ICCV 2021 论文列表

Transformers with powerful global relation modeling abilities have been introduced to fundamental computer vision tasks recently. As a typical example, the Vision Transformer (ViT) directly applies a pure transformer architecture on image classification, by simply splitting images into tokens with a fixed length, and employing transformers to learn relations between these tokens. However, such naive tokenization could destruct object structures, assign grids to uninterested regions such as background, and introduce interference signals. To mitigate the above issues, in this paper, we propose an iterative and progressive sampling strategy to locate discriminative regions. At each iteration, embeddings of the current sampling step are fed into a transformer encoder layer, and a group of sampling offsets is predicted to update the sampling locations for the next step. The progressive sampling is differentiable. When combined with the Vision Transformer, the obtained PS-ViT network can adaptively learn where to look. The proposed PS-ViT is both effective and efficient. When trained from scratch on ImageNet, PS-ViT performs 3.8% higher than the vanilla ViT in terms of top-1 accuracy with about 4x fewer parameters and 10x fewer FLOPs. Code is available at https://github.com/yuexy/PS-ViT.

We address the task of Re-Identification (Re-ID) in multi-target multi-camera (MTMC) tracking where we track multiple pedestrians using multiple overlapping uncalibrated (unknown pose) cameras. Since the videos are temporally synchronized and spatially overlapping, we can see a person from multiple views and associate their trajectory across cameras. In order to find the correct association between pedestrians visible from multiple views during the same time window, we extract a visual feature from a tracklet (sequence of pedestrian images) that encodes its similarity and dissimilarity to all other candidate tracklets. We propose a inter-tracklet (person to person) attention mechanism that learns a representation for a target tracklet while taking into account other tracklets across multiple views. Furthermore, to encode the gait and motion of a person, we introduce second intra-tracklet (person-specific) attention module with position embeddings. This second module employs a transformer encoder to learn a feature from a sequence of features over one tracklet. Experimental results on WILDTRACK and our new dataset `ConstructSite' confirm the superiority of our model over state-of-the-art ReID methods (5% and 10% performance gain respectively) in the context of uncalibrated MTMC tracking. While our model is designed for overlapping cameras, we also obtain state-of-the-art results on two other benchmark datasets (MARS and DukeMTMC) with non-overlapping cameras.

The past year has witnessed the rapid development of applying the Transformer module to vision problems. While some researchers have demonstrated that Transformer-based models enjoy a favorable ability of fitting data, there are still growing number of evidences showing that these models suffer over-fitting especially when the training data is limited. This paper offers an empirical study by performing step-by-step operations to gradually transit a Transformer-based model to a convolution-based model. The results we obtain during the transition process deliver useful messages for improving visual recognition. Based on these observations, we propose a new architecture named Visformer, which is abbreviated from the 'Vision-friendly Transformer'. With the same computational complexity, Visformer outperforms both the Transformer-based and convolution-based models in terms of ImageNet classification accuracy, and the advantage becomes more significant when the model complexity is lower or the training set is smaller. The code is available at https://github.com/danczs/Visformer.

We present a novel framework for mesh reconstruction from unstructured point clouds by taking advantage of the learned visibility of the 3D points in the virtual views and traditional graph-cut based mesh generation. Specifically, we first propose a three-step network that explicitly employs depth completion for visibility prediction. Then the visibility information of multiple views is aggregated to generate a 3D mesh model by solving an optimization problem considering visibility in which a novel adaptive visibility weighting term in surface determination is also introduced to suppress line of sight with a large incident angle. Compared to other learning-based approaches, our pipeline only exercises the learning on a 2D binary classification task, i.e., points visible or not in a view, which is much more generalizable and practically more efficient and capable to deal with a large number of points. Experiments demonstrate that our method with favorable transferability and robustness, and achieve competing performances w.r.t. state-of-the-art learning-based approaches on small complex objects and outperforms on large indoor and outdoor scenes. Code is available at https://github.com/GDAOSU/vis2mesh.

Most existing human matting algorithms tried to separate pure human-only foreground from the background. In this paper, we propose a Virtual Multi-modality Foreground Matting (VMFM) method to learn human-object interactive foreground (human and objects interacted with him or her) from a raw RGB image. The VMFM method requires no additional inputs, e.g. trimap or known background. We reformulate foreground matting as a self-supervised multi-modality problem: factor each input image into estimated depth map, segmentation mask, and interaction heatmap using three auto-encoders. In order to fully utilize the characteristics of each modality, we first train a dual encoder-to-decoder network to estimate the same alpha matte. Then we introduce a self-supervised method: Complementary Learning(CL) to predict deviation probability map and exchange reliable gradients across modalities without label. We conducted extensive experiments to analyze the effectiveness of each modality and the significance of different components in complementary learning. We demonstrate that our model outperforms the state-of-the-art methods.

The light transport matrix (LTM) is an instrumental tool in line-of-sight (LOS) imaging, describing how light interacts with the scene and enabling applications such as relighting or separation of illumination components. We introduce a framework to estimate the LTM of non-line-of-sight (NLOS) scenarios, coupling recent virtual forward light propagation models for NLOS imaging with the LOS light transport equation. We design computational projector-camera setups, and use these virtual imaging systems to estimate the transport matrix of hidden scenes. We introduce the specific illumination functions to compute the different elements of the matrix, overcoming the challenging wide-aperture conditions of NLOS setups. Our NLOS light transport matrix allows us to (re)illuminate specific locations of a hidden scene, and separate direct, first-order indirect, and higher-order indirect illumination of complex cluttered hidden scenes, similar to existing LOS techniques.

Change captioning is the task of identifying the change and describing it with a concise caption. Despite recent advancements, filtering out insignificant changes still remains as a challenge. Namely, images from different camera perspectives can cause issues; a mere change in viewpoint should be disregarded while still capturing the actual changes. In order to tackle this problem, we present a new Viewpoint-Agnostic change captioning network with Cycle Consistency (VACC) that requires only one image each for the before and after scene, without depending on any other information. We achieve this by devising a new difference encoder module which can encode viewpoint information and model the difference more effectively. In addition, we propose a cycle consistency module that can potentially improve the performance of any change captioning networks in general by matching the composite feature of the generated caption and before image with the after image feature. We evaluate the performance of our proposed model across three datasets for change captioning, including a novel dataset we introduce here that contains images with changes under extreme viewpoint shifts. Through our experiments, we show the excellence of our method with respect to the CIDEr, BLEU-4, METEOR and SPICE scores. Moreover, we demonstrate that attaching our proposed cycle consistency module yields a performance boost for existing change captioning networks, even with varying image encoding mechanisms.

In this paper we propose a novel method for image matching based on dense local features and tailored for visual geolocalization. Dense local features matching is robust against changes in illumination and occlusions, but not against viewpoint shifts which are a fundamental aspect of geolocalization. Our method, called GeoWarp, directly embeds invariance to viewpoint shifts in the process of extracting dense features. This is achieved via a trainable module which learns from the data an invariance that is meaningful for the task of recognizing places. We also devise a new self-supervised loss and two new weakly supervised losses to train this module using only unlabeled data and weak labels. GeoWarp is implemented efficiently as a re-ranking method that can be easily embedded into pre-existing visual geolocalization pipelines. Experimental validation on standard geolocalization benchmarks demonstrates that GeoWarp boosts the accuracy of state-of-the-art retrieval architectures. The code and trained models will be released upon acceptance of this paper.

In structure-from-motion the viewing graph is a graph where vertices correspond to cameras and edges represent fundamental matrices. We provide a new formulation and an algorithm for establishing whether a viewing graph is solvable, i.e. it uniquely determines a set of projective cameras. Known theoretical conditions either do not fully characterize the solvability of all viewing graphs, or are exceedingly hard to compute for they involve solving a system of polynomial equations with a large number of unknowns. The main result of this paper is a method for reducing the number of unknowns by exploiting the cycle consistency. We advance the understanding of the solvability by (i) finishing the classification of all previously undecided minimal graphs up to 9 nodes, (ii) extending the practical solvability testing up to minimal graphs with up to 90 nodes, and (iii) definitely answering an open research question by showing that the finite solvability is not equivalent to the solvability. Finally, we present an experiment on real data showing that unsolvable graphs are appearing in practical situations.

Understanding the 3D world without supervision is currently a major challenge in computer vision as the annotations required to supervise deep networks for tasks in this domain are expensive to obtain on a large scale. In this paper, we address the problem of unsupervised viewpoint estimation. We formulate this as a self-supervised learning task, where image reconstruction provides the supervision needed to predict the camera viewpoint. Specifically, we make use of pairs of images of the same object at training time, from unknown viewpoints, to self-supervise training by combining the viewpoint information from one image with the appearance information from the other. We demonstrate that using a perspective spatial transformer allows efficient viewpoint learning, outperforming existing unsupervised approaches on synthetic data, and obtains competitive results on the challenging PASCAL3D+ dataset.

Label distributions in real-world are oftentimes long-tailed and imbalanced, resulting in biased models towards dominant labels. While long-tailed recognition has been extensively studied for image classification tasks, limited effort has been made for video domain. In this paper, we introduce VideoLT, a large-scale long-tailed video recognition dataset, as a step toward real-world video recognition. VideoLT contains 256,218 untrimmed videos, annotated into 1,004 classes with a long-tailed distribution. Through extensive studies, we demonstrate that state-of-the-art methods used for long-tailed image recognition do not perform well in the video domain due to the additional temporal dimension in video data. This motivates us to propose FrameStack, a simple yet effective method for long-tailed video recognition task. In particular, FrameStack performs sampling at the frame-level in order to balance class distributions, and the sampling ratio is dynamically determined using knowledge derived from the network during training. Experimental results demonstrate that FrameStack can improve classification performance without sacrificing overall accuracy. Code and dataset are available at: https://github.com/17Skye17/VideoLT.

Video-based person re-identification (reID) aims to retrieve person videos with the same identity as a query person across multiple cameras. Spatial and temporal distractors in person videos, such as background clutter and partial occlusions over frames, respectively, make this task much more challenging than image-based person reID. We observe that spatial distractors appear consistently in a particular location, and temporal distractors show several patterns, e.g., partial occlusions occur in the first few frames, where such patterns provide informative cues for predicting which frames to focus on (i.e., temporal attentions). Based on this, we introduce a novel Spatial and Temporal Memory Networks (STMN). The spatial memory stores features for spatial distractors that frequently emerge across video frames, while the temporal memory saves attentions which are optimized for typical temporal patterns in person videos. We leverage the spatial and temporal memories to refine frame-level person representations and to aggregate the refined frame-level features into a sequence-level person representation, respectively, effectively handling spatial and temporal distractors in person videos. We also introduce a memory spread loss preventing our model from addressing particular items only in the memories. Experimental results on standard benchmarks, including MARS, DukeMTMC-VideoReID, and LS-VID, demonstrate the effectiveness of our method.

Temporal action localization (TAL) in videos is a challenging task, especially due to the large variation in action temporal scales. Short actions usually occupy a major proportion in the datasets, but tend to have the lowest performance. In this paper, we confront the challenge of short actions and propose a multi-level cross-scale solution dubbed as video self-stitching graph network (VSGN). We have two key components in VSGN: video self-stitching (VSS) and cross-scale graph pyramid network (xGPN). In VSS, we focus on a short period of a video and magnify it along the temporal dimension to obtain a larger scale. We stitch the original clip and its magnified counterpart in one input sequence to take advantage of the complementary properties of both scales. The xGPN component further exploits the cross-scale correlations by a pyramid of cross-scale graph networks, each containing a hybrid module to aggregate features from across scales as well as within the same scale. Our VSGN not only enhances the feature representations, but also generates more positive anchors for short actions and more short training samples. Experiments demonstrate that VSGN obviously improves the localization performance of short actions as well as achieving the state-of-the-art overall performance on THUMOS-14 and ActivityNet-v1.3.

Video Question Answering (Video QA) aims to give an answer to the question through semantic reasoning between visual and linguistic information. Recently, handling large amounts of multi-modal video and language information of a video is considered important in the industry. However, the current video QA models use deep features, suffered from significant computational complexity and insufficient representation capability both in training and testing. Existing features are extracted using pre-trained networks after all the frames are decoded, which is not always suitable for video QA tasks. In this paper, we develop a novel deep neural network to provide video QA features obtained from coded video bit-stream to reduce the complexity. The proposed network includes several dedicated deep modules to both the video QA and the video compression system, which is the first attempt at the video QA task. The proposed network is predominantly model-agnostic. It is integrated into the state-of-the-art networks for improved performance without any computationally expensive motion-related deep models. The experimental results demonstrate that the proposed network outperforms the previous studies at lower complexity.

Human pose is a useful feature for fine-grained sports action understanding. However, pose estimators are often unreliable when run on sports video due to domain shift and factors such as motion blur and occlusions. This leads to poor accuracy when downstream tasks, such as action recognition, depend on pose. End-to-end learning circumvents pose, but requires more labels to generalize. We introduce Video Pose Distillation (VPD), a weakly-supervised technique to learn features for new video domains, such as individual sports that challenge pose estimation. Under VPD, a student network learns to extract robust pose features from RGB frames in the sports video, such that, whenever pose is considered reliable, the features match the output of a pretrained teacher pose detector. Our strategy retains the best of both pose and end-to-end worlds, exploiting the rich visual patterns in raw video frames, while learning features that agree with the athletes' pose and motion in the target video domain to avoid over-fitting to patterns unrelated to athletes' motion. VPD features improve performance on few-shot, fine-grained action recognition, retrieval, and detection tasks in four real-world sports video datasets, without requiring additional ground-truth pose annotations.

In this paper, we propose a novel solution for object-matching based semi-supervised video object segmentation, where the target object masks in the first frame are provided. Existing object-matching based methods focus on the matching between the raw object features of the current frame and the first/previous frames. However, two issues are still not solved by these object-matching based methods. As the appearance of the video object changes drastically over time, 1) unseen parts/details of the object present in the current frame, resulting in incomplete annotation in the first annotated frame (e.g., view/scale changes). 2) even for the seen parts/details of the object in the current frame, their positions change relatively (e.g., pose changes/camera motion), leading to a misalignment for the object matching. To obtain the complete information of the target object, we propose a novel object-based dynamic memory network that exploits visual contents of all the past frames. To solve the misalignment problem caused by position changes of visual contents, we propose an adaptive object alignment module by incorporating a region translation function that aligns object proposals towards templates in the feature space. Our method achieves state-of-the-art results on latest benchmark datasets DAVIS 2017 (J of 81.4% and F of 87.5% on the validation set) and YouTube-VOS (the overall score of 82.7% on the validation set) with a very efficient inference time (0.16 second/frame on DAVIS 2017 validation set). Code is available at: https://github.com/liang4sx/DMN-AOA.

Learning temporally consistent foreground opacity from videos, i.e., video matting, has drawn great attention due to the blossoming of video conferencing. Previous approaches are built on top of image matting models, which fail in maintaining the temporal coherence when being adapted to videos. They either utilize the optical flow to smooth frame-wise prediction, where the performance is dependent on the selected optical flow model; or naively combine feature maps from multiple frames, which does not model well the correspondence of pixels in adjacent frames. In this paper, we propose to enhance the temporal coherence by Consistency-Regularized Graph Neural Networks (CRGNN) with the aid of a synthesized video matting dataset. CRGNN utilizes Graph Neural Networks (GNN) to relate adjacent frames such that pixels or regions that are incorrectly predicted in one frame can be corrected by leveraging information from its neighboring frames. To generalize our model from synthesized videos to real-world videos, we propose a consistency regularization technique to enforce the consistency on the alpha and foreground when blending them with different backgrounds. To evaluate the efficacy of CRGNN, we further collect a real-world dataset with annotated alpha mattes. Compared with state-of-the-art methods that require hand-crafted trimaps or backgrounds for modeling training, CRGNN generates favorably results with the help of unlabeled real training dataset.

Video instance segmentation (VIS) aims to segment and associate all instances of predefined classes for each frame in videos. Prior methods usually obtain segmentation for a frame or clip first, and merge the incomplete results by tracking or matching. These methods may cause error accumulation in the merging step. Contrarily, we propose a new paradigm -- Propose-Reduce, to generate complete sequences for input videos by a single step. We further build a sequence propagation head on the existing image-level instance segmentation network for long-term propagation. To ensure robustness and high recall of our proposed framework, multiple sequences are proposed where redundant sequences of the same instance are reduced. We achieve state-of-the-art performance on two representative benchmark datasets -- we obtain 47.6% in terms of AP on YouTube-VIS validation set and 70.4% for J&F on DAVIS-UVOS validation set.

In this paper, we address the problem of video geo-localization by proposing a Geo-Temporal Feature Learning (GTFL) Network to simultaneously learn the discriminative features between the query videos and gallery images for estimating the geo-spatial trajectory of a query video. Based on a transformer encoder architecture, our GTFL model encodes query and gallery data separately, via two dedicated branches. The proposed GPS Loss and Clip Triplet Loss exploit the geographical and temporal proximity between the frames and the clips to jointly learn the query and gallery features. We also propose a deep learning approach to trajectory smoothing by predicting the outliers in the estimated GPS positions and learning the offsets to smooth the trajectory. We build a large dataset from four different regions of USA; New York, San Francisco, Berkeley and Bay Area using BDD driving videos as query, and by collecting corresponding Google StreetView (GSV) Images for gallery. Extensive evaluations of proposed method on this new dataset are provided. Code and dataset details will be made publicly available.