 41]. This is an important direction, as training models to decompose scenes into objects allows for an explicit abstraction stage promoting feature reuse. However, abstract visual reasoning tasks such as Bongard problems [3] expose philosophically [20] — and in this pa- per, experimentally — that such an approach might work against the creation of models that possess the ability to ab- stract and deploy useful concepts. This observation also en- gages a current debate in the literature regarding the scala- bility of built-in knowledge and inductive biases [24, 36]. Humans display flexibility in how they decompose scenes, and perceive such scenes at a level of abstraction informed by past experiences and appropriate to present goals [8,9]. Scene understanding in humans is therefore un- dergirded by something other than the perception of static objects [22], and the idea that scene modelling research can separate perception and higher cognition into a pipeline of self-contained modules is strongly critiqued [5]. Noticing other shortcomings of deep-learnt approaches to computer vision, including brittleness to out-of- distribution (OOD) data, a small number of abstraction datasets inspired by Raven’s Progressive Matrices have been recently released [23]. Further motivations to this direction include a) the expectation that tasks with such an extended history in general psychometric testing would be useful to import into computer vision research, and b) the opinion that the more broadly applicable a model’s ab- stracted concepts become, the more robust that model will 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. 19082 be under OOD conditions [29, 35]. While the applicabil- ity of such concepts should ideally be evaluated by these datasets, common approaches to dataset creation feature conceptual schemas consisting of simple objects that can be neatly dropped into scenes, and extracted by scene de- composition stages [35]. This seems to require little in the way of contextual perception, such as Hofstadter’s notion of “conceptual slippage” [16]. We observe that the world’s writing systems present a diverse resource of characters that are amenable to con- tent analysis, and can assemble novel reasoning problems of their own. We introduce the Unicode Analogies (UA) challenge, consisting of character-based progressive ma- trix problems (PMPs) to benchmark fluid conceptualisation ability in vision systems. The characters in UA are poly- semic, and may instantiate any number of concepts, with the salient concept only revealing itself given context (Fig. 1). By generating training and testing problems from disjoint sets of characters, we challenge these systems by present- ing tasks much harder to complete without robust feature extraction, while remaining largely solvable by human par- ticipants. In doing so, we contribute a dataset that unlike others in this area, operates on a rich conceptual schema that invites fine-grained experimentation, and is easily ex- tensible to new user-defined concepts. Over five key ex- periments, we explore human and model performance on a number of dataset splits generated by UA, demonstrate that state-of-the-art solvers are still far from achieving the founding goals their datasets were created for, and encour- age new solvers to overcome these limitations.

Somepalli_Diffusion_Art_or_Digital_Forgery_Investigating_Data_Replication_in_Diffusion_CVPR_2023

Abstract Cutting-edge diffusion models produce images with high quality and customizability, enabling them to be used for commercial art and graphic design purposes. But do diffu- sion models create unique works of art, or are they repli- cating content directly from their training sets? In this work, we study image retrieval frameworks that enable us to compare generated images with training samples and de- tect when content has been replicated. Applying our frame- works to diffusion models trained on multiple datasets in- cluding Oxford flowers, Celeb-A, ImageNet, and LAION, we discuss how factors such as training set size impact rates of content replication. We also identify cases where diffusion models, including the popular Stable Diffusion model, bla- tantly copy from their training data. Project page: https: //somepago.github.io/diffrep.html

1. Introduction The rapid rise of diffusion models has led to new gen- erative tools with the potential to be used for commer- cial art and graphic design. The power of the diffusion paradigm stems in large part from its reliance on simple de- noising networks that maintain their stability when trained on huge web-scale datasets containing billions of image- caption pairs. These mega-datasets have the power to forge commercial models like DALL ·E[54] and Stable Diffusion [56], but also bring with them a number of legal and ethical risks [ 7]. Because these datasets are too large for careful hu- man curation, the origins and intellectual property rights of the data sources are largely unknown. This fact, combined with the ability of large models to memorize their training data [ 9,10,22], raises questions about the originality of dif- fusion outputs. There is a risk that diffusion models might, without notice, reproduce data from the training set directly, or present a collage of multiple training images. We informally refer to the reproduction of training 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. 6048 images, either in part or in whole, as content replication . In principle, replicating partial or complete information from the training data has implications for the ethical and legal use of diffusion models in terms of attributions to artists and photographers. Replicants are either a benefit or a hazard; there may be situations where content replication is acceptable, desirable, or fair use, and others where it is “stealing.” While these ethical boundaries are unclear at this time, we focus on the scientific question of whether replication actually happens with modern state-of-the-art diffusion models, and to what degree . Our contributions are as follows. We begin with a study of how to detect content replication, and we consider a range of image similarity metrics developed in the self- supervised learning and image retrieval communities. We benchmark the performance of different image feature ex- tractors using real and purpose-built synthetic datasets and show that state-of-the-art instance retrieval models work well for this task. Armed with new and existing tools, we search for data replication behavior in a range of diffusion models with different dataset properties. We show that for small and medium dataset sizes, replication happens fre- quently, while for a model trained on the large and diverse ImageNet dataset, replication seems undetectable. This latter finding may lead one to believe that replica- tion is not a problem for large-scale models. However, the even larger Stable Diffusion model exhibits clear replica- tion in various forms (Fig 1). Furthermore, we believe that the rate of content replication we identify in Stable Diffu- sionlikely underestimates the true rate because the model is trained on a 2B image split of LAION, but we only search for matches in the smaller 12M “Aesthetics v2 6+” subset. The level of image similarity required for something to count as “replication” is subjective and may depend on both the amount of diversity within the image’s class as well as the observer. Some replication behaviors we uncover are unambiguous, while in other instances they fall into a gray area. Rather than choosing an arbitrary definition, we fo- cus on presenting quantitative and qualitative results to the reader, leaving each person to draw their own conclusions based on their role and stake in the process of generative AI.

Tukra_Improving_Visual_Representation_Learning_Through_Perceptual_Understanding_CVPR_2023

Abstract We present an extension to masked autoencoders (MAE) which improves on the representations learnt by the model by explicitly encouraging the learning of higher scene-level features. We do this by: (i) the introduction of a perceptual similarity term between generated and real images (ii) in- corporating several techniques from the adversarial train- ing literature including multi-scale training and adaptive discriminator augmentation. The combination of these re- sults in not only better pixel reconstruction but also repre- sentations which appear to capture better higher-level de- tails within images. More consequentially, we show how our method, Perceptual MAE, leads to better performance when used for downstream tasks outperforming previous methods. We achieve 78.1%top-1 accuracy linear probing on ImageNet-1K and up to 88.1%when fine-tuning, with similar results for other downstream tasks, all without use of additional pre-trained models or data.

1. Introduction Self-supervision provides a powerful framework for training deep neural networks without relying on explicit supervision or labels where learning proceeds by predicting one part of the input data from another. Approaches based on denoising autoencoders [45], where the input is masked and the missing parts reconstructed, have shown to be ef- fective for pre-training in NLP with BERT [8], and more recently similar techniques have been applied for learning visual representations from images [1,4,17,30]. Such meth- ods effectively use image reconstruction as a pretext task on the basis that by learning to predict missing patches useful representations can be learnt for downstream tasks. One challenge when applying such techniques to images is that, unlike language where words contain some level of semantic meaning by design, the pixels in images are natu- ral signals containing high-frequency variations. Therefore, image-based denoising autoencoders have been adapted to avoid learning trivial solutions to reconstruction based on local textures or patterns. BEiT [1] uses an intermediarycodebook of patches such that pixels are not reconstructed directly, whilst MAE [17] masks a high proportion of the image to force the model to learn how to reconstruct whole scenes with limited context. In this paper, we build upon MAE and ask how we can move beyond the implicit conditioning of high masking ra- tios to explicitly incorporate the learning of higher-order ‘semantic’ features into the learning objective. To do this, we focus on introducing scene-level information by adding a perceptual loss term [22]. This works by constraining feature map similarity with a second pre-trained network, a technique which has been shown empirically in the gener- ative modelling literature to improve perceptual reconstruc- tion quality [54]. In addition, this also provides a mecha- nism to incorporate relevant scene-level cues contained in the second network (which could be e.g.a strong ImageNet classifier or a pre-trained language-vision embedding). One of the benefits of MAE is that it can rapidly learn strong representations using only self-supervision from the images in the target pre-training set. To maintain this prop- erty, we introduce a second idea: tying the features not with a separate network, but with an adversarial discriminator trained in parallel to distinguish between real and generated images. Both ideas combined result in not only a lower re- construction error, but also learnt representations which bet- ter capture details of the scene layout and object boundaries (see Figure 3) without either explicit supervision or the use of hand-engineered inductive biases. Finally, we build on these results and show that tech- niques from the generative modelling literature such as multi-scale gradients [24] and adaptive discriminator aug- mentation [25] can lead to further improvements in the learnt representation, and this also translates into a further boost in performance across downstream tasks. We hypoth- esise that the issues that these methods were designed to overcome, such as mode collapse during adversarial train- ing and incomplete learning of the underlying data distribu- tion, are related to overfitting on low-level image features. Our contributions can be summarized as follows: (i) we introduce a simple and self-contained technique to improve the representations learnt by masked image modelling 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. 14486 on perceptual loss and adversarial learning, (ii) we per- form a rigorous evaluation of this method and variants, and set new state-of-the-art for masked image modelling with- out additional data for classification on ImageNet-1K, ob- ject detection on MS COCO and semantic segmentation on ADE20K, (iii) we demonstrate this approach can further draw on powerful pre-trained models when available, re- sulting in a further boost in performance and (iv) we show our approach leads qualitatively to more ‘object-centric’ representations and stronger performance with frozen fea- tures (in the linear probe setting) compared to MAE.

Sui_ScanDMM_A_Deep_Markov_Model_of_Scanpath_Prediction_for_360deg_CVPR_2023

Abstract Scanpath prediction for 360◦images aims to produce dy- namic gaze behaviors based on the human visual perception mechanism. Most existing scanpath prediction methods for 360◦images do not give a complete treatment of the time- dependency when predicting human scanpath, resulting in inferior performance and poor generalizability. In this pa- per, we present a scanpath prediction method for 360◦im- ages by designing a novel Deep Markov Model (DMM) architecture, namely ScanDMM. We propose a semantics- guided transition function to learn the nonlinear dynam- ics of time-dependent attentional landscape. Moreover, a state initialization strategy is proposed by considering the starting point of viewing, enabling the model to learn the dynamics with the correct “launcher”. We further demon- strate that our model achieves state-of-the-art performance on four 360◦image databases, and exhibit its generalizabil- ity by presenting two applications of applying scanpath pre- diction models to other visual tasks – saliency detection and image quality assessment, expecting to provide profound in- sights into these fields.

1. Introduction 360◦images, also referred to as omnidirectional, sphere or virtual reality (VR) images, have been a popular type of visual data in many applications, providing us with immer- sive experiences. Nevertheless, how people explore virtual environments in 360◦images has not been well understood. The scanpath prediction model that aims at generating real- istic gaze trajectories has obtained increasing attention due to its significant influence in understanding users’ viewing behaviors in VR scenes, as well as in developing VR ren- dering, display, compression, and transmission [51]. Scanpath prediction has been explored for many years in 2D images [29]. However, 360◦images are different greatly from 2D images, as a larger space is offered to interact with *Corresponding author (email: [email protected]) Saliency -based models Generative models Noise Our model 𝐳𝟎𝐳𝟏𝐳𝟐 ෤𝐱𝟏෤𝐱𝟐···෠𝐒෠𝐒 ෠𝐒 Semantics Image GTFigure 1. Existing scanpath prediction models for 360◦images could be classified to two types: saliency-based models [2, 70, 71] and generative models [42, 43]. The scanpaths produced by saliency-based models, taking the study [2] as an example, com- monly exhibits unstable behavior with large displacements and scarce focal regions. Generative models, taking the study [43] as an example, shows less attention to regions of interest. The pro- posed ScanDMM can produce more realistic scanpaths that focus on regions of interests. – humans are allowed to use both head and gaze move- ments to explore viewports of interest in the scene. In such a case, viewing conditions, e.g., the starting point of viewing, has an important impact on humans’ scanpaths [20, 21, 52], and leads to complex and varied scanpaths among humans. This is inherently different from what happens in 2D visu- als since humans can directly guide their attention to the regions of interest. Therefore, scanpath prediction for 360◦ images is a more complex task. Current 360◦image scanpath prediction methods could 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. 6989 be roughly divided into two categories: saliency-based [2, 70, 71] and generative methods [42, 43]. The basic idea of the former one is to sample predicted gaze points from a saliency map. The performance of such methods is highly dependent on that of the saliency maps. Furthermore, con- structing a satisfactory sampling strategy to account for time-dependent visual behavior is non-trivial – the results of SaltiNet [2] exhibit unstable behavior with large displace- ments and scarce focal regions (see Fig. 1). The latter group of methods utilizes the advance of generative models, e.g., Generative Adversarial Network (GAN), to predict realis- tic scanpaths. However, such methods show less attention to regions of interests (see Fig. 1). In addition, the GAN- based methods are less flexible in determining the length of scanpaths and commonly suffer from unstable training. None of above-mentioned studies give a complete treat- ment of the time-dependency of viewing behavior, which is critical for modeling dynamic gaze behaviors in 360◦im- ages. For time-series data, a popular approach is to leverage sequential models, e.g., recurrent neural networks (RNNs), as exemplified in gaze prediction for 360◦videos [17, 35, 45]. However, such deterministic models are prone to over- fitting, particularly on small 360◦databases. More impor- tantly, they typically make simplistic assumptions, e.g., one choice is to concatenate the saliency map to the model’s hid- den states [17, 45], which assumes that the network learns how the states evolve by learning from saliency maps. Nev- ertheless, the neuroscience research [62] reveals that in ad- dition to top-down and bottom-up features, prior history and scene semantics are essential sources for guiding visual at- tention. Moreover, to be identified as interests or rejected as distractors, items must be compared to target templates held in memory [62]. Inspired by this, we argue that hu- mans’ scanpaths in 360◦scenes are complex nonlinear dy- namic attentional landscapes over time as a function of in- terventions of scene semantics on visual working memory. We present a probabilistic approach to learning the visual states that encode the time-dependent attentional landscape by specifying how these states evolve under the guidance of scene semantics and visual working memory. We instanti- ate our approach in the Deep Markov Model (DMM) [28], namely ScanDMM. Our contributions can be summarized as follows: • We present a novel method for time-dependent vi- sual attention modeling for 360◦images. Specifically, we model the mechanism of visual working memory by maintaining and updating the visual state in the Markov chain. Furthermore, a semantics-guided tran- sition function is built to learn the nonlinear dynamics of the states, in which we model the interventions of scene semantics on visual working memory. • We propose a practical strategy to initialize the visual state, facilitating our model to focus on learning thedynamics of states with correct “launcher”, as well as enabling us to assign a specific starting point for scan- path generation. Moreover, ScanDMM is capable of producing 1,000variable-length scanpaths within one second, which is critical for real-world applications. • We apply the proposed ScanDMM to two other com- puter vision tasks – saliency detection and image qual- ity assessment, which demonstrates our model equips with strong generalizability and is expected to provide insights into other vision tasks.

Sun_Masked_Motion_Encoding_for_Self-Supervised_Video_Representation_Learning_CVPR_2023

Abstract How to learn discriminative video representation from unlabeled videos is challenging but crucial for video anal- ysis. The latest attempts seek to learn a representation model by predicting the appearance contents in the masked regions. However, simply masking and recovering ap- pearance contents may not be sufficient to model tempo- ral clues as the appearance contents can be easily recon- structed from a single frame. To overcome this limitation, we present Masked Motion Encoding (MME), a new pre- training paradigm that reconstructs both appearance and motion information to explore temporal clues. In MME, we focus on addressing two critical challenges to improve the representation performance: 1) how to well represent the possible long-term motion across multiple frames; and 2) how to obtain fine-grained temporal clues from sparsely sampled videos. Motivated by the fact that human is able to recognize an action by tracking objects’ position changes and shape changes, we propose to reconstruct a motion trajectory that represents these two kinds of change in the masked regions. Besides, given the sparse video input, we enforce the model to reconstruct dense motion trajectories in both spatial and temporal dimensions. Pre-trained with our MME paradigm, the model is able to anticipate long- term and fine-grained motion details. Code is available at https://github.com/XinyuSun/MME .

1. Introduction Video representation learning plays a critical role in video analysis like action recognition [15,32,79], action lo- calization [12,81], video retrieval [4,82], videoQA [40], etc. Learning video representation is very difficult for two rea- sons. Firstly, it is extremely difficult and labor-intensive to annotate videos, and thus relying on annotated data to learn *Equal contribution. Email: {csxinyusun, phchencs }@gmail.com †Corresponding author. Email: [email protected] representations is not scalable. Also, the complex spatial-temporal contents with a large data volume are diffi- cult to be represented simultaneously. How to perform self- supervised videos representation learning only using unla- beled videos has been a prominent research topic [7,13,49]. Taking advantage of spatial-temporal modeling using a flexible attention mechanism, vision transformers [3, 8, 25, 26, 53] have shown their superiority in representing video. Prior works [5, 37, 84] have successfully introduced the mask-and-predict scheme in NLP [9,23] to pre-train an im- age transformer. These methods vary in different recon- struction objectives, including raw RGB pixels [37], hand- crafted local patterns [75], and VQ-V AE embedding [5], all above are static appearance information in images. Based on previous successes, some researchers [64,72,75] attempt to extend this scheme to the video domain, where they mask 3D video regions and reconstruct appearance information. However, these methods suffers from two limitations. First , as the appearance information can be well reconstructed in a single image with an extremely high masking ratio (85% in MAE [37]), it is also feasible to be reconstructed in the tube-masked video frame-by-frame and neglect to learn im- portant temporal clues. This can be proved by our ablation study (cf. Section 4.2.1). Second , existing works [64, 75] often sample frames sparsely with a fixed stride, and then mask some regions in these sampled frames. The recon- struction objectives only contain information in the sparsely sampled frames, and thus are hard to provide supervision signals for learning fine-grained motion details, which is critical to distinguish different actions [3, 8]. In this paper, we aim to design a new mask-and-predict paradigm to tackle these two issues. Fig. 1(a) shows two key factors to model an action, i.e., position change and shape change. By observing the position change of the person, we realize he is jumping in the air, and by observing the shape change that his head falls back and then tucks to his chest, we are aware that he is adjusting his posture to cross the bar. We believe that anticipating these changes helps the model better understand an action. 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. 2235 𝑡ℎ 𝑤 (b) Appearance reconstruction vs. motion trajectory reconstruction.Appearance reconstructionMotion trajectory reconstruction𝑡0 𝑡0𝑡2 ViT-based AutoencoderInput video (a) Two key factors to recognize a high jump action .Position change Shape change𝑓𝑒𝑛𝑐𝑓𝑑𝑒𝑐 𝑡ℎ 𝑤 𝑡1 Figure 1. Illustration of motion trajectory reconstruction for Masked Motion Encoding. (a) Position change and shape change over time are two key factors to recognize an action, we leverage them to represent the motion trajectory. (b) Compared with the current appearance reconstruction task, our motion trajectory reconstruction takes into account both appearance and motion information. Based on this observation, instead of predicting the ap- pearance contents, we propose to predict motion trajectory , which represents impending position and shape changes, for the mask-and-predict task. Specifically, we use a dense grid to sample points as different object parts, and then track these points using optical flow in adjacent frames to gener- ate trajectories, as shown in Fig. 1(b). The motion trajectory contains information in two aspects: the position features that describe relative movement; and the shape features that describe shape changes of the tracked object along the tra- jectory. To predict this motion trajectory, the model has to learn to reason the semantics of masked objects based on the visible patches, and then learn the correlation of objects among different frames and try to estimate their accurate motions. We name the proposed mask-and-predict task as Masked Motion Encoding (MME). Moreover, to help the model learn fine-grained motion details, we further propose to interpolate the motion trajec- tory. Taking sparsely sampled video as input, the model is asked to reconstruct spatially and temporally dense motion trajectories. This is inspired by the video frame interpo- lation task [77] where a deep model can reconstruct dense video at the pixel level from sparse video input. Differ- ent from it, we aim to reconstruct the fine-grained motion details of moving objects, which has higher-level motion information and is helpful for understanding actions. Our main contributions are as follows: • Existing mask-and-predict task based on appearance reconstruction is hard to learn important temporal clues, which is critical for representing video content. Our Masked Motion Encoding (MME) paradigm over- comes this limitation by asking the model to recon- struct motion trajectory. • Our motion interpolation scheme takes a sparsely sam- pled video as input and then predicts dense motion tra- jectory in both spatial and temporal dimensions. This scheme endows the model to capture long-term and fine-grained motion clues from sparse video input.Extensive experimental results on multiple standard video recognition benchmarks prove that the representations learned from the proposed mask-and-predict task achieve state-of-the-art performance on downstream action recog- nition tasks. Specifically, pre-trained on Kinetics-400 [10], our MME brings the gain of 2.3% on Something-Something V2 [34], 0.9% on Kinetics-400, 0.4% on UCF101 [59], and 4.7% on HMDB51 [44].

Turki_SUDS_Scalable_Urban_Dynamic_Scenes_CVPR_2023

Abstract We extend neural radiance fields (NeRFs) to dynamic large-scale urban scenes. Prior work tends to reconstruct single video clips of short durations (up to 10 seconds). Two reasons are that such methods (a) tend to scale linearly with the number of moving objects and input videos because a separate model is built for each and (b) tend to require su- pervision via 3D bounding boxes and panoptic labels, ob- tained manually or via category-specific models. As a step towards truly open-world reconstructions of dynamic cities, we introduce two key innovations: (a) we factorize the scene into three separate hash table data structures to efficiently encode static, dynamic, and far-field radiance fields, and (b) we make use of unlabeled target signals consisting of RGB images, sparse LiDAR, off-the-shelf self-supervised 2D descriptors, and most importantly, 2D optical flow. Op- erationalizing such inputs via photometric, geometric, and feature-metric reconstruction losses enables SUDS to de- compose dynamic scenes into the static background, indi- vidual objects, and their motions. When combined with our multi-branch table representation, such reconstructions can be scaled to tens of thousands of objects across 1.2 million frames from 1700 videos spanning geospatial footprints of hundreds of kilometers, (to our knowledge) the largest dy- namic NeRF built to date. We present qualitative initial re- sults on a variety of tasks enabled by our representations, including novel-view synthesis of dynamic urban scenes, unsupervised 3D instance segmentation, and unsupervised 3D cuboid detection. To compare to prior work, we also evaluate on KITTI and Virtual KITTI 2, surpassing state-of- the-art methods that rely on ground truth 3D bounding box annotations while being 10x quicker to train.

1. Introduction Scalable geometric reconstructions of cities have trans- formed our daily lives, with tools such as Google Maps and Streetview [ 6] becoming fundamental to how we navigate and interact with our environments. A watershed moment *Work done as an intern at Argo AI. Figure 1. SUDS. We scale neural reconstructions to city scale by dividing the area into multiple cells and training hash table rep- resentations for each. We show our full city-scale reconstruction above and the derived representations below. Unlike prior meth- ods, our approach handles dynamism across multiple videos, dis- entangling dynamic objects from static background and modeling shadow effects. We use unlabeled inputs to learn scene flow and semantic predictions, enabling category- and object-level scene manipulation. in the development of such technology was the ability to scale structure-from-motion (SfM) algorithms to city-scale footprints [ 4]. Since then, the advent of Neural Radiance Fields (NeRFs) [ 33] has transformed this domain by allow- ing for photorealistic interaction with a reconstructed scene via view synthesis. Recent works have attempted to scale such represen- tations to neighborhood-scale reconstructions for virtual drive-throughs [ 47] and photorealistic fly-throughs [ 52]. However, these maps remain static and frozen in time. This makes capturing bustling human environments— 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. 12375 complete with moving vehicles, pedestrians, and objects— impossible, limiting the usefulness of the representation. Challenges. One possible solution is a dynamic NeRF that conditions on time or warps a canonical space with a time-dependent deformation [ 38]. However, reconstruct- ing dynamic scenes is notoriously challenging because the problem is inherently under-constrained, particularly when input data is constrained to limited viewpoints, as is typi- cal from egocentric video capture [ 20]. One attractive so- lution is to scale up reconstructions to many videos, per- haps collected at different days (e.g., by an autonomous ve- hicle fleet). However, this creates additional challenges in jointly modeling fixed geometry that holds for all time (such as buildings), geometry that is locally static but transient across the videos (such as a parked car), and geometry that is truly dynamic (such as a moving person). SUDS. In this paper, we propose SUDS: Scalable Ur- ban Dynamic Scenes, a 4D representation that targets both scale anddynamism . Our key insight is twofold; (1) SUDS makes use of a rich suite of informative but freely avail- able input signals, such as LiDAR depth measurements and optical flow. Other dynamic scene representations [ 27,37] require supervised inputs such as panoptic segmentation la- bels or bounding boxes, which are difficult to acquire with high accuracy for our in-the-wild captures. (2) SUDS de- composes the world into 3 components: a static branch that models stationary topography that is consistent across videos, a dynamic branch that handles both transient (e.g., parked cars) and truly dynamic objects (e.g., pedestrians), and an environment map that handles far-field objects and sky. We model each branch using a multi-resolution hash table with scene partitioning, allowing SUDS to scale to an entire city spanning over 100 km2. Contributions. We make the following contributions: (1) to our knowledge, we build the first large-scale dynamic NeRF, (2) we introduce a scalable three-branch hash table representation for 4D reconstruction, (3) we present state- of-the-art reconstruction on 3 different datasets. Finally, (4) we showcase a variety of downstream tasks enabled by our representation, including free-viewpoint synthesis, 3D scene flow estimation, and even unsupervised instance seg- mentation and 3D cuboid detection.

Tastan_CaPriDe_Learning_Confidential_and_Private_Decentralized_Learning_Based_on_Encryption-Friendly_CVPR_2023

Abstract Large volumes of data required to train accurate deep neural networks (DNNs) are seldom available with any sin- gle entity. Often, privacy concerns prevent entities from sharing data with each other or with a third-party learning service provider. While cross-silo federated learning (FL) allows collaborative learning of large DNNs without shar- ing the data itself, most existing cross-silo FL algorithms have an unacceptable utility-privacy trade-off. In this work, we propose a framework called Confidential andPrivate Decentralized (CaPriDe) learning, which optimally lever- ages the power of fully homomorphic encryption (FHE) to enable collaborative learning without compromising on the confidentiality and privacy of data. In CaPridDe learn- ing, participating entities release their private data in an encrypted form allowing other participants to perform in- ference in the encrypted domain. The crux of CaPriDe learning is mutual knowledge distillation between multi- ple local models through a novel distillation loss, which is an approximation of the Kullback-Leibler (KL) divergence between the local predictions and encrypted inferences of other participants on the same data that can be computed in the encrypted domain. Extensive experiments on three datasets show that CaPriDe learning can improve the ac- curacy of local models without any central coordination, provide strong guarantees of data confidentiality and pri- vacy, and has the ability to handle statistical heterogene- ity. Constraints on the model architecture (arising from the need to be FHE-friendly), limited scalability, and compu- tational complexity of encrypted domain inference are the main limitations of the proposed approach. The code can be found at https://github.com/tnurbek/capride-learning.

1. Introduction Rapid strides have been made in machine learning (ML) (in particular, deep learning) over the past decade. How- ever, in many important application domains such as health-care and finance, the absence of large, centralized datasets is a significant obstacle to realizing the full benefits of deep learning algorithms. Data in these applications often resides in silos and is governed by strict regulations (e.g., HIPAA, GDPR, etc.) because of its privacy sensitive nature [22]. Competing business interests of data owners and the lack of appropriate incentives for data sharing further accentuate the problem. To overcome these issues, there is a need for collaborative learning algorithms that ideally satisfy the fol- lowing requirements [15]: (i) confidentiality - no sharing of raw data, (ii) privacy - minimal leakage of information via the knowledge exchange mechanism (e.g., gradients or pre- dictions), (iii) utility - gain in accuracy (over the individual models) resulting from the collaboration, even in the pres- ence of statistical heterogeneity, (iv) efficiency - minimize computational complexity and communication burden, (v) robustness - handle unintentional failures and attacks em- anating from malicious entities, and (vi) fairness - utility should be proportional to the individual contributions. Federated learning (FL) [15] is a special case of collab- orative learning, which works under the orchestration of a central server. FL allows multiple entities to collaboratively solve a ML problem by sharing of focused updates (e.g., gradients), instead of raw data. Specifically, cross-silo FL (typically between 2-100 participants) has been touted as a promising solution to address the data fragmentation prob- lem in finance [8] and healthcare [16,23]. Most prior cross- silo FL algorithms assume that all the parties are collec- tively training a single model with a common architecture , which is too restrictive in practice. Furthermore, knowl- edge exchange usually happens through sharing of gradi- ents or model parameters . Recent gradient inversion at- tacks [9, 14] demonstrate that it is indeed possible to re- cover high fidelity information from individual gradient up- dates, thus violating the privacy requirement. While dif- ferential privacy (DP) [24], secure multi-party computation (MPC) [26], and trusted execution environment (TEE) [21] have been proposed as potential remedies to safeguard pri- vacy in FL, none of the existing solutions offer an accept- able privacy-utility trade-off with reasonable efficiency. As This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8084 noted in [5], sharing of high dimensional gradients/model parameters is a fundamental privacy limitation of standard FL methods, which cannot be addressed by simply wrap- ping around a DP, MPC, or TEE solution. Confidential and private collaborative learning (CaPC) [7] is the only method that claims to achieve both confi- dentiality and privacy in a collaborative setting. In CaPC learning, each participant is able to query other parties via a private inference protocol based on 2-party computation (2PC). However, the answering parties cannot directly re- veal the inference logits to the querying party because it leaks information about their local models (e.g., through model extraction attacks). To circumvent this problem, dif- ferential privacy, private aggregation of teacher ensembles, and secure argmax computation through a central privacy guardian (PG) are employed to output only the predicted la- bel to the querying party. A drawback of the CaPC approach is that it achieves all privacy guarantees using the help of a semi-trusted PG. Moreover, the use of differential pri- vacy reduces the performance of FL algorithms, especially if strong privacy guarantees are required. Finally, the use of MPC requires multiple rounds of communication between the parties for each query. All other approaches that have been proposed to achieve knowledge transfer through distil- lation [18] require a non-private labeled/unlabeled/synthetic dataset that can be shared among participants. In this work, we propose a new protocol called Confidential andPrivateDecentralized (CaPriDe) learning, where participants learn from each other in a collaborative setting while preserving their confidentiality and privacy (see Figure 1). Unlike the 2PC protocols used in CaPC, we leverage fully homomorphic encryption (FHE) to en- able participants to publish their encrypted data. Since the published data is encrypted using the data owner’s pub- lic key and only the owner can decrypt the data (using the secret key), confidentiality is preserved. However, the CaPriDe framework allows other collaborators to perform encrypted inference on the published (encrypted) data by applying their own local model. Mutual knowledge distilla- tion (KD) [13] between the data owner’s local model and the local models of the collaborators is used to transfer knowl- edge between models and ensure a collaborative gain. KD is typically achieved through minimizing the distillation loss between the student and teacher responses (logits). Since the collaborators in CaPriDe learning make predictions on encrypted data, a loss function that can be computed with- out any decryption is required. Hence, we propose a new distillation loss, which is an approximation of the KL di- vergence that can be securely computed. Only an encrypted value of the distillation loss aggregated over an entire batch is sent back to the data owner, which ensures strong privacy. Data owners can decrypt this aggregate loss value and use it for updating their local models. Our contributions are:• We introduce the CaPriDe learning framework, which exploits FHE-based encrypted inference and knowl- edge distillation to achieve confidential and private collaborative learning without any central orchestra- tion and any need for non-private shared data. • To enable CaPriDe learning, we propose an encryption-friendly distillation loss that estimates theapproximate KL divergence between two model predictions and design a protocol to securely compute this loss in the encrypted domain. • We conduct extensive experiments to show that CaPriDe learning enables participants to achieve col- laborative gain, even in the non-iid setting. To prove feasibility, we implement encrypted inference using the Tile Tensors library with a FHE-friendly model.

Song_Robust_Single_Image_Reflection_Removal_Against_Adversarial_Attacks_CVPR_2023

Abstract This paper addresses the problem of robust deep single- image reflection removal (SIRR) against adversarial at- tacks. Current deep learning based SIRR methods have shown significant performance degradation due to unno- ticeable distortions and perturbations on input images. For a comprehensive robustness study, we first conduct diverse adversarial attacks specifically for the SIRR problem, i.e. towards different attacking targets and regions. Then we propose a robust SIRR model, which integrates the cross- scale attention module, the multi-scale fusion module, and the adversarial image discriminator. By exploiting the multi-scale mechanism, the model narrows the gap between features from clean and adversarial images. The image dis- criminator adaptively distinguishes clean or noisy inputs, and thus further gains reliable robustness. Extensive ex- periments on Nature, SIR2, and Real datasets demonstrate that our model remarkably improves the robustness of SIRR across disparate scenes.

1. Introduction Single image reflection removal (SIRR) is a classic topic in the low-level image processing area, namely a kind of image restoration. When taking an image through a trans- parent surface, a reflection layer would be blended with the original photography ( i.e. the transmission layer), resulting in imaging corruptions. The SIRR is devoted to recovering a clear transmission image by removing the reflection layer. However, the SIRR is fundamentally ill-posed [42] that there could be an infinite number of transmission and re- flection decompositions from a blended image. Therefore, traditional methods often exploit manual priors to optimize the layer separation, such as gradient sparsity prior [25] and ∗Equal contribution Corresponding Author: <[email protected] >. /uni00000026/uni0000004f/uni00000048/uni00000044/uni00000051 /uni00000030/uni00000036/uni00000028/uni00000042/uni00000029/uni00000035 /uni0000002f/uni00000033/uni0000002c/uni00000033/uni00000036/uni00000042/uni00000029/uni00000035 /uni00000030/uni00000036/uni00000028/uni00000042/uni00000035/uni00000035 /uni0000002f/uni00000033/uni0000002c/uni00000033/uni00000036/uni00000042/uni00000035/uni00000035 /uni00000030/uni00000036/uni00000028/uni00000042/uni00000031/uni00000035 /uni0000002f/uni00000033/uni0000002c/uni00000033/uni00000036/uni00000042/uni00000031/uni00000035/uni00000033/uni00000036/uni00000031/uni00000035/uni0000000b/uni00000047/uni00000025/uni0000000c /uni00000015/uni00000013/uni00000011/uni0000001c/uni0000001a /uni00000015/uni00000013/uni00000011/uni00000016/uni00000015/uni00000014/uni0000001b/uni00000011/uni00000018/uni00000016 /uni00000014/uni00000013/uni00000011/uni00000016/uni00000018/uni00000014/uni0000001c/uni00000011/uni0000001c/uni0000001b /uni00000014/uni00000018/uni00000011/uni00000014/uni00000018/uni00000015/uni00000013/uni00000011/uni0000001b/uni00000017 /uni00000014/uni0000001b/uni00000011/uni0000001c/uni0000001b/uni00000015/uni00000013/uni00000011/uni00000019/uni00000019 /uni00000014/uni0000001b/uni00000011/uni00000018/uni00000015/uni00000014/uni0000001b/uni00000011/uni00000018/uni00000017 /uni00000014/uni00000014/uni00000011/uni00000015/uni00000014/uni0000001c/uni00000011/uni0000001c/uni0000001c /uni00000014/uni00000019/uni00000011/uni00000015/uni00000016/uni00000025/uni00000044/uni00000056/uni00000048/uni0000004f/uni0000004c/uni00000051/uni00000048 /uni00000015/uni00000013/uni00000011/uni00000013/uni00000017/uni00000032/uni00000058/uni00000055/uni00000056/uni00000003/uni0000005a/uni00000003/uni00000044/uni00000047/uni00000059/uni00000011 /uni00000032/uni00000058/uni00000055/uni00000056/uni00000003/uni0000005a/uni00000012/uni00000052/uni00000003/uni00000044/uni00000047/uni00000059/uni00000011Figure 1. The PSNR measurements of our approach under dif- ferent kinds of adversarial attacks. ‘Clean’ indicates no-attacks, ‘MSE FR’ represents attacking on FullRegion with MSE objec- tive, and so on. The testing result is from Nature dataset [26]. relative smoothness prior [27]. These priors are often vio- lated when facing complex scenes. Recently, deep learning based methods [10, 16, 55] have attracted considerable at- tention to tackle the SIRR problem. By learning semantic and contextual features, deep SIRR methods have achieved much better quality of the recovered images. However, deep neural networks are often vulnerable to visually imperceptible adversarial perturbations [14, 29]. The prediction can be totally invalid even with slight and unnoticeable attacks on inputs. Similarly, such vulnerability is also an important issue for the deep SIRR problem, and the robustness of current methods has not been thoroughly studied. There have been no benchmarks and evaluations for the robustness of deep SIRR models against intended attacks. Meanwhile, general defense methods [44] have not been applied to SIRR models. Accordingly, the robust SIRR model is still a crucial and desiderate research problem. In this paper, we first investigate the robustness of deep SIRR methods. We apply the widely-used and powerful attack method PGD [29] to generate adversarial samples. For completeness of the robustness evaluation, we present various attack modes by referring [3, 48]. Specifically, we employ different attack objectives i.e. mean squared er- ror (MSE) and learned perceptual image patch similarity (LPIPS) [52], as well as different attack regions, i.e. 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. 24688 full image region (FR), the reflection region (RR), and the non-reflection region (NR). Through a systematic analy- sis, the most effective attack mode and currently the most robust SIRR model are identified. Then we conduct ad- versarial training based on this model to enhance its ro- bustness. In order to develop a furthermore robust SIRR model, we borrow the wisdom of multi-scale image pro- cessing [19] and adversarial discriminating [56] from pre- vious defense methods. Consequently, we build a new robust SIRR model based on the image transformer [41], which integrates the cross-scale attention module, the multi- scale fusion module, and the adversarial image discrimi- nator. The proposed method obtains significant improve- ments in robustness. Fig. 1 reveals the PSNR changes of our model prediction under distinct attack modes on the Na- ture dataset [26]. It is notable that our model yields limited degradations against perturbed images when compared with input clean images. Overall, our main contributions can be summarized be- low. (1) We present a comprehensive evaluation of existing deep SIRR methods in terms of their robustness against var- ious adversarial attacks on diverse datasets. Extensive ex- perimental studies suggest presently the most effective at- tack and the most robust SIRR model. (2) We propose a novel transformer-based SIRR model, which integrates sev- eral relatively-robust modules to defend against adversarial attacks. The model can mitigate the effects of perturbations and distinguish clean or polluted inputs as well. (3) We carry out sufficient experiments to analyze the robustness of the proposed method, which achieves state-of-the-art sta- bility against adversarial images. The model performs supe- rior reflection removal robustness on distorted images while maintaining favorable accuracy on original clean images.

Sun_Indiscernible_Object_Counting_in_Underwater_Scenes_CVPR_2023

Abstract Recently, indiscernible scene understanding has at- tracted a lot of attention in the vision community. We further advance the frontier of this field by systematically studying a new challenge named indiscernible object count- ing (IOC), the goal of which is to count objects that are blended with respect to their surroundings. Due to a lack of appropriate IOC datasets, we present a large-scale dataset IOCfish5K which contains a total of 5,637 high- resolution images and 659,024 annotated center points. Our dataset consists of a large number of indiscernible ob- jects (mainly fish) in underwater scenes, making the an- notation process all the more challenging. IOCfish5K is superior to existing datasets with indiscernible scenes be- cause of its larger scale, higher image resolutions, more annotations, and denser scenes. All these aspects make it the most challenging dataset for IOC so far, supporting progress in this area. For benchmarking purposes, we se- lect 14 mainstream methods for object counting and care- fully evaluate them on IOCfish5K. Furthermore, we propose IOCFormer , a new strong baseline that combines density and regression branches in a unified framework and can effectively tackle object counting under concealed scenes. Experiments show that IOCFormer achieves state-of-the- art scores on IOCfish5K. The resources are available at github.com/GuoleiSun/Indiscernible-Object-Counting.

1. Introduction Object counting – to estimate the number of object in- stances in an image – has always been an essential topic in computer vision. Understanding the counts of each cate- gory in a scene can be of vital importance for an intelligent agent to navigate in its environment. The task can be the end goal or can be an auxiliary step. As to the latter, counting objects has been proven to help instance segmentation [14], action localization [54], and pedestrian detection [83]. As to the former, it is a core algorithm in surveillance [78], crowd *Corresponding author ([email protected]) Figure 1. Illustration of different counting tasks. Top left : Generic Object Counting (GOC), which counts objects of various classes innatural scenes .Top right : Dense Object Counting (DOC), which counts objects of a foreground class in scenes packed with instances .Down : Indiscernible Object Counting (IOC), which counts objects of a foreground class in indiscernible scenes . Can you find all fishes in the given examples? For GOC, DOC, and IOC, the images shown are from PASCAL VOC [18], Shang- haiTech [91], and the new IOCfish5K dataset, respectively. monitoring [6], wildlife conservation [56], diet patterns un- derstanding [55] and cell population analysis [1]. Previous object counting research mainly followed two directions: generic/common object counting (GOC) [8, 14, 32, 68] and dense object counting (DOC) [28, 36, 50, 57, 64, 67, 91]. The difference between these two sub-tasks lies in the studied scenes, as shown in Fig. 1. GOC tack- les the problem of counting object(s) of various categories in natural/common scenes [8], i.e., images from PASCAL VOC [18] and COCO [41]. The number of objects to be es- timated is usually small, i.e., less than 10. DOC, on the other hand, mainly counts objects of a foreground class in crowded scenes. The estimated count can be hundreds 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. 13791 Dataset YearIndiscernible Scene#Ann. IMG Avg. Resolution Free ViewCount StatisticsWebTotal Min Ave Max UCSD [6] 2008 ✗ 2,000 158×238 ✗ 49,885 11 25 46 Link Mall [10] 2012 ✗ 2,000 480×640 ✗ 62,325 13 31 53 Link UCF CC50 [27] 2013 ✗ 50 2101×2888 ✓ 63,974 94 1,279 4,543 Link WorldExpo’10 [90] 2016 ✗ 3,980 576×720 ✗ 199,923 1 50 253 Link ShanghaiTech B [91] 2016 ✗ 716 768×1024 ✗ 88,488 9 123 578 Link ShanghaiTech A [91] 2016 ✗ 482 589×868 ✓ 241,677 33 501 3,139 Link UCF-QNRF [28] 2018 ✗ 1,535 2013×2902 ✓ 1,251,642 49 815 12,865 Link Crowd surv [87] 2019 ✗ 13,945 840×1342 ✗ 386,513 2 35 1420 Link GCC (synthetic) [80] 2019 ✗ 15,212 1080×1920 ✗ 7,625,843 0 501 3,995 Link JHU-CROWD++ [65] 2019 ✗ 4,372 910×1430 ✓ 1,515,005 0 346 25,791 Link NWPU-Crowd [79] 2020 ✗ 5,109 2191×3209 ✓ 2,133,375 0 418 20,033 Link NC4K [51] 2021 ✓ 4,121 530×709 ✓ 4,584 1 1 8 Link CAMO++ [33] 2021 ✓ 5,500 N/A ✓ 32,756 N/A 6 N/A Link COD [19] 2022 ✓ 5,066 737×964 ✓ 5,899 1 1 8 Link IOCfish5K (Ours) 2023 ✓ 5,637 1080×1920 ✓ 659,024 0 117 2,371 Link Table 1. Statistics of existing datasets for dense object counting (DOC) and indiscernible object counting (IOC). or even tens of thousands. The counted objects are often persons (crowd counting) [36, 39, 88], vehicles [26, 57] or plants [50]. Thanks to large-scale datasets [10,18,28,65,79, 91] and deep convolutional neural networks (CNNs) trained on them, significant progress has been made both for GOC and DOC. However, to the best of our knowledge, there is no previous work on counting indiscernible objects. Under indiscernible scenes, foreground objects have a similar appearance, color, or texture to the background and are thus difficult to be detected with a traditional visual system. The phenomenon exists in both natural and arti- ficial scenes [20, 33]. Hence, scene understanding for in- discernible scenes has attracted increasing attention since the appearance of some pioneering works [20, 34]. Various tasks have been proposed and formalized: camouflaged ob- ject detection (COD) [20], camouflaged instance segmen- tation (CIS) [33] and video camouflaged object detection (VCOD) [12, 31]. However, no previous research has fo- cused on counting objects in indiscernible scenes, which is an important aspect. In this paper, we study the new indiscernible object counting (IOC ) task, which focuses on counting foreground objects in indiscernible scenes. Fig. 1 illustrates this chal- lenge. Tasks such as image classification [17, 24], seman- tic segmentation [11, 42] and instance segmentation [3, 23] all owe their progress to the availability of large-scale datasets [16, 18, 41]. Similarly, a high-quality dataset for IOC would facilitate its advancement. Although existing datasets [20, 33, 51] with instance-level annotations can be used for IOC, they have the following limitations: 1) the total number of annotated objects in these datasets is lim- ited, and image resolutions are low; 2) they only contain scenes/images with a small instance count; 3) the instance- level mask annotations can be converted to point supervi- sion by computing the centers of mass, but the computed points do not necessarily fall inside the objects. To facilitate the research on IOC, we construct a large- scale dataset, IOCfish5K. We collect 5,637 images withindiscernible scenes and annotate them with 659,024 cen- ter points. Compared with the existing datasets, the pro- posed IOCfish5K has several advantages: 1) it is the largest- scale dataset for IOC in terms of the number of images, image resolution, and total object count; 2) the images in IOCfish5K are carefully selected and contain diverse indis- cernible scenes; 3) the point annotations are accurate and located at the center of each object. Our dataset is com- pared with existing DOC and IOC datasets in Table 1, and example images are shown in Fig. 2. Based on the proposed IOCfish5K dataset, we provide a systematic study on 14 mainstream baselines [32,36,39,40, 45, 47, 52, 66, 73, 76, 89, 91]. We find that methods which perform well on existing DOC datasets do not necessarily preserve their competitiveness on our challenging dataset. Hence, we propose a simple and effective approach named IOCFormer. Specifically, we combine the advantages of density-based [76] and regression-based [39] counting ap- proaches. The former can estimate the object density across the image, while the latter directly regresses the co- ordinates of points, which is straightforward and elegant. IOCFormer contains two branches: density and regression. The density-aware features from the density branch help make indiscernible objects stand out through the proposed density-enhanced transformer encoder (DETE). Then the refined features are passed through a conventional trans- former decoder, after which predicted object points are gen- erated. Experiments show that IOCFormer outperforms all considered algorithms, demonstrating its effectiveness on IOC. To summarize, our contributions are three-fold. • We propose the new indiscernible object counting (IOC) task. To facilitate research on IOC, we con- tribute a large-scale dataset IOCfish5K, containing 5,637 images and 659,024 accurate point labels. • We select 14 classical and high-performing approaches for object counting and evaluate them on the proposed IOCfish5K for benchmarking purposes. 13792 • We propose a novel baseline, namely IOCFormer, which integrates density-based and regression-based methods in a unified framework. In addition, a novel density-based transformer encoder is proposed to grad- ually exploit density information from the density branch to help detect indiscernible objects.

Tian_Trainable_Projected_Gradient_Method_for_Robust_Fine-Tuning_CVPR_2023

Abstract Recent studies on transfer learning have shown that se- lectively fine-tuning a subset of layers or customizing dif- ferent learning rates for each layer can greatly improve ro- bustness to out-of-distribution (OOD) data and retain gen- eralization capability in the pre-trained models. However, most of these methods employ manually crafted heuristics or expensive hyper-parameter searches, which prevent them from scaling up to large datasets and neural networks. To solve this problem, we propose Trainable Projected Gradi- ent Method (TPGM) to automatically learn the constraint imposed for each layer for a fine-grained fine-tuning reg- ularization. This is motivated by formulating fine-tuning as a bi-level constrained optimization problem. Specifi- cally, TPGM maintains a set of projection radii, i.e., dis- tance constraints between the fine-tuned model and the pre- trained model, for each layer, and enforces them through weight projections. To learn the constraints, we propose a bi-level optimization to automatically learn the best set of projection radii in an end-to-end manner. Theoretically, we show that the bi-level optimization formulation is the key to learning different constraints for each layer. Em- pirically, with little hyper-parameter search cost, TPGM outperforms existing fine-tuning methods in OOD perfor- mance while matching the best in-distribution (ID) per- formance. For example, when fine-tuned on DomainNet- Real and ImageNet, compared to vanilla fine-tuning, TPGM shows 22% and 10% relative OOD improvement respec- tively on their sketch counterparts. Code is available at https://github.com/PotatoTian/TPGM .

1. Introduction Improving out-of-distribution (OOD) robustness such that a vision model can be trusted reliably across a variety of conditions beyond the in-distribution (ID) training data has been a central research topic in deep learning. For example, *Work partially done during internship at Meta.Pre-trained !!Fine-tuned !"Projected !"#(!)#(%&')#(%)… Param. 0Param. $−1Param. $Figure 1. Illustration of TPGM. TPGM learns different weight projection radii, , for each layer between a fine-tuned model ✓t and a pre-trained model ✓0and enforces the constraints through projection to obtain a projected model ˜✓. domain adaptation [ 39,46], domain generalization [ 25,50], and out-of-distribution calibration [ 33] are examples of re- lated fields. More recently, large pre-trained models, such as CLIP [ 28] (pre-trained on 400M image-text pairs), have demonstrated large gains in OOD robustness, thanks to the ever-increasing amount of pre-training data as well as ef- fective architectures and optimization methods. However, fine-tuning such models to other tasks generally results in worse OOD generalization as the model over-fits to the new data and forgets the pre-trained features [ 28].A natural goal is to preserve the generalization capability acquired by the pre-trained model when fine-tuning it to a downstream task. A recent empirical study shows that aggressive fine- tuning strategies such as using a large learning rate can de- crease OOD robustness [ 41]. We hypothesize that the for- getting of the generalization capability of the pre-trained model in the course of fine-tuning is due to unconstrained optimization on the new training data [ 44]. This conjec- ture is not surprising, because several prior works, even though they did not focus on OOD robustness, have dis- covered that encouraging a close distance to the pre-trained model weights can improve ID generalization, i.e., avoid- ing over-fitting to the training data [ 9,44]. Similarly, if suit- able distance constraints are enforced, we expect the model to behave more like the pre-trained model and thus retain 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. 7836 more of its generalization capability. The question is where to enforce distance constraints and how to optimize them? Several works have demonstrated the importance of treating each layer differently during fine-tuning. For ex- ample, a new work [ 22] discovers that selectively fine- tuning a subset of layers can lead to improved robustness to distribution shift. Another work [ 31] shows that opti- mizing a different learning rate for each layer is benefi- cial for few-shot learning. Therefore, we propose to en- force a different constraint for each layer. However, ex- isting works either use manually crafted heuristics or ex- pensive hyper-parameter search, which prevent them from scaling up to large datasets and neural networks. For exam- ple, the prior work [ 31] using evolutionary search for hyper- parameters can only scale up to a custom 6-layer ConvNet and a ResNet-12 for few-shot learning. The computation and time for searching hyper-parameters become increas- ingly infeasible for larger datasets, let alone scaling up the combinatorial search space to all layers. For example, a ViT-base [ 37] model has 154 trainable parameter groups in- cluding both weights, biases, and embeddings1. This leads to a search space with more than 1045combinations even if we allow only two choices per constraint parameter, which makes the search prohibitively expensive. tian2021geometricTo solve this problem, we propose a trainable projected gradient method (TPGM) to support layer-wise regularization optimization. Specifically, TPGM adopts trainable weight projection constraints , which we refer to as projection radii , and incorporates them in the forward pass of the main model to optimize. Intuitively, as shown in Fig. 1, TPGM maintains a set of weight projection radii i.e., the distance between the pre-trained model (✓0) and the current fine-tuned model ( ✓t), for each layer of a neural network and updates them. The projection radii control how much ”freedom” each layer has to grow. For example, if the model weights increase outside of the norm ball defined by andk·k, the projection operator will project them back to be within the constraints. To learn the weight projection radii in a principled manner, we propose to use alternating optimization between the model weights and the projection radii, motivated by formulating fine-tuning as a bi-level constrained problem (Sec. 3.1). We theoretically show that the bi-level formulation is the key to learning constraints for each layer (Sec. 3.4). Empirically, we conduct thorough experiments on large- scale datasets, DomainNet [ 27] and ImageNet [ 4], using dif- ferent architectures. Under the premise of preserving ID performance, i.e., OOD robustness should not come at the expense of worse ID accuracy, TPGM outperforms exist- ing approaches with little effort for hyper-parameter tun- ing. Further analysis of the learned projection radii reveals 1For example, for a linear layer y=Wx+b, we need to use separate distance constraints for Wandb.that lower layers (layers closer to the input) in a network require stronger regularization while higher layers (layers closer to the output) need more flexibility. This observation is in line with the common belief that lower layers learn more general features while higher layers specialize to each dataset [ 26,29,41,45]. Therefore, when conducting trans- fer learning such as fine-tuning, we need to treat each layer differently. Our contributions are summarized below. •We propose a trainable projected gradient method (TPGM) for fine-tuning to automatically learn the distance constraints for each layer in a neural network during fine-tuning. •We conduct experiments on different datasets and architectures to show significantly improved OOD generalization while matching ID performance. •We theoretically study TPGM using linear models to show that bi-level optimization is the key to learning different constraints for each layer

Sun_Learning_Semantic-Aware_Disentangled_Representation_for_Flexible_3D_Human_Body_Editing_CVPR_2023

Abstract 3D human body representation learning has received in- creasing attention in recent years. However, existing works cannot flexibly, controllably and accurately represent hu- man bodies, limited by coarse semantics and unsatisfac- tory representation capability, particularly in the absence of supervised data. In this paper, we propose a human body representation with fine-grained semantics and high reconstruction-accuracy in an unsupervised setting. Specif- ically, we establish a correspondence between latent vectors and geometric measures of body parts by designing a part- aware skeleton-separated decoupling strategy, which facili- tates controllable editing of human bodies by modifying the corresponding latent codes. With the help of a bone-guided auto-encoder and an orientation-adaptive weighting strat- egy, our representation can be trained in an unsupervised manner. With the geometrically meaningful latent space, it can be applied to a wide range of applications, from human body editing to latent code interpolation and shape style transfer. Experimental results on public datasets demon- strate the accurate reconstruction and flexible editing abil- ities of the proposed method. The code will be available *Corresponding authorathttp://cic.tju.edu.cn/faculty/likun/ projects/SemanticHuman .

1. Introduction Learning low-dimensional representations of human bodies plays an important role in various applications in- cluding human body reconstruction [4, 19, 32, 37], gener- ation [7, 30, 31] and editing [35, 36, 39]. Existing meth- ods [2, 18, 22, 25, 29] are either too semantically coarse to enable personalized human body editing, or suffer from poor reconstruction performance due to limited representa- tion capability. This paper aims to develop a fine-grained semantic-aware human body representation with flexible representation ability. Since human bodies are rich in variations of poses and shapes, traditional linear models [1, 25, 29, 35, 36] can- not handle complex nonlinear structures of human body meshes accurately. Therefore, parametric models have been proposed for better representation. The landmark works SCAPE [2] and SMPL [22] represent human bodies by the shape and pose parameters. However, the semantics of their shape parameters are not sufficiently precise, making it im- possible to flexibly edit the body shape. Furthermore, 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. 16985 representation ability of these methods is limited by the lin- ear shape space of human body shapes, and hence their re- construction accuracy is often unsatisfactory. With the success of deep learning, the encoder-decoder architecture has demonstrated excellent representation ca- pability [7,10,13,14,26]. Such methods improve the recon- struction precision by constructing different convolution- like operators for feature extraction on irregular meshes. However, these works lack disentangled representation and fail to obtain promising results for geometrically complex human body parts. Several works [3,9,11,18,38] pursue the disentanglement of latent representations, i.e., each latent code has clear semantics. But these methods either require paired supervised data or have poor performance on the re- construction, which significantly affects their generalization and robustness. In addition, the semantics of the above rep- resentations are coarse, which only enables person-level at- tribute transfer and cannot be applied to part-level flexible editing. In this paper, we aim to build a human body represen- tation with fine-grained semantics and high reconstruction- accuracy in an unsupervised setting, which needs to over- come two main challenges. First, how to disentangle the human body to reconstruct precise semantics is a key but difficult problem. Although it is straightforward to decom- pose a human body into articulated parts for part-level edit- ing, the hidden space of each part is still coupled. Secondly, providing paired supervised data requires a lot of manual effort, and it is very challenging to make the representation disentangled without sacrificing reconstruction accuracy in an unsupervised manner. To address these challenges, we propose SemanticHu- man, an editable human body representation with fine- grained semantics and high reconstruction-precision, which facilitates controllable human body editing without paired supervised data. To reconstruct fine-grained semantics, we design a part-aware skeleton-separated decoupling strategy with anatomical priors of the human body. Specifically, we disentangle body part variations into bone-related vari- ations ( e.g., length and orientation variations) and bone- independent variations ( e.g., circumference variations). In contrast to the previous pose and shape disentanglement on the entire person [2, 18, 22], this part-aware skeleton- separated decoupling strategy establishes a correspondence between latent vectors and geometric properties of body parts, which benefits part-level controllable editing. To ensure high reconstruction accuracy and fine-grained semantics of the representation by unsupervised learn- ing, we propose a bone-guided autoencoder architecture and an orientation-adaptive geometry-preserving loss. The bone-guided auto-encoder fuses the geometric features of body parts with their joint information to achieve accu- rate and efficient modeling of human bodies. Besides,an orientation-adaptive weighting strategy is introduced to compute the geometry-preserving loss, which can provide effective geometric regularization for unsupervised disen- tanglement and part-level editing. Experimental results on two public datasets with different mesh connectivities demonstrate the high reconstruction-precision and control- lable editing capability of the proposed method. An ex- ample is given in Fig. 1. The code will be available athttp://cic.tju.edu.cn/faculty/likun/ projects/SemanticHuman . Our main contributions are summarized as follows: • We propose a semantic-aware and editable human body representation with fine-grained representation ability. The latent space of our approach facilitates per- sonalized editing of human bodies by modifying their latent vectors. • We propose a part-aware skeleton-separated decou- pling strategy exploiting structural priors of the human body to learn geometrically meaningful latent codes with fine-grained semantics. • We propose a bone-guided auto-encoder architecture and an orientation-adaptive geometry-preserving loss to ensure the robust and effective disentanglement of the representation learned without supervision.

Tomar_TeSLA_Test-Time_Self-Learning_With_Automatic_Adversarial_Augmentation_CVPR_2023

Abstract Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with auto- matic Adversarial augmentation dubbed TeSLA for adapt- ing a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning meth- ods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowl- edge distillation by simulating high entropy augmented im- ages. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several de- sirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available at https://github.com/devavratTomar/TeSLA.

1. Introduction Deep neural networks (DNNs) perform exceptionally well when the training ( source ) and test ( target ) data fol- low the same distribution. However, distribution shifts are inevitable in real-world settings and propose a major chal- lenge to the performance of deep networks after deployment. Also, access to the labeled training data may be infeasible at test time due to privacy concerns or transmission band- width. In such scenarios, source-free domain adaptation (SFDA ) [1, 23, 25] and test-time adaptation (TTA ) meth- ods [19, 28, 39] aim to adapt the pre-trained source model to the unlabeled distributionally shifted target domain while easing access to source data. While SFDA methods have access to all full target data through multiple training epochs Soft-pseudo Label Unaugmented Soft-Pseudo Label Augmented Augmentation Direction Correctly Classified Incorrectly Classified Augmented Image Uncertainty Region(b) Hard Image to Hard Augmentation Decision Boundary Decision Boundary (a) Easy Image to Hard Augmentation Figure 1. Knowledge distillation with adversarial augmenta- tions. (a) Easy images with confident soft-pseudo labels and (b) Hard images with unconfident soft-pseudo labels are adversarially augmented and pushed to the uncertainty region (high entropy) near the decision boundary. The model is updated for (a)to match its output on the augmented views with non-augmented views of Easy test images using KL-Divergence Lkd̸= 0, while not updated for (b)asLkd∼0between Hard images and their augmented views. (offline setup), TTA methods usually process test images in an online streaming fashion and represent a more realistic domain adaptation. However, most of these methods are applied: (i) only to classification tasks, (ii) evaluated on the non-real-world domain shifts, e.g., the non-measurement shift; (iii) destroy model calibration—entropy minimizing with overconfident predictions [45] on incorrectly classified samples, and (iv) use specialized network architectures or rely on the source dataset feature statistics [28]. We address these issues by proposing a new test-time adaptation method with automatic adversarial augmentation called TeSLA , under which we further define realistic TTA protocols. Self-learning methods often supervise the model adaptation on the unlabeled test images using their predicted pseudo-labels. As the model can easily overfit on its own pseudo-labels, a weight-averaged teacher model (slowly up- dated by the student model) is employed for obtaining the pseudo-labels [41, 47]. The student model is then trained with cross-entropy ( CE) loss between the one-hot pseudo- labels and its predictions on the test images. In this paper, we instead propose to minimize flipped cross-entropy between the student model’s predictions and the soft pseudo-labels (notice the reverse order) with the negative entropy of its marginalized predictions over the test images. In Sec. 3, we show that the proposed formulation is an equivalence to mutual information maximization implicitly corrected 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. 20341 by the teacher-student knowledge distillation via pseudo- labels, yielding performance improvement on various test- time adaptation protocols and compared to the basic CE optimization (cf. ablation in Fig. 5-(1)). Motivated by teacher-student knowledge distillation, an- other central tenet of our method is to assist the student model during adaptation in improving its performance on the hard-to-classify (high entropy) test images. For this purpose, we propose learning automatic adversarial aug- mentations (see Fig. 1) as a proxy for simulating images in the uncertainty region of the feature space. The model is then updated to ensure consistency between predictions of high entropy augmented images and soft-pseudo labels from the respective non-augmented versions. Consequently, the model is self-distilled on Easy test images with confident soft-pseudo labels (Fig. 1a). In contrast, the model update onHard test images is discarded (Fig. 1b), resulting in better class-wise feature separation. In summary, our contributions are: (i) we propose a novel test-time self-learning method based on flipped cross-entropy (f-CE) through the tight connection with the mutual informa- tion between the model’s predictions and the test images; (ii) we propose an efficient plug-in test-time automatic adversar- ial augmentation module used for online knowledge distilla- tion from the teacher to the student network that consistently improves the performance of test-time adaptation methods, including ours; and (iii) TeSLA achieves new state-of-the-art results on several benchmarks, from common image corrup- tion to realistic measurement shifts for classification and segmentation tasks. Furthermore, TeSLA outperforms ex- isting TTA methods in terms of calibration and uncertainty metrics while making no assumptions about the network architecture and source domain’s information, e.g., feature statistics or training strategy.