CVPR 2022 论文列表

Convolution and self-attention are two powerful techniques for representation learning, and they are usually considered as two peer approaches that are distinct from each other. In this paper, we show that there exists a strong underlying relation between them, in the sense that the bulk of computations of these two paradigms are in fact done with the same operation. Specifically, we first show that a traditional convolution with kernel size k x k can be decomposed into k^2 individual 1x1 convolutions, followed by shift and summation operations. Then, we interpret the projections of queries, keys, and values in selfattention module as multiple 1x1 convolutions, followed by the computation of attention weights and aggregation of the values. Therefore, the first stage of both two modules comprises the similar operation. More importantly, the first stage contributes a dominant computation complexity (square of the channel size) comparing to the second stage. This observation naturally leads to an elegant integration of these two seemingly distinct paradigms, i.e., a mixed model that enjoys the benefit of both self-Attention and Convolution (ACmix), while having minimum computational overhead compared to the pure convolution or selfattention counterpart. Extensive experiments show that our model achieves consistently improved results over competitive baselines on image recognition and downstream tasks. Code and pre-trained models will be released at https://github.com/LeapLabTHU/ACmix and https://gitee.com/mindspore/models.

Relative pose estimation using the 5-point or 7-point Random Sample Consensus (RANSAC) algorithms can fail even when no outliers are present and there are enough inliers to support a hypothesis. These cases arise due to numerical instability of the 5- and 7-point minimal problems. This paper characterizes these instabilities, both in terms of minimal world scene configurations that lead to infinite condition number in epipolar estimation, and also in terms of the related minimal image feature pair correspondence configurations. The instability is studied in the context of a novel framework for analyzing the conditioning of minimal problems in multiview geometry, based on Riemannian manifolds. Experiments with synthetic and real-world data reveal that RANSAC does not only serve to filter out outliers, but RANSAC also selects for well-conditioned image data, sufficiently separated from the ill-posed locus that our theory predicts. These findings suggest that, in future work, one could try to accelerate and increase the success of RANSAC by testing only well-conditioned image data.

Many recent self-supervised frameworks for visual representation learning are based on certain forms of Siamese networks. Such networks are conceptually symmetric with two parallel encoders, but often practically asymmetric as numerous mechanisms are devised to break the symmetry. In this work, we conduct a formal study on the importance of asymmetry by explicitly distinguishing the two encoders within the network -- one produces source encodings and the other targets. Our key insight is keeping a relatively lower variance in target than source generally benefits learning. This is empirically justified by our results from five case studies covering different variance-oriented designs, and is aligned with our preliminary theoretical analysis on the baseline. Moreover, we find the improvements from asymmetric designs generalize well to longer training schedules, multiple other frameworks and newer backbones. Finally, the combined effect of several asymmetric designs achieves a state-of-the-art accuracy on ImageNet linear probing and competitive results on downstream transfer. We hope our exploration will inspire more research in exploiting asymmetry for Siamese representation learning.

Deep neural networks are able to memorize noisy labels easily with a softmax cross entropy (CE) loss. Previous studies attempted to address this issue focus on incorporating a noise-robust loss function to the CE loss. However, the memorization issue is alleviated but still remains due to the non-robust CE loss. To address this issue, we focus on learning robust contrastive representations of data on which the classifier is hard to memorize the label noise under the CE loss. We propose a novel contrastive regularization function to learn such representations over noisy data where the label noise does not dominate the representation learning. By theoretically investigating the representations induced by the proposed regularization function, we reveal that the learned representations keep information related to true labels and discard information related to corrupted labels from images. Moreover, our theoretical results also indicate that the learned representations are robust to the label noise. Experiments on benchmark datasets demonstrate that the efficacy of our method.

While real world challenges typically define visual categories with language words or phrases, most visual classification methods define categories with numerical indicies. However, the language specification of the classes provides an especially useful prior for biased and noisy datasets, where it can help disambiguate what features are task-relevant. Recently, large-scale multimodal models have been shown to recognize a wide variety of high-level concepts from a language specification even without additional image training data, but they are often unable to distinguish classes for more fine-grained tasks. CNNs, in contrast, can extract subtle image features that are required for fine-grained discrimination, but will overfit to any bias or noise in datasets. Our insight is to use high-level language specification as advice for constraining the prediction evidence to task-relevant features, instead of distractors. To do this, we ground task-relevant words or phrases with attention maps from a pretrained large-scale model. We then use this grounding to supervise a classifier's spatial attention away from distracting context. We show that supervising spatial attention in this way improves performance on classification tasks with biased and noisy data, including 3-15% worst-group accuracy improvements and 41-45% relative improvements on fairness metrics.

In the real world, humans have the ability to accumulate new knowledge in any conditions. However, deeplearning suffers from the phenomenon so-called catastrophic forgetting of the previously observed knowledge after learning a new task. Many recent methods focus on preventing catastrophic forgetting under a typical assumption of thetrain and test data following a similar distribution. In thiswork, we consider the more realistic scenario of continuallearning under domain shifts where the model is able to gen-eralize its inference to a an unseen domain. To this end, wepropose to make use of sample correlations of the learning tasks in the classifiers where the subsequent optimization isperformed over similarity measures obtained in a similar fashion to the Mahalanobis distance computation. In addition, we also propose an approach based on the exponential moving average of the parameters for better knowledge distillation, allowing a further adaptation to the old model. We demonstrate in our experiments that, to a great extent, the past continual learning algorithms fail to handle the forgetting issue under multiple distributions, while our proposed approach identifies the task under domain shift where insome cases can boost up the performance up to 10% on the challenging datasets e.g., DomainNet and OfficeHome.

Metrics for evaluating generative models aim to measure the discrepancy between real and generated images. The oftenused Frechet Inception Distance (FID) metric, for example, extracts "high-level" features using a deep network from the two sets. However, we find that the differences in "low-level" preprocessing, specifically image resizing and compression, can induce large variations and have unforeseen consequences. For instance, when resizing an image, e.g., with a bilinear or bicubic kernel, signal processing principles mandate adjusting prefilter width depending on the downsampling factor, to antialias to the appropriate bandwidth. However, commonly used implementations use a fixed-width prefilter, resulting in aliasing artifacts. Such aliasing leads to corruptions in the feature extraction downstream. Next, lossy compression, such as JPEG, is commonly used to reduce the file size of an image. Although designed to minimally degrade the perceptual quality of an image, the operation also produces variations downstream. Furthermore, we show that if compression is used on real training images, FID can actually improve if the generated images are also subsequently compressed. This paper shows that choices in low-level image processing have been an under-appreciated aspect of generative modeling. We identify and characterize variations in generative modeling development pipelines, provide recommendations based on signal processing principles, and release a reference implementation to facilitate future comparisons.

Trajectory prediction is a critical component for autonomous vehicles (AVs) to perform safe planning and navigation. However, few studies have analyzed the adversarial robustness of trajectory prediction or investigated whether the worst-case prediction can still lead to safe planning. To bridge this gap, we study the adversarial robustness of trajectory prediction models by proposing a new adversarial attack that perturbs normal vehicle trajectories to maximize the prediction error. Our experiments on three models and three datasets show that the adversarial prediction increases the prediction error by more than 150%. Our case studies show that if an adversary drives a vehicle close to the target AV following the adversarial trajectory, the AV may make an inaccurate prediction and even make unsafe driving decisions. We also explore possible mitigation techniques via data augmentation and trajectory smoothing.

Prior work has studied different visual modalities in isolation and developed separate architectures for recognition of images, videos, and 3D data. Instead, in this paper, we propose a single model which excels at classifying images, videos, and single-view 3D data using exactly the same model parameters. Our 'OMNIVORE' model leverages the flexibility of transformer-based architectures and is trained jointly on classification tasks from different modalities. OMNIVORE is simple to train, uses off-the-shelf standard datasets, and performs at-par or better than modality-specific models of the same size. A single OMNIVORE model obtains 86.0% on ImageNet, 84.1% on Kinetics, and 67.1% on SUN RGB-D. After finetuning, our models outperform prior work on a variety of vision tasks and generalize across modalities. OMNIVORE's shared visual representation naturally enables cross-modal recognition without access to correspondences between modalities. We hope our results motivate researchers to model visual modalities together.

A well-known challenge in applying deep-learning methods to omnidirectional images is spherical distortion. In dense regression tasks such as depth estimation, where structural details are required, using a vanilla CNN layer on the distorted 360 image results in undesired information loss. In this paper, we propose a 360 monocular depth estimation pipeline, OmniFusion, to tackle the spherical distortion issue. Our pipeline transforms a 360 image into less-distorted perspective patches (i.e. tangent images) to obtain patch-wise predictions via CNN, and then merge the patch-wise results for final output. To handle the discrepancy between patch-wise predictions which is a major issue affecting the merging quality, we propose a new framework with the following key components. First, we propose a geometry-aware feature fusion mechanism that combines 3D geometric features with 2D image features to compensate for the patch-wise discrepancy. Second, we employ the self-attention-based transformer architecture to conduct a global aggregation of patch-wise information, which further improves the consistency. Last, we introduce an iterative depth refinement mechanism, to further refine the estimated depth based on the more accurate geometric features. Experiments show that our method greatly mitigates the distortion issue, and achieves state-of-the-art performances on several 360 monocular depth estimation benchmark datasets.

We consider the problem of omni-supervised object detection, which can use unlabeled, fully labeled and weakly labeled annotations, such as image tags, counts, points, etc., for object detection. This is enabled by a unified architecture, Omni-DETR, based on the recent progress on student-teacher framework and end-to-end transformer based object detection. Under this unified architecture, different types of weak labels can be leveraged to generate accurate pseudo labels, by a bipartite matching based filtering mechanism, for the model to learn. In the experiments, Omni-DETR has achieved state-of-the-art results on multiple datasets and settings. And we have found that weak annotations can help to improve detection performance and a mixture of them can achieve a better trade-off between annotation cost and accuracy than the standard complete annotation. These findings could encourage larger object detection datasets with mixture annotations. The code is available at https://github.com/amazon-research/omni-detr.

RGBD-based real-time dynamic 3D reconstruction suffers from inaccurate inter-frame motion estimation as errors may accumulate with online tracking. This problem is even more severe for single-view-based systems due to strong occlusions. Based on these observations, we propose OcclusionFusion, a novel method to calculate occlusion-aware 3D motion to guide the reconstruction. In our technique, the motion of visible regions is first estimated and combined with temporal information to infer the motion of the occluded regions through an LSTM-involved graph neural network. Furthermore, our method computes the confidence of the estimated motion by modeling the network output with a probabilistic model, which alleviates untrustworthy motions and enables robust tracking. Experimental results on public datasets and our own recorded data show that our technique outperforms existing single-view-based real-time methods by a large margin. With the reduction of the motion errors, the proposed technique can handle long and challenging motion sequences. Please check out the project page for sequence results: https://wenbin-lin.github.io/OcclusionFusion.

The occlusion problem heavily degrades the localization performance of face alignment. Most current solutions for this problem focus on annotating new occlusion data, introducing boundary estimation, and stacking deeper models to improve the robustness of neural networks. However, the performance degradation of models remains under extreme occlusion (average occlusion of over 50%) because of missing a large amount of facial context information. We argue that exploring neural networks to model the facial hierarchies is a more promising method for dealing with extreme occlusion. Surprisingly, in recent studies, little effort has been devoted to representing the facial hierarchies using neural networks. This paper proposes a new network architecture called GlomFace to model the facial hierarchies against various occlusions, which draws inspiration from the viewpoint-invariant hierarchy of facial structure. Specifically, GlomFace is functionally divided into two modules: the part-whole hierarchical module and the whole-part hierarchical module. The former captures the part-whole hierarchical dependencies of facial parts to suppress multi-scale occlusion information, whereas the latter injects structural reasoning into neural networks by building the whole-part hierarchical relations among facial parts. As a result, GlomFace has a clear topological interpretation due to its correspondence to the facial hierarchies. Extensive experimental results indicate that the proposed GlomFace performs comparably to existing state-of-the-art methods, especially in cases of extreme occlusion. Models are available at https://github.com/zhuccly/GlomFace-Face-Alignment.

Matching cost construction is a key step in light field (LF) depth estimation, but was rarely studied in the deep learning era. Recent deep learning-based LF depth estimation methods construct matching cost by sequentially shifting each sub-aperture image (SAI) with a series of predefined offsets, which is complex and time-consuming. In this paper, we propose a simple and fast cost constructor to construct matching cost for LF depth estimation. Our cost constructor is composed by a series of convolutions with specifically designed dilation rates. By applying our cost constructor to SAI arrays, pixels under predefined disparities can be integrated and matching cost can be constructed without using any shifting operation. More importantly, the proposed cost constructor is occlusion-aware and can handle occlusions by dynamically modulating pixels from different views. Based on the proposed cost constructor, we develop a deep network for LF depth estimation. Our network ranks first on the commonly used 4D LF benchmark in terms of the mean square error (MSE), and achieves a faster running time than other state-of-the-art methods.

Top-down methods for monocular human mesh recovery have two stages: (1) detect human bounding boxes; (2) treat each bounding box as an independent single-human mesh recovery task. Unfortunately, the single-human assumption does not hold in images with multi-human occlusion and crowding. Consequently, top-down methods have difficulties in recovering accurate 3D human meshes under severe person-person occlusion. To address this, we present Occluded Human Mesh Recovery (OCHMR) - a novel top-down mesh recovery approach that incorporates image spatial context to overcome the limitations of the single-human assumption. The approach is conceptually simple and can be applied to any existing top-down architecture. Along with the input image, we condition the top-down model on spatial context from the image in the form of body-center heatmaps. To reason from the predicted body centermaps, we introduce Contextual Normalization (CoNorm) blocks to adaptively modulate intermediate features of the top-down model. The contextual conditioning helps our model disambiguate between two severely overlapping human bounding-boxes, making it robust to multi-person occlusion. Compared with state-of-the-art methods, OCHMR achieves superior performance on challenging multi-person benchmarks like 3DPW, CrowdPose, and OCHuman. Specifically, our proposed contextual reasoning architecture applied to the SPIN model with ResNet-50 backbone results in 75.2 PMPJPE on 3DPW-PC, 23.6 AP on CrowdPose, and 37.7 AP on OCHuman datasets, a significant improvement of 6.9 mm, 6.4 AP, and 20.8 AP respectively over the baseline. Code and models will be released.

While 3D object detection in LiDAR point clouds is well-established in academia and industry, the explainability of these models is a largely unexplored field. In this paper, we propose a method to generate attribution maps for the detected objects in order to better understand the behavior of such models. These maps indicate the importance of each 3D point in predicting the specific objects. Our method works with black-box models: We do not require any prior knowledge of the architecture nor access to the model's internals, like parameters, activations or gradients. Our efficient perturbation-based approach empirically estimates the importance of each point by testing the model with randomly generated subsets of the input point cloud. Our sub-sampling strategy takes into account the special characteristics of LiDAR data, such as the depth-dependent point density. We show a detailed evaluation of the attribution maps and demonstrate that they are interpretable and highly informative. Furthermore, we compare the attribution maps of recent 3D object detection architectures to provide insights into their decision-making processes.

Recent advances in image editing techniques have posed serious challenges to the trustworthiness of multimedia data, which drives the research of image tampering detection. In this paper, we propose ObjectFormer to detect and localize image manipulations. To capture subtle manipulation traces that are no longer visible in the RGB domain, we extract the high-frequency features of the images and combine them with RGB features as multimodal patch embeddings. In order to detect forgery traces, we use a set of learnable object prototypes as mid-level representations to model the object-level consistencies among different regions, which are further used to refine patch embeddings to capture the patch-level consistencies. We conduct extensive experiments on various datasets and the results verify the effectiveness of the proposed method, outperforming state-of-the-art tampering detection and localization methods.

Objects play a crucial role in our everyday activities. Though multisensory object-centric learning has shown great potential lately, the modeling of objects in prior work is rather unrealistic. ObjectFolder 1.0 is a recent dataset that introduces 100 virtualized objects with visual, auditory, and tactile sensory data. However, the dataset is small in scale and the multisensory data is of limited quality, hampering generalization to real-world scenarios. We present ObjectFolder 2.0, a large-scale, multisensory dataset of common household objects in the form of implicit neural representations that significantly enhances ObjectFolder 1.0 in three aspects. First, our dataset is 10 times larger in the amount of objects and orders of magnitude faster in rendering time. Second, we significantly improve the multisensory rendering quality for all three modalities. Third, we show that models learned from virtual objects in our dataset successfully transfer to their real-world counterparts in three challenging tasks: object scale estimation, contact localization, and shape reconstruction. ObjectFolder 2.0 offers a new path and testbed for multisensory learning in computer vision and robotics. The dataset is available at https://github.com/rhgao/ObjectFolder.

Action recognition is a challenging task since the attributes of objects as well as their relationships change constantly in the video. Existing methods mainly use object-level graphs or scene graphs to represent the dynamics of objects and relationships, but ignore modeling the fine-grained relationship transitions directly. In this paper, we propose an Object-Relation Reasoning Graph (OR2G) for reasoning about action in videos. By combining an object-level graph (OG) and a relation-level graph (RG), the proposed OR2G catches the attribute transitions of objects and reasons about the relationship transitions between objects simultaneously. In addition, a graph aggregating module (GAM) is investigated by applying the multi-head edge-to-node message passing operation. GAM feeds back the information from the relation node to the object node and enhances the coupling between the object-level graph and the relation-level graph. Experiments in video action recognition demonstrate the effectiveness of our approach when compared with the state-of-the-art methods.