CVPR 2022 论文列表

In this paper, we study Multiscale Vision Transformers (MViTv2) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTv2s' pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViTv2 has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 58.7 boxAP on COCO object detection as well as 86.1% on Kinetics-400 video classification. Code and models are available at https://github.com/facebookresearch/mvit.

Deep learning has made significant impacts on multi-view stereo systems. State-of-the-art approaches typically involve building a cost volume, followed by multiple 3D convolution operations to recover the input image's pixel-wise depth. While such end-to-end learning of plane-sweeping stereo advances public benchmarks' accuracy, they are typically very slow to compute. We present MVS2D, a highly efficient multi-view stereo algorithm that seamlessly integrates multi-view constraints into single-view networks via an attention mechanism. Since MVS2Donly builds on 2D convolutions, it is at least 2x faster than all the notable counterparts. Moreover, our algorithm produces precise depth estimations and 3D reconstructions, achieving state-of-the-art results on challenging benchmarks ScanNet, SUN3D, RGBD, and the classical DTU dataset. our algorithm also out-performs all other algorithms in the setting of inexact camera poses. Our code is released at https://github.com/zhenpeiyang/MVS2D

Accurate long-term trajectory prediction in complex scenes, where multiple agents (e.g., pedestrians or vehicles) interact with each other and the environment while attempting to accomplish diverse and often unknown goals, is a challenging stochastic forecasting problem. In this work, we propose MUSE-VAE, a new probabilistic modeling framework based on a cascade of Conditional VAEs, which tackles the long-term, uncertain trajectory prediction task using a coarse-to-fine multi-factor forecasting architecture. In its Macro stage, the model learns a joint pixel-space representation of two key factors, the underlying environment and the agent movements, to predict the long and short term motion goals. Conditioned on them, the Micro stage learns a fine-grained spatio-temporal representation for the prediction of individual agent trajectories. The VAE backbones across the two stages make it possible to naturally account for the joint uncertainty at both levels of granularity. As a result, MUSE-VAE offers diverse and simultaneously more accurate predictions compared to the current state-of-the-art. We demonstrate these assertions through a comprehensive set of experiments on nuScenes and SDD benchmarks as well as PFSD, a new synthetic dataset, which challenges the forecasting ability of models on complex agent-environment interaction scenarios.

Many recent semi-supervised learning (SSL) studies build teacher-student architecture and train the student network by the generated supervisory signal from the teacher. Data augmentation strategy plays a significant role in the SSL framework since it is hard to create a weak-strong augmented input pair without losing label information. Especially when extending SSL to semi-supervised object detection (SSOD), many strong augmentation methodologies related to image geometry and interpolation-regularization are hard to utilize since they possibly hurt the location information of the bounding box in the object detection task. To address this, we introduce a simple yet effective data augmentation method, Mix/UnMix (MUM), which unmixes feature tiles for the mixed image tiles for the SSOD framework. Our proposed method makes mixed input image tiles and reconstructs them in the feature space. Thus, MUM can enjoy the interpolation-regularization effect from non-interpolated pseudo-labels and successfully generate a meaningful weak-strong pair. Furthermore, MUM can be easily equipped on top of various SSOD methods. Extensive experiments on MS-COCO and PASCAL VOC datasets demonstrate the superiority of MUM by consistently improving the mAP performance over the baseline in all the tested SSOD benchmark protocols. The code is released at https://github.com/JongMokKim/mix-unmix.

Human-Object Interaction (HOI) detection is the task of identifying a set of <human, object, interaction> triplets from an image. Recent work proposed transformer encoder-decoder architectures that successfully eliminated the need for many hand-designed components in HOI detection through end-to-end training. However, they are limited to single-scale feature resolution, providing suboptimal performance in scenes containing humans, objects, and their interactions with vastly different scales and distances. To tackle this problem, we propose a Multi-Scale TRansformer (MSTR) for HOI detection powered by two novel HOI-aware deformable attention modules called Dual-Entity attention and Entity-conditioned Context attention. While existing deformable attention comes at a huge cost in HOI detection performance, our proposed attention modules of MSTR learn to effectively attend to sampling points that are essential to identify interactions. In experiments, we achieve the new state-of-the-art performance on two HOI detection benchmarks.

Transformers have offered a new methodology of designing neural networks for visual recognition. Compared to convolutional networks, Transformers enjoy the ability of referring to global features at each stage, yet the attention module brings higher computational overhead that obstructs the application of Transformers to process high-resolution visual data. This paper aims to alleviate the conflict between efficiency and flexibility, for which we propose a specialized token for each region that serves as a messenger (MSG). Hence, by manipulating these MSG tokens, one can flexibly exchange visual information across regions and the computational complexity is reduced. We then integrate the MSG token into a multi-scale architecture named MSG-Transformer. In standard image classification and object detection, MSG-Transformer achieves competitive performance and the inference on both GPU and CPU is accelerated. Code is available at https://github.com/hustvl/MSG-Transformer.

The key challenge of zero-shot learning (ZSL) is how to infer the latent semantic knowledge between visual and attribute features on seen classes, and thus achieving a desirable knowledge transfer to unseen classes. Prior works either simply align the global features of an image with its associated class semantic vector or utilize unidirectional attention to learn the limited latent semantic representations, which could not effectively discover the intrinsic semantic knowledge (e.g., attribute semantics) between visual and attribute features. To solve the above dilemma, we propose a Mutually Semantic Distillation Network (MSDN), which progressively distills the intrinsic semantic representations between visual and attribute features for ZSL. MSDN incorporates an attribute->visual attention sub-net that learns attribute-based visual features, and a visual->attribute attention sub-net that learns visual-based attribute features. By further introducing a semantic distillation loss, the two mutual attention sub-nets are capable of learning collaboratively and teaching each other throughout the training process. The proposed MSDN yields significant improvements over the strong baselines, leading to new state-of-the-art performances on three popular challenging benchmarks. Our source codes, pre-trained models, and more results have been available at the anonymous project website: https://anonymous.4open.science/r/MSDN.

Establishing superior-quality correspondences in an image pair is pivotal to many subsequent computer vision tasks. Using Euclidean distance between correspondences to find neighbors and extract local information is a common strategy in previous works. However, most such works ignore similar sparse semantics information between two given images and cannot capture local topology among correspondences well. Therefore, to deal with the above problems, Multiple Sparse Semantics Dynamic Graph Network (MS^ 2 DG-Net) is proposed, in this paper, to predict probabilities of correspondences as inliers and recover camera poses. MS^ 2 DG-Net dynamically builds sparse semantics graphs based on sparse semantics similarity between two given images, to capture local topology among correspondences, while maintaining permutation-equivariant. Extensive experiments prove that MS^ 2 DG-Net outperforms state-of-the-art methods in outlier removal and camera pose estimation tasks on the public datasets with heavy outliers. Source code:https://github.com/changcaiyang/MS2DG-Net

Action detection is an essential and challenging task, especially for densely labelled datasets of untrimmed videos. The temporal relation is complex in those datasets, including challenges like composite action, and co-occurring action. For detecting actions in those complex videos, efficiently capturing both short-term and long-term temporal information in the video is critical. To this end, we propose a novel ConvTransformer network for action detection. This network comprises three main components: (1) Temporal Encoder module extensively explores global and local temporal relations at multiple temporal resolutions. (2) Temporal Scale Mixer module effectively fuses the multi-scale features to have a unified feature representation. (3) Classification module is used to learn the instance center-relative position and predict the frame-level classification scores. The extensive experiments on multiple datasets, including Charades, TSU and MultiTHUMOS, confirm the effectiveness of our proposed method. Our network outperforms the state-of-the-art methods on all the three datasets.

Dense computer vision tasks such as object detection and segmentation require effective multi-scale feature representation for detecting or classifying objects or regions with varying sizes. While Convolutional Neural Networks (CNNs) have been the dominant architectures for such tasks, recently introduced Vision Transformers (ViTs) aim to replace them as a backbone. Similar to CNNs, ViTs build a simple multi-stage structure (i.e., fine-to-coarse) for multi-scale representation with single-scale patches. In this work, with a different perspective from existing Transformers, we explore multi-scale patch embedding and multi-path structure, constructing the Multi-Path Vision Transformer (MPViT). MPViT embeds features of the same size (i.e., sequence length) with patches of different scales simultaneously by using overlapping convolutional patch embedding. Tokens of different scales are then independently fed into the Transformer encoders via multiple paths and the resulting features are aggregated, enabling both fine and coarse feature representations at the same feature level. Thanks to the diverse, multi-scale feature representations, our MPViTs scaling from tiny (5M) to base (73M) consistently achieve superior performance over state-of-the-art Vision Transformers on ImageNet classification, object detection, instance segmentation, and semantic segmentation. These extensive results demonstrate that MPViT can serve as a versatile backbone network for various vision tasks.

Despite the promising progress having been made, the two challenges of multi-view clustering (MVC) are still waiting for better solutions: i) Most existing methods are either not qualified or require additional steps for incomplete multi-view clustering and ii) noise or outliers might significantly degrade the overall clustering performance. In this paper, we propose a novel unified framework for incomplete and complete MVC named multi-view probabilistic clustering (MPC). MPC equivalently transforms multi-view pairwise posterior matching probability into composition of each view's individual distribution, which tolerates data missing and might extend to any number of views. Then graph-context-aware refinement with path propagation and co-neighbor propagation is used to refine pairwise probability, which alleviates the impact of noise and outliers. Finally, MPC also equivalently transforms probabilistic clustering's objective to avoid complete pairwise computation and adjusts clustering assignments by maximizing joint probability iteratively. Extensive experiments on multiple benchmarks for incomplete and complete MVC show that MPC significantly outperforms previous state-of-the-art methods in both effectiveness and efficiency.

With the rapid development of display technology, it has become an urgent need to obtain realistic 3D surfaces with as high-quality as possible. Due to the unstructured and irregular nature of 3D object data, it is usually difficult to obtain high-quality surface details and geometry textures at a low cost. In this article, we propose an effective multimodal-driven deep neural network to perform 3D surface super-resolution in 2D normal domain, which is simple, accurate, and robust to the above difficulty. To leverage the multimodal information from different perspectives, we jointly consider the texture, depth, and normal modalities to simultaneously restore fine-grained surface details as well as preserve geometry structures. To better utilize the cross-modality information, we explore a two-bridge normal method with a transformer structure for feature alignment, and investigate an affine transform module for fusing multimodal features. Extensive experimental results on public and our newly constructed photometric stereo dataset demonstrate that the proposed method delivers promising surface geometry details compared with nine competitive schemes.

Test-time adaptation approaches have recently emerged as a practical solution for handling domain shift without access to the source domain data. In this paper, we propose and explore a new multi-modal extension of test-time adaptation for 3D semantic segmentation. We find that, directly applying existing methods usually results in performance instability at test time, because multi-modal input is not considered jointly. To design a framework that can take full advantage of multi-modality, where each modality provides regularized self-supervisory signals to other modalities, we propose two complementary modules within and across the modalities. First, Intra-modal Pseudo-label Generation (Intra-PG) is introduced to obtain reliable pseudo labels within each modality by aggregating information from two models that are both pre-trained on source data but updated with target data at different paces. Second, Intermodal Pseudo-label Refinement (Inter-PR) adaptively selects more reliable pseudo labels from different modalities based on a proposed consistency scheme. Experiments demonstrate that our regularized pseudo labels produce stable self-learning signals in numerous multi-modal test-time adaptation scenarios for 3D semantic segmentation. Visit our project website at https://www.nec-labs.com/ mas/MM-TTA.

Most of the research to date focuses on bilingual sign language translation (BSLT). However, such models are inefficient in building multilingual sign language translation systems. To solve this problem, we introduce the multilingual sign language translation (MSLT) task. It aims to use a single model to complete the translation between multiple sign languages and spoken languages. Then, we propose MLSLT, the first MSLT model, which contains two novel dynamic routing mechanisms for controlling the degree of parameter sharing between different languages. Intra-layer language-specific routing controls the proportion of data flowing through shared parameters and language-specific parameters from the token level through a soft gate within the layer, and inter-layer language-specific routing controls and learns the data flow path of different languages at the language level through a soft gate between layers. In order to evaluate the performance of MLSLT, we collect the first publicly available multilingual sign language understanding dataset, Spreadthesign-Ten (SP-10), which contains up to 100 language pairs, e.g., CSL->en, GSG->zh. Experimental results show that the average performance of MLSLT outperforms the baseline MSLT model and the combination of multiple BSLT models in many cases. In addition, we also explore zero-shot translation in sign language and find that our model can achieve comparable performance to the supervised BSLT model on some language pairs. Dataset and more details are at https://mlslt.github.io/.

Convolutional Neural Networks (CNNs) have been regarded as the go-to models for visual recognition. More recently, convolution-free networks, based on multi-head self-attention (MSA) or multi-layer perceptrons (MLPs), become more and more popular. Nevertheless, it is not trivial when utilizing these newly-minted networks for video recognition due to the large variations and complexities in video data. In this paper, we present MLP-3D networks, a novel MLP-like 3D architecture for video recognition. Specifically, the architecture consists of MLP-3D blocks, where each block contains one MLP applied across tokens (i.e., token-mixing MLP) and one MLP applied independently to each token (i.e., channel MLP). By deriving the novel grouped time mixing (GTM) operations, we equip the basic token-mixing MLP with the ability of temporal modeling. GTM divides the input tokens into several temporal groups and linearly maps the tokens in each group with the shared projection matrix. Furthermore, we devise several variants of GTM with different grouping strategies, and compose each variant in different blocks of MLP-3D network by greedy architecture search. Without the dependence on convolutions or attention mechanisms, our MLP-3D networks achieves 68.5%/81.4% top-1 accuracy on Something-Something V2 and Kinetics-400 datasets, respectively. Despite with fewer computations, the results are comparable to state-of-the-art widely-used 3D CNNs and video transformers.

Although achieving significant progress, existing deep generative inpainting methods still show low generalization across different scenes. As a result, the generated images usually contain artifacts or the filled pixels differ greatly from the ground truth, making them far from real-world applications. Image-level predictive filtering is a widely used restoration technique by predicting suitable kernels adaptively according to different input scenes. Inspired by this inherent advantage, we explore the possibility of addressing image inpainting as a filtering task. To this end, we first study the advantages and challenges of the image-level predictive filtering for inpainting: the method can preserve local structures and avoid artifacts but fails to fill large missing areas. Then, we propose the semantic filtering by conducting filtering on deep feature level, which fills the missing semantic information but fails to recover the details. To address the issues while adopting the respective advantages, we propose a novel filtering technique, i.e., Multi-level Interactive Siamese Filtering (MISF) containing two branches: kernel prediction branch (KPB) and semantic & image filtering branch (SIFB). These two branches are interactively linked: SIFB provides multi-level features for KPB while KPB predicts dynamic kernels for SIFB. As a result, the final method takes the advantage of effective semantic & image-level filling for high-fidelity inpainting. Moreover, we discuss the relationship between MISF and the naive encoder-decoder-based inpainting, inferring that MISF provides novel dynamic convolutional operations to enhance the high generalization capability across scenes. We validate our method on three challenging datasets, i.e., Dunhuang, Places2, and CelebA. Our method outperforms state-of-the-art baselines on four metrics, i.e., L1, PSNR, SSIM, and LPIPS.

Estimating 3D human poses from monocular videos is a challenging task due to depth ambiguity and self-occlusion. Most existing works attempt to solve both issues by exploiting spatial and temporal relationships. However, those works ignore the fact that it is an inverse problem where multiple feasible solutions (i.e., hypotheses) exist. To relieve this limitation, we propose a Multi-Hypothesis Transformer (MHFormer) that learns spatio-temporal representations of multiple plausible pose hypotheses. In order to effectively model multi-hypothesis dependencies and build strong relationships across hypothesis features, the task is decomposed into three stages: (i) Generate multiple initial hypothesis representations; (ii) Model self-hypothesis communication, merge multiple hypotheses into a single converged representation and then partition it into several diverged hypotheses; (iii) Learn cross-hypothesis communication and aggregate the multi-hypothesis features to synthesize the final 3D pose. Through the above processes, the final representation is enhanced and the synthesized pose is much more accurate. Extensive experiments show that MHFormer achieves state-of-the-art results on two challenging datasets: Human3.6M and MPI-INF-3DHP. Without bells and whistles, its performance surpasses the previous best result by a large margin of 3% on Human3.6M. Code and models are available at https://github.com/Vegetebird/MHFormer.

As humans, we navigate a multimodal world, building a holistic understanding from all our senses. We introduce MERLOT Reserve, a model that represents videos jointly over time -- through a new training objective that learns from audio, subtitles, and video frames. Given a video, we replace snippets of text and audio with a MASK token; the model learns by choosing the correct masked-out snippet. Our objective learns faster than alternatives, and performs well at scale: we pretrain on 20 million YouTube videos. Empirical results show that MERLOT Reserve learns strong multimodal representations. When finetuned, it sets state-of-the-art on Visual Commonsense Reasoning (VCR), TVQA, and Kinetics-600; outperforming prior work by 5%, 7%, and 1.5% respectively. Ablations show that these tasks benefit from audio pretraining -- even VCR, a QA task centered around images (without sound). Moreover, our objective enables out-of-the-box prediction, revealing strong multimodal commonsense understanding. In a fully zero-shot setting, our model obtains competitive results on four video tasks, even outperforming supervised approaches on the recently proposed Situated Reasoning (STAR) benchmark. We analyze why audio enables better vision-language representations, suggesting significant opportunities for future research. We conclude by discussing ethical and societal implications of multimodal pretraining.

Visual Emotion Analysis (VEA) is attracting increasing attention. One of the biggest challenges of VEA is to bridge the affective gap between visual clues in a picture and the emotion expressed by the picture. As the granularity of emotions increases, the affective gap increases as well. Existing deep approaches try to bridge the gap by directly learning discrimination among emotions globally in one shot without considering the hierarchical relationship among emotions at different affective levels and the affective level of emotions to be classified. In this paper, we present the Multi-level Dependent Attention Network (MDAN) with two branches, to leverage the emotion hierarchy and the correlation between different affective levels and semantic levels. The bottom-up branch directly learns emotions at the highest affective level and strictly follows the emotion hierarchy while predicting emotions at lower affective levels. In contrast, the top-down branch attempt to disentangle the affective gap by one-to-one mapping between semantic levels and affective levels, namely, Affective Semantic Mapping. At each semantic level, a local classifier learns discrimination among emotions at the corresponding affective level. Finally, We integrate global learning and local learning into a unified deep framework and optimize the network simultaneously. Moreover, to properly extract and leverage channel dependencies and spatial attention while disentangling the affective gap, we carefully designed two attention modules: the Multi-head Cross Channel Attention module and the Level-dependent Class Activation Map module. Finally, the proposed deep framework obtains new state-of-the-art performance on six VEA benchmarks, where it outperforms existing state-of-the-art methods by a large margin, e.g., +3.85% on the WEBEmo dataset at 25 classes classification accuracy.