CVPR 2022 Papers

Recently, large-scale synthetic datasets are shown to be very useful for generalizable person re-identification. However, synthesized persons in existing datasets are mostly cartoon-like and in random dress collocation, which limits their performance. To address this, in this work, an automatic approach is proposed to directly clone the whole outfits from real-world person images to virtual 3D characters, such that any virtual person thus created will appear very similar to its real-world counterpart. Specifically, based on UV texture mapping, two cloning methods are designed, namely registered clothes mapping and homogeneous cloth expansion. Given clothes keypoints detected on person images and labeled on regular UV maps with clear clothes structures, registered mapping applies perspective homography to warp real-world clothes to the counterparts on the UV map. As for invisible clothes parts and irregular UV maps, homogeneous expansion segments a homogeneous area on clothes as a realistic cloth pattern or cell, and expand the cell to fill the UV map. Furthermore, a similarity-diversity expansion strategy is proposed, by clustering person images, sampling images per cluster, and cloning outfits for 3D character generation. This way, virtual persons can be scaled up densely in visual similarity to challenge model learning, and diversely in population to enrich sample distribution. Finally, by rendering the cloned characters in Unity3D scenes, a more realistic virtual dataset called ClonedPerson is created, with 5,621 identities and 887,766 images. Experimental results show that the model trained on ClonedPerson has a better generalization performance, superior to that trained on other popular real-world and synthetic person re-identification datasets. The ClonedPerson project is available at https://github.com/Yanan-Wang-cs/ClonedPerson.

Hyperbolic space can naturally embed hierarchies, unlike Euclidean space. Hyperbolic Neural Networks (HNNs) exploit such representational power by lifting Euclidean features into hyperbolic space for classification, outperforming Euclidean neural networks (ENNs) on datasets with known semantic hierarchies. However, HNNs underperform ENNs on standard benchmarks without clear hierarchies, greatly restricting HNNs' applicability in practice. Our key insight is that HNNs' poorer general classification performance results from vanishing gradients during backpropagation, caused by their hybrid architecture connecting Euclidean features to a hyperbolic classifier. We propose an effective solution by simply clipping the Euclidean feature magnitude while training HNNs. Our experiments demonstrate that clipped HNNs become super-hyperbolic classifiers: They are not only consistently better than HNNs which already outperform ENNs on hierarchical data, but also on-par with ENNs on MNIST, CIFAR10, CIFAR100 and ImageNet benchmarks, with better adversarial robustness and out-of-distribution detection.

Scanning Transmission Electron Microscopes (STEMs) acquire 2D images of a 3D sample on the scale of individual cell components. Unfortunately, these 2D images can be too noisy to be fused into a useful 3D structure and facilitating good denoisers is challenging due to the lack of clean-noisy pairs. Additionally, representing detailed 3D structure can be difficult even for clean data when using regular 3D grids. Addressing these two limitations, we suggest a differentiable image formation model for STEM, allowing to learn a joint model of 2D sensor noise in STEM together with an implicit 3D model. We show, that the combination of these models are able to successfully disentangle 3D signal and noise without supervision and outperform at the same time several baselines on synthetic and real data.

Today's VidSGG models are all proposal-based methods, i.e., they first generate numerous paired subject-object snippets as proposals, and then conduct predicate classification for each proposal. In this paper, we argue that this prevalent proposal-based framework has three inherent drawbacks: 1) The ground-truth predicate labels for proposals are partially correct. 2) They break the high-order relations among different predicate instances of a same subject-object pair. 3) VidSGG performance is upper-bounded by the quality of the proposals. To this end, we propose a new classification-then-grounding framework for VidSGG, which can avoid all the three overlooked drawbacks. Meanwhile, under this framework, we reformulate the video scene graphs as temporal bipartite graphs, where the entities and predicates are two types of nodes with time slots, and the edges denote different semantic roles between these nodes. This formulation takes full advantage of our new framework. Accordingly, we further propose a novel BIpartite Graph based SGG model: BIG. It consists of a classification stage and a grounding stage, where the former aims to classify the categories of all the nodes and the edges, and the latter tries to localize the temporal location of each relation instance. Extensive ablations on two VidSGG datasets have attested to the effectiveness of our framework and BIG. Code is available at https://github.com/Dawn-LX/VidSGG-BIG.

Class-incremental learning (CIL) has been widely studied under the setting of starting from a small number of classes (base classes). Instead, we explore an understudied real-world setting of CIL that starts with a strong model pre-trained on a large number of base classes. We hypothesize that a strong base model can provide a good representation for novel classes and incremental learning can be done with small adaptations. We propose a 2-stage training scheme, i) feature augmentation - cloning part of the backbone and fine-tuning it on the novel data, and ii) fusion - combining the base and novel classifiers into a unified classifier. Experiments show that the proposed method significantly outperforms state-of-the-art CIL methods on the large-scale ImageNet dataset (e.g. +10% overall accuracy than the best). We also propose and analyze understudied practical CIL scenarios, such as base-novel overlap with distribution shift. Our proposed method is robust and generalizes to all analyzed CIL settings.

Domain adaptive semantic segmentation aims to learn a model with the supervision of source domain data, and produce satisfactory dense predictions on unlabeled target domain. One popular solution to this challenging task is self-training, which selects high-scoring predictions on target samples as pseudo labels for training. However, the produced pseudo labels often contain much noise because the model is biased to source domain as well as majority categories. To address the above issues, we propose to directly explore the intrinsic pixel distributions of target domain data, instead of heavily relying on the source domain. Specifically, we simultaneously cluster pixels and rectify pseudo labels with the obtained cluster assignments. This process is done in an online fashion so that pseudo labels could co-evolve with the segmentation model without extra training rounds. To overcome the class imbalance problem on long-tailed categories, we employ a distribution alignment technique to enforce the marginal class distribution of cluster assignments to be close to that of pseudo labels. The proposed method, namely Class-balanced Pixel-level Self-Labeling (CPSL), improves the segmentation performance on target domain over state-of-the-arts by a large margin, especially on long-tailed categories.

Pseudo-label-based semi-supervised learning (SSL) has achieved great success on raw data utilization. However, its training procedure suffers from confirmation bias due to the noise contained in self-generated artificial labels. Moreover, the model's judgment becomes noisier in real-world applications with extensive out-of-distribution data. To address this issue, we propose a general method named Class-aware Contrastive Semi-Supervised Learning (CCSSL), which is a drop-in helper to improve the pseudo-label quality and enhance the model's robustness in the real-world setting. Rather than treating real-world data as a union set, our method separately handles reliable in-distribution data with class-wise clustering for blending into downstream tasks and noisy out-of-distribution data with image-wise contrastive for better generalization. Furthermore, by applying target re-weighting, we successfully emphasize clean label learning and simultaneously reduce noisy label learning. Despite its simplicity, our proposed CCSSL has significant performance improvements over the state-of-the-art SSL methods on the standard datasets CIFAR100 and STL10. On the real-world dataset Semi-iNat 2021, we improve FixMatch by 9.80% and CoMatch by 3.18%. Code is available https://github.com/TencentYoutuResearch/Classification-SemiCLS.

Deep learning models are known to suffer from the problem of catastrophic forgetting when they incrementally learn new classes. Continual learning for semantic segmentation (CSS) is an emerging field in computer vision. We identify a problem in CSS: A model tends to be confused between old and new classes that are visually similar, which makes it forget the old ones. To address this gap, we propose REMINDER - a new CSS framework and a novel class similarity knowledge distillation (CSW-KD) method. Our CSW-KD method distills the knowledge of a previous model on old classes that are similar to the new one. This provides two main benefits: (i) selectively revising old classes that are more likely to be forgotten, and (ii) better learning new classes by relating them with the previously seen classes. Extensive experiments on Pascal-VOC 2012 and ADE20k datasets show that our approach outperforms state-of-the-art methods on standard CSS settings by up to 7.07% and 8.49%, respectively.

Extracting class activation maps (CAM) is arguably the most standard step of generating pseudo masks for weakly-supervised semantic segmentation (WSSS). Yet, we find that the crux of the unsatisfactory pseudo masks is the binary cross-entropy loss (BCE) widely used in CAM. Specifically, due to the sum-over-class pooling nature of BCE, each pixel in CAM may be responsive to multiple classes co-occurring in the same receptive field. As a result, given a class, its hot CAM pixels may wrongly invade the area belonging to other classes, or the non-hot ones may be actually a part of the class. To this end, we introduce an embarrassingly simple yet surprisingly effective method: Reactivating the converged CAM with BCE by using softmax cross-entropy loss (SCE), dubbed ReCAM. Given an image, we use CAM to extract the feature pixels of every single class, and use them with the class label to learn another fully-connected layer (after the backbone) with SCE. Once converged, we extract ReCAM in the same way as in CAM. Thanks to the contrastive nature of SCE, the pixel response is disentangled into different classes and hence less mask ambiguity is expected. The evaluation on both PASCAL VOC and MS COCO shows that ReCAM can not only generates high-quality masks, but also supports plug-and-play in any CAM variant with little overhead.

Current stereo matching techniques are challenged by restricted searching space, occluded regions and sheer size. While single image depth estimation is spared from these challenges and can achieve satisfactory results with the extracted monocular cues, the lack of stereoscopic relationship renders the monocular prediction less reliable on its own, especially in highly dynamic or cluttered environments. To address these issues in both scenarios, we present an optic-chiasm-inspired self-supervised binocular depth estimation method, wherein a vision transformer (ViT) with gated positional cross-attention (GPCA) layers is designed to enable feature-sensitive pattern retrieval between views while retaining the extensive context information aggregated through self-attentions. Monocular cues from a single view are thereafter conditionally rectified by a blending layer with the retrieved pattern pairs. This crossover design is biologically analogous to the optic-chasma structure in the human visual system and hence the name, ChiTransformer. Our experiments show that this architecture yields substantial improvements over state-of-the-art self-supervised stereo approaches by 11%, and can be used on both rectilinear and non-rectilinear (e.g., fisheye) images.

It is well known that Winograd convolution algorithms speed up the widely used small-size convolutions. However, the problem of quantization of Winograd convolutions is challenging - while quantization of slower Winograd algorithms does not cause problems, quantization of faster Winograd algorithms often leads to a significant drop in the quality of models. We introduce a novel class of Winograd algorithms that balances the filter and input channels in the Winograd domain. Unlike traditional Winograd convolutions, the proposed convolution balances the ranges of input channels on the forward pass by scaling the input tensor using special balancing coefficients (the filter channels are balanced offline). As a result of balancing, the inputs and filters of the Winograd convolution are much easier to quantize. Thus, the proposed technique allows us to obtain models with quantized Winograd convolutions, the quality of which is significantly higher than the quality of models with traditional quantized Winograd convolutions. Moreover, we propose a special direct algorithm for calculating the balancing coefficients, which does not require additional model training. This algorithm makes it easy to obtain the post-training quantized balanced Winograd convolutions - one should just feed a few data samples to the model without training to calibrate special parameters. In addition, it is possible to initialize the balancing coefficients using this algorithm and further train them as trainable variables during Winograd quantization-aware training for greater quality improvement.

Certified patch defenses can guarantee robustness of an image classifier to arbitrary changes within a bounded contiguous region. But, currently, this robustness comes at a cost of degraded standard accuracies and slower inference times. We demonstrate how using vision transformers enables significantly better certified patch robustness that is also more computationally efficient and does not incur a substantial drop in standard accuracy. These improvements stem from the inherent ability of the vision transformer to gracefully handle largely masked images.

Classifying all cells in an organ is a relevant and difficult problem from plant developmental biology. We here abstract the problem into a new benchmark for node classification in a geo-referenced graph. Solving it requires learning the spatial layout of the organ including symmetries. To allow the convenient testing of new geometrical learning methods, the benchmark of Arabidopsis thaliana ovules is made available as a PyTorch data loader, along with a large number of precomputed features. Finally, we benchmark eight recent graph neural network architectures, finding that DeeperGCN currently works best on this problem.

Visual representations underlie object recognition tasks, but they often contain both robust and non-robust features. Our main observation is that image classifiers may perform poorly on out-of-distribution samples because spurious correlations between non-robust features and labels can be changed in a new environment. By analyzing procedures for out-of-distribution generalization with a causal graph, we show that standard classifiers fail because the association between images and labels is not transportable across settings. However, we then show that the causal effect, which severs all sources of confounding, remains invariant across domains. This motivates us to develop an algorithm to estimate the causal effect for image classification, which is transportable (i.e., invariant) across source and target environments. Without observing additional variables, we show that we can derive an estimand for the causal effect under empirical assumptions using representations in deep models as proxies. Theoretical analysis, empirical results, and visualizations show that our approach captures causal invariances and improves overall generalization.

Human-Object Interactions (HOI) detection, which aims to localize a human and a relevant object while recognizing their interaction, is crucial for understanding a still image. Recently, tranformer-based models have significantly advanced the progress of HOI detection. However, the capability of these models has not been fully explored since the Object Query of the model is always simply initialized as just zeros, which would affect the performance. In this paper, we try to study the issue of promoting transformerbased HOI detectors by initializing the Object Query with category-aware semantic information. To this end, we innovatively propose the Category-Aware Transformer Network (CATN). Specifically, the Object Query would be initialized via category priors represented by an external object detection model to yield a better performance. Moreover, such category priors can be further used for enhancing the representation ability of features via the attention mechanism. We have firstly verified our idea via the Oracle experiment by initializing the Object Query with the groundtruth category information. And then extensive experiments have been conducted to show that a HOI detection model equipped with our idea outperforms the baseline by a large margin to achieve a new state-of-the-art result.

Instance contrast for unsupervised representation learning has achieved great success in recent years. In this work, we explore the idea of instance contrastive learning in unsupervised domain adaptation (UDA) and propose a novel Category Contrast technique (CaCo) that introduces semantic priors on top of instance discrimination for visual UDA tasks. By considering instance contrastive learning as a dictionary look-up operation, we construct a semantics-aware dictionary with samples from both source and target domains where each target sample is assigned a (pseudo) category label based on the category priors of source samples. This allows category contrastive learning (between target queries and the category-level dictionary) for category-discriminative yet domain-invariant feature representations: samples of the same category (from either source or target domain) are pulled closer while those of different categories are pushed apart simultaneously. Extensive UDA experiments in multiple visual tasks (e.g., segmentation, classification and detection) show that CaCo achieves superior performance as compared with state-of-the-art methods. The experiments also demonstrate that CaCo is complementary to existing UDA methods and generalizable to other learning setups such as unsupervised model adaptation, open-/partial-set adaptation etc.

Despite most existing anomaly detection studies assume the availability of normal training samples only, a few labeled anomaly examples are often available in many real-world applications, such as defect samples identified during random quality inspection, lesion images confirmed by radiologists in daily medical screening, etc. These anomaly examples provide valuable knowledge about the application-specific abnormality, enabling significantly improved detection of similar anomalies in some recent models. However, those anomalies seen during training often do not illustrate every possible class of anomaly, rendering these models ineffective in generalizing to unseen anomaly classes. This paper tackles open-set supervised anomaly detection, in which we learn detection models using the anomaly examples with the objective to detect both seen anomalies ('gray swans') and unseen anomalies ('black swans'). We propose a novel approach that learns disentangled representations of abnormalities illustrated by seen anomalies, pseudo anomalies, and latent residual anomalies (i.e., samples that have unusual residuals compared to the normal data in a latent space), with the last two abnormalities designed to detect unseen anomalies. Extensive experiments on nine real-world anomaly detection datasets show superior performance of our model in detecting seen and unseen anomalies under diverse settings. Code and data are available at: https://github.com/choubo/DRA.

The goal of person search is to localize a target person from a gallery set of scene images, which is extremely challenging due to large scale variations, pose/viewpoint changes, and occlusions. In this paper, we propose the Cascade Occluded Attention Transformer (COAT) for end-to-end person search. Our three-stage cascade design focuses on detecting people in the first stage, while later stages simultaneously and progressively refine the representation for person detection and re-identification. At each stage the occluded attention transformer applies tighter intersection over union thresholds, forcing the network to learn coarse-to-fine pose/scale invariant features. Meanwhile, we calculate each detection's occluded attention to differentiate a person's tokens from other people or the background. In this way, we simulate the effect of other objects occluding a person of interest at the token-level. Through comprehensive experiments, we demonstrate the benefits of our method by achieving state-of-the-art performance on two benchmark datasets.