CVPR 2022 论文列表

The ultimate goal of semi-supervised object detection (SSOD) is to facilitate the utilization and deployment of detectors in actual applications with the help of a large amount of unlabeled data. Although a few works have proposed various self-training-based methods or consistency-regularization-based methods, they all target anchor-based detectors, while ignoring the dependency on anchor-free detectors of the actual industrial deployment. To this end, in this paper, we intend to bridge the gap on anchor-free SSOD algorithm by proposing a DenSe Learning (DSL) based algorithm for SSOD. It is mainly achieved by introducing several novel techniques, including (1) Adaptive Ignoring strategy with MetaNet for assigning multi-level and accurate dense pixel-wise pseudo-labels, (2) Aggregated Teacher for producing stable and precise pseudo-labels, and (3) uncertainty consistency regularization among scales and shuffled patches for improving the generalization of the detector. In order to verify the effectiveness of our proposed method, extensive experiments have been conducted over the popular datasets MS-COCO [??] and PASCAL-VOC [??], achieving state-of-the-art performances. Codes will be available at \textcolor[rgb] 1,0,0 xxxxxxxxx .

Neural radiance fields (NeRF) encode a scene into a neural representation that enables photo-realistic rendering of novel views. However, a successful reconstruction from RGB images requires a large number of input views taken under static conditions - typically up to a few hundred images for room-size scenes. Our method aims to synthesize novel views of whole rooms from an order of magnitude fewer images. To this end, we leverage dense depth priors in order to constrain the NeRF optimization. First, we take advantage of the sparse depth data that is freely available from the structure from motion (SfM) preprocessing step used to estimate camera poses. Second, we use depth completion to convert these sparse points into dense depth maps and uncertainty estimates, which are used to guide NeRF optimization. Our method enables data-efficient novel view synthesis on challenging indoor scenes, using as few as 18 images for an entire scene.

In Neural Architecture Search (NAS), reducing the cost of architecture evaluation remains one of the most crucial challenges. Among a plethora of efforts to bypass training of each candidate architecture to convergence for evaluation, the Neural Tangent Kernel (NTK) is emerging as a promising theoretical framework that can be utilized to estimate the performance of a neural architecture at initialization. In this work, we revisit several at-initialization metrics that can be derived from the NTK and reveal their key shortcomings. Then, through the empirical analysis of the time evolution of NTK, we deduce that modern neural architectures exhibit highly non-linear characteristics, making the NTK-based metrics incapable of reliably estimating the performance of an architecture without some amount of training. To take such non-linear characteristics into account, we introduce Label-Gradient Alignment (LGA), a novel NTK-based metric whose inherent formulation allows it to capture the large amount of non-linear advantage present in modern neural architectures. With minimal amount of training, LGA obtains a meaningful level of rank correlation with the post-training test accuracy of an architecture. Lastly, we demonstrate that LGA, complemented with few epochs of training, successfully guides existing search algorithms to achieve competitive search performances with significantly less search cost. The code is available at: https://github.com/nutellamok/DemystifyingNTK.

Co-salient object detection, with the target of detecting co-existed salient objects among a group of images, is gaining popularity. Recent works use the attention mechanism or extra information to aggregate common co-salient features, leading to incomplete even incorrect responses for target objects. In this paper, we aim to mine comprehensive co-salient features with democracy and reduce background interference without introducing any extra information. To achieve this, we design a democratic prototype generation module to generate democratic response maps, covering sufficient co-salient regions and thereby involving more shared attributes of co-salient objects. Then a comprehensive prototype based on the response maps can be generated as a guide for final prediction. To suppress the noisy background information in the prototype, we propose a self-contrastive learning module, where both positive and negative pairs are formed without relying on additional classification information. Besides, we also design a democratic feature enhancement module to further strengthen the co-salient features by readjusting attention values. Extensive experiments show that our model obtains better performance than previous state-of-the-art methods, especially on challenging real-world cases (e.g., for CoCA, we obtain a gain of 2.0% for MAE, 5.4% for maximum F-measure, 2.3% for maximum E-measure, and 3.7% for S-measure) under the same settings. Code will be released soon.

Batch normalization (BN) is a milestone technique in deep learning. It normalizes the activation using mini-batch statistics during training but the estimated population statistics during inference. This paper focuses on investigating the estimation of population statistics. We define the estimation shift magnitude of BN to quantitatively measure the difference between its estimated population statistics and expected ones. Our primary observation is that the estimation shift can be accumulated due to the stack of BN in a network, which has detriment effects for the test performance. We further find a batch-free normalization (BFN) can block such an accumulation of estimation shift. These observations motivate our design of XBNBlock that replace one BN with BFN in the bottleneck block of residual-style networks. Experiments on the ImageNet and COCO benchmarks show that XBNBlock consistently improves the performance of different architectures, including ResNet and ResNeXt, by a significant margin and seems to be more robust to distribution shift.

Recently, Vision Transformers have achieved impressive results on various Vision tasks. Yet, their generalization ability under different distribution shifts is poorly understood. In this work, we provide a comprehensive study on the out-of-distribution generalization of Vision Transformers. To support a systematic investigation, we first present a taxonomy of distribution shifts by categorizing them into five conceptual levels: corruption shift, background shift, texture shift, destruction shift, and style shift. Then we perform extensive evaluations of Vision Transformer variants under different levels of distribution shifts and compare their generalization ability with Convolutional Neural Network (CNN) models. Several important observations are obtained: 1) Vision Transformers generalize better than CNNs under multiple distribution shifts. With the same or less amount of parameters, Vision Transformers are ahead of corresponding CNNs by more than 5% in top-1 accuracy under most types of distribution shift. In particular, Vision Transformers lead by more than 10% under the corruption shifts. 2) larger Vision Transformers gradually narrow the in-distribution (ID) and out-of-distribution (OOD) performance gap. To further improve the generalization of Vision Transformers, we design the enhanced Vision Transformers through self-supervised learning, information theory, and adversarial learning. By investigating these three types of generalization-enhanced Transformers, we observe the gradient-sensitivity of Vision Transformers and design a smoother learning strategy to achieve a stable training process. With modified training schemes, we achieve improvements on performance towards out-of-distribution data by 4% from vanilla Vision Transformers. We comprehensively compare these three types of generalization-enhanced Vision Transformers with their corresponding CNN models and observe that: 1) For the enhanced model, larger Vision Transformers still benefit more from the out-of-distribution generalization. 2) generalization-enhanced Vision Transformers are more sensitive to the hyper-parameters than their corresponding CNN models. We hope our comprehensive study could shed light on the design of more generalizable learning systems.

Convolutional neural network inference on video data requires powerful hardware for real-time processing. Given the inherent coherence across consecutive frames, large parts of a video typically change little. By skipping identical image regions and truncating insignificant pixel updates, computational redundancy can in theory be reduced significantly. However, these theoretical savings have been difficult to translate into practice, as sparse updates hamper computational consistency and memory access coherence; which are key for efficiency on real hardware. With DeltaCNN, we present a sparse convolutional neural network framework that enables sparse frame-by-frame updates to accelerate video inference in practice. We provide sparse implementations for all typical CNN layers and propagate sparse feature updates end-to-end - without accumulating errors over time. DeltaCNN is applicable to all convolutional neural networks without retraining. To the best of our knowledge, we are the first to significantly outperform the dense reference, cuDNN, in practical settings, achieving speedups of up to 7x with only marginal differences in accuracy.

Color constancy is an essential function in digital photography and a fundamental process for many computer vision applications. Accordingly, many methods have been proposed, and some recent ones have used deep neural networks to handle more complex scenarios. However, both the traditional and latest methods still impose strict assumptions on their target scenes in explicit or implicit ways. This paper shows that the degree of linear polarization dramatically solves the color constancy problem because it allows us to find achromatic pixels stably. Because we only rely on the physics-based polarization model, we significantly reduce the assumptions compared to existing methods. Furthermore, we captured a wide variety of scenes with ground-truth illuminations for evaluation, and the proposed approach achieved state-of-the-art performance with a low computational cost. Additionally, the proposed method can estimate illumination colors from chromatic pixels and manage multi-illumination scenes. Lastly, the evaluation scenes and codes are publicly available to encourage more development in this field.

In this paper, we study the problem of stereo matching from a pair of images with different resolutions, e.g., those acquired with a tele-wide camera system. Due to the difficulty of obtaining ground-truth disparity labels in diverse real-world systems, we start from an unsupervised learning perspective. However, resolution asymmetry caused by unknown degradations between two views hinders the effectiveness of the generally assumed photometric consistency. To overcome this challenge, we propose to impose the consistency between two views in a feature space instead of the image space, named feature-metric consistency. Interestingly, we find that, although a stereo matching network trained with the photometric loss is not optimal, its feature extractor can produce degradation-agnostic and matching-specific features. These features can then be utilized to formulate a feature-metric loss to avoid the photometric inconsistency. Moreover, we introduce a self-boosting strategy to optimize the feature extractor progressively, which further strengthens the feature-metric consistency. Experiments on both simulated datasets with various degradations and a self-collected real-world dataset validate the superior performance of the proposed method over existing solutions.

Point cloud scene flow estimation is of practical importance for dynamic scene navigation in autonomous driving. Since scene flow labels are hard to obtain, current methods train their models on synthetic data and transfer them to real scenes. However, large disparities between existing synthetic datasets and real scenes lead to poor model transfer. We make two major contributions to address that. First, we develop a point cloud collector and scene flow annotator for GTA-V engine to automatically obtain diverse realistic training samples without human intervention. With that, we develop a large-scale synthetic scene flow dataset GTA-SF. Second, we propose a mean-teacher-based domain adaptation framework that leverages self-generated pseudo-labels of the target domain. It also explicitly incorporates shape deformation regularization and surface correspondence refinement to address distortions and misalignments in domain transfer. Through extensive experiments, we show that our GTA-SF dataset leads to a consistent boost in model generalization to three real datasets (i.e., Waymo, Lyft and KITTI) as compared to the most widely used FT3D dataset. Moreover, our framework achieves superior adaptation performance on six source-target dataset pairs, remarkably closing the average domain gap by 60%. Data and codes are available at https://github.com/leolyj/DCA-SRSFE

Video transformers have recently emerged as an effective alternative to convolutional networks for action classification. However, most prior video transformers adopt either global space-time attention or hand-defined strategies to compare patches within and across frames. These fixed attention schemes not only have high computational cost but, by comparing patches at predetermined locations, they neglect the motion dynamics in the video. In this paper, we introduce the Deformable Video Transformer (DVT), which dynamically predicts a small subset of video patches to attend for each query location based on motion information, thus allowing the model to decide where to look in the video based on correspondences across frames. Crucially, these motion-based correspondences are obtained at zero-cost from information stored in the compressed format of the video. Our deformable attention mechanism is optimized directly with respect to classification performance, thus eliminating the need for suboptimal hand-design of attention strategies. Experiments on four large-scale video benchmarks (Kinetics-400, Something-Something-V2, EPIC-KITCHENS and Diving-48) demonstrate that, compared to existing video transformers, our model achieves higher accuracy at the same or lower computational cost, and it attains state-of-the-art results on these four datasets.

We describe a method to extract persistent elements of a dynamic scene from an input video. We represent each scene element as a Deformable Sprite consisting of three components: 1) a 2D texture image for the entire video, 2) per-frame masks for the element, and 3) non-rigid deformations that map the texture image into each video frame. The resulting decomposition allows for applications such as consistent video editing. Deformable Sprites are a type of video auto-encoder model that is optimized on individual videos, and does not require training on a large dataset, nor does it rely on pre-trained models. Moreover, our method does not require object masks or other user input, and discovers moving objects of a wider variety than previous work. We evaluate our approach on standard video datasets and show qualitative results on a diverse array of Internet videos.

We present a deformable prototypical part network (Deformable ProtoPNet), an interpretable image classifier that integrates the power of deep learning and the interpretability of case-based reasoning. This model classifies input images by comparing them with prototypes learned during training, yielding explanations in the form of "this looks like that." However, while previous methods use spatially rigid prototypes, we address this shortcoming by proposing spatially flexible prototypes. Each prototype is made up of several prototypical parts that adaptively change their relative spatial positions depending on the input image. Consequently, a Deformable ProtoPNet can explicitly capture pose variations and context, improving both model accuracy and the richness of explanations provided. Compared to other case-based interpretable models using prototypes, our approach achieves state-of-the-art accuracy and gives an explanation with greater context. The code is available at https://github.com/jdonnelly36/Deformable-ProtoPNet.

To operate in real-world high-stakes environments, deep learning systems have to endure noises that have been continuously thwarting their robustness. Data-end defense, which improves robustness by operations on input data instead of modifying models, has attracted intensive attention due to its high feasibility in practice. However, previous data-end defenses show low generalization against diverse noises and weak transferability across multiple models. Motivated by the fact that robust recognition depends on both local and global features, we propose a defensive patch generation framework to address these problems by helping models better exploit these features. For the generalization against diverse noises, we inject class-specific identifiable patterns into a confined local patch prior, so that defensive patches could preserve more recognizable features towards specific classes, leading models for better recognition under noises. For the transferability across multiple models, we guide the defensive patches to capture more global feature correlations within a class, so that they could activate model-shared global perceptions and transfer better among models. Our defensive patches show great potentials to improve model robustness in practice by simply sticking them around target objects. Extensive experiments show that we outperform others by large margins (improve 20+% accuracy for both adversarial and corruption robustness on average in the digital and physical world).

Lidars and cameras are critical sensors that provide complementary information for 3D detection in autonomous driving. While prevalent multi-modal methods simply decorate raw lidar point clouds with camera features and feed them directly to existing 3D detection models, our study shows that fusing camera features with deep lidar features instead of raw points, can lead to better performance. However, as those features are often augmented and aggregated, a key challenge in fusion is how to effectively align the transformed features from two modalities. In this paper, we propose two novel techniques: InverseAug that inverses geometric-related augmentations, e.g., rotation, to enable accurate geometric alignment between lidar points and image pixels, and LearnableAlign that leverages cross-attention to dynamically capture the correlations between image and lidar features during fusion. Based on InverseAug and LearnableAlign, we develop a family of generic multi-modal 3D detection models named DeepFusion, which is more accurate than previous methods. For example, DeepFusion improves PointPillars, CenterPoint, and 3D-MAN baselines on Pedestrian detection for 6.7, 8.9, and 6.2 LEVEL_2 APH, respectively. Notably, our models achieve state-of-the-art performance on Waymo Open Dataset, and show strong model robustness against input corruptions and out-of-distribution data. Code will be publicly available at https://github.com/tensorflow/lingvo.

In recent years, with the advances of generative models, many powerful face manipulation systems have been developed based on Deep Neural Networks (DNNs), called DeepFakes. If DeepFakes are not controlled timely and properly, they would become a real threat to both celebrities and ordinary people. Precautions such as adding perturbations to the source inputs will make DeepFake results look distorted from the perspective of human eyes. However, previous method doesn't explore whether the disrupted images can still spoof DeepFake detectors. This is critical for many applications where DeepFake detectors are used to discriminate between DeepFake data and real data due to the huge cost of examining a large amount of data manually. We argue that the detectors do not share a similar perspective as human eyes, which might still be spoofed by the disrupted data. Besides, the existing disruption methods rely on iteration-based perturbation generation algorithms, which is time-consuming. In this paper, we propose a novel DeepFake disruption algorithm called "DeepFake Disrupter". By training a perturbation generator, we can add the human-imperceptible perturbations to source images that need to be protected without any backpropagation update. The DeepFake results of these protected source inputs would not only look unrealistic by the human eye but also can be distinguished by DeepFake detectors easily. For example, experimental results show that by adding our trained perturbations, fake images generated by StarGAN can result in a 10 20% increase in F1-score evaluated by various DeepFake detectors.

Face identification (FI) is ubiquitous and drives many high-stake decisions made by the law enforcement. State-of-the-art FI approaches compare two images by taking the cosine similarity between their image embeddings. Yet, such approach suffers from poor out-of-distribution (OOD) generalization to new types of images (e.g., when a query face is masked, cropped or rotated) not included in the training set or the gallery. Here, we propose a re-ranking approach that compares two faces using the Earth Mover's Distance on the deep, spatial features of image patches. Our extra comparison stage explicitly examines image similarity at a fine-grained level (e.g., eyes to eyes) and is more robust to OOD perturbations and occlusions than traditional FI. Interestingly, without finetuning feature extractors, our method consistently improves the accuracy on all tested OOD queries: masked, cropped, rotated, and adversarial while obtaining similar results on in-distribution images.

Deep Learning (DL) has shown great promise in the unsupervised task of clustering. That said, while in classical (i.e., non-deep) clustering the benefits of the nonparametric approach are well known, most deep-clustering methods are parametric: namely, they require a predefined and fixed number of clusters, denoted by K. When K is unknown, however, using model-selection criteria to choose its optimal value might become computationally expensive, especially in DL as the training process would have to be repeated numerous times. In this work, we bridge this gap by introducing an effective deep-clustering method that does not require knowing the value of K as it infers it during the learning. Using a split/merge framework, a dynamic architecture that adapts to the changing K, and a novel loss, our proposed method outperforms existing nonparametric methods (both classical and deep ones). While the very few existing deep nonparametric methods lack scalability, we demonstrate ours by being the first to report the performance of such a method on ImageNet. We also demonstrate the importance of inferring K by showing how methods that fix it deteriorate in performance when their assumed K value gets further from the ground-truth one, especially on imbalanced datasets. Our code is available at https://github.com/BGU-CS-VIL/DeepDPM.