CVPR 2021 论文列表

Pedestrian detection in a crowd is a challenging task due to a high number of mutually-occluding human instances, which brings ambiguity and optimization difficulties to the current IoU-based ground truth assignment procedure in classical object detection methods. In this paper, we develop a unique perspective of pedestrian detection as a variational inference problem. We formulate a novel and efficient algorithm for pedestrian detection by modeling the dense proposals as a latent variable while proposing a customized Auto-Encoding Variational Bayes (AEVB) algorithm. Through the optimization of our proposed algorithm, a classical detector can be fashioned into a variational pedestrian detector. Experiments conducted on CrowdHuman and CityPersons datasets show that the proposed algorithm serves as an efficient solution to handle the dense pedestrian detection problem for the case of single-stage detectors. Our method can also be flexibly applied to two-stage detectors, achieving notable performance enhancement.

Active Learning for discriminative models has largely been studied with the focus on individual samples, with less emphasis on how classes are distributed or which classes are hard to deal with. In this work, we show that this is harmful. We propose a method based on the Bayes' rule, that can naturally incorporate class imbalance into the Active Learning framework. We derive that three terms should be considered together when estimating the probability of a classifier making a mistake for a given sample; i) probability of mislabelling a class, ii) likelihood of the data given a predicted class, and iii) the prior probability on the abundance of a predicted class. Implementing these terms requires a generative model and an intractable likelihood estimation. Therefore, we train a Variational Auto Encoder (VAE) for this purpose. To further tie the VAE with the classifier and facilitate VAE training, we use the classifiers' deep feature representations as input to the VAE. By considering all three probabilities, among them especially the data imbalance, we can substantially improve the potential of existing methods under limited data budget. We show that our method can be applied to classification tasks on multiple different datasets -- including one that is a real-world dataset with heavy data imbalance -- significantly outperforming the state of the art.

In this paper, we present a new dataset with the target of advancing the scene parsing task from images to videos. Our dataset aims to perform Video Scene Parsing in the Wild (VSPW), which covers a wide range of real-world scenarios and categories. To be specific, our VSPW is featured from the following aspects: 1) Well-trimmed long-temporal clips. Each video contains a complete shot, lasting around 5 seconds on average. 2) Dense annotation. The pixel-level annotations are provided at a high frame rate of 15 f/s. 3) High resolution. Over 96% of the captured videos are with high spatial resolutions from 720P to 4K. We totally annotate 3,337 videos, including 239,934 frames from 124 categories. To the best of our knowledge, our VSPW is the first attempt to tackle the challenging video scene parsing task in the wild by considering diverse scenarios. Based on VSPW, we design a generic Temporal Context Blending (TCB) network, which can effectively harness long-range contextual information from the past frames to help segment the current one. Extensive experiments show that our TCB network improves both the segmentation performance and temporal stability comparing with image-/video-based state-of-the-art methods. We hope that the scale, diversity, long-temporal, and high frame rate of our VSPW can significantly advance the research of video scene parsing and beyond.

Visual localization is of great importance in robotics and computer vision. Recently, scene coordinate regression based methods have shown good performance in visual localization in small static scenes. However, it still estimates camera poses from many inferior scene coordinates. To address this problem, we propose a novel visual localization framework that establishes 2D-to-3D correspondences between the query image and the 3D map with a series of learnable scene-specific landmarks. In the landmark generation stage, the 3D surfaces of the target scene are over-segmented into mosaic patches whose centers are regarded as the scene-specific landmarks. To robustly and accurately recover the scene-specific landmarks, we propose the Voting with Segmentation Network (VS-Net) to segment the pixels into different landmark patches with a segmentation branch and estimate the landmark locations within each patch with a landmark location voting branch. Since the number of landmarks in a scene may reach up to 5000, training a segmentation network with such a large number of classes is both computation and memory costly for the commonly used cross-entropy loss. We propose a novel prototype-based triplet loss with hard negative mining, which is able to train semantic segmentation networks with a large number of labels efficiently. Our proposed VS-Net is extensively tested on multiple public benchmarks and can outperform state-of-the-art visual localization methods. Code and models are available at https://github.com/zju3dv/VS-Net.

Accuracy of many visiolinguistic tasks has benefited significantly from the application of vision-and-language (V&L) BERT. However, its application for the task of vision-and-language navigation (VLN) remains limited. One reason for this is the difficulty adapting the BERT architecture to the partially observable Markov decision process present in VLN, requiring history-dependent attention and decision making. In this paper we propose a recurrent BERT model that is time-aware for use in VLN. Specifically, we equip the BERT model with a recurrent function that maintains cross-modal state information for the agent. Through extensive experiments on R2R and REVERIE we demonstrate that our model can replace more complex encoder-decoder models to achieve state-of-the-art results. Moreover, our approach can be generalised to other transformer-based architectures, supports pre-training, and is capable of solving navigation and referring expression tasks simultaneously.

The task of image-based virtual try-on aims to transfer a target clothing item onto the corresponding region of a person, which is commonly tackled by fitting the item to the desired body part and fusing the warped item with the person. While an increasing number of studies have been conducted, the resolution of synthesized images is still limited to low (e.g., 256x192), which acts as the critical limitation against satisfying online consumers. We argue that the limitation stems from several challenges: as the resolution increases, the artifacts in the misaligned areas between the warped clothes and the desired clothing regions become noticeable in the final results; the architectures used in existing methods have low performance in generating high-quality body parts and maintaining the texture sharpness of the clothes. To address the challenges, we propose a novel virtual try-on method called VITON-HD that successfully synthesizes 1024x768 virtual try-on images. Specifically, we first prepare the segmentation map to guide our virtual try-on synthesis, and then roughly fit the target clothing item to a given person's body. Next, we propose ALIgnment-Aware Segment (ALIAS) normalization and ALIAS generator to handle the misaligned areas and preserve the details of 1024x768 inputs. Through rigorous comparison with existing methods, we demonstrate that VITON-HD highly surpasses the baselines in terms of synthesized image quality both qualitatively and quantitatively.

In this paper, we present ViP-DeepLab, a unified model attempting to tackle the long-standing and challenging inverse projection problem in vision, which we model as restoring the point clouds from perspective image sequences while providing each point with instance-level semantic interpretations. Solving this problem requires the vision models to predict the spatial location, semantic class, and temporally consistent instance label for each 3D point. ViP-DeepLab approaches it by jointly performing monocular depth estimation and video panoptic segmentation. We name this joint task as Depth-aware Video Panoptic Segmentation, and propose a new evaluation metric along with two derived datasets for it, which will be made available to the public. On the individual sub-tasks, ViP-DeepLab also achieves state-of-the-art results, outperforming previous methods by 5.1% VPQ on Cityscapes-VPS, ranking 1st on the KITTI monocular depth estimation benchmark, and 1st on KITTI MOTS pedestrian. The datasets and the evaluation codes are made publicly available.

Cross-view image geo-localization aims to determine the locations of street-view query images by matching with GPS-tagged reference images from aerial view. Recent works have achieved surprisingly high retrieval accuracy on city-scale datasets. However, these results rely on the assumption that there exists a reference image exactly centered at the location of any query image, which is not applicable for practical scenarios. In this paper, we redefine this problem with a more realistic assumption that the query image can be arbitrary in the area of interest and the reference images are captured before the queries emerge. This assumption breaks the one-to-one retrieval setting of existing datasets as the queries and reference images are not perfectly aligned pairs, and there may be multiple reference images covering one query location. To bridge the gap between this realistic setting and existing datasets, we propose a new large-scale benchmark --VIGOR-- for cross-View Image Geo-localization beyond One-to-one Retrieval. We benchmark existing state-of-the-art methods and propose a novel end-to-end framework to localize the query in a coarse-to-fine manner. Apart from the image-level retrieval accuracy, we also evaluate the localization accuracy in terms of the actual distance (meters) using the raw GPS data. Extensive experiments are conducted under different application scenarios to validate the effectiveness of the proposed method. The results indicate that cross-view geo-localization in this realistic setting is still challenging, fostering new research in this direction. Our dataset and code will be publicly available.

Disentangled representations support a range of downstream tasks including causal reasoning, generative modeling, and fair machine learning. Unfortunately, disentanglement has been shown to be impossible without the incorporation of supervision or inductive bias. Given that supervision is often expensive or infeasible to acquire, we choose to incorporate structural inductive bias and present an unsupervised, deep State-Space-Model for Video Disentanglement (VDSM). The model disentangles latent time-varying and dynamic factors via the incorporation of hierarchical structure with a dynamic prior and a Mixture of Experts decoder. VDSM learns separate disentangled representations for the identity of the object or person in the video, and for the action being performed. We evaluate VDSM across a range of qualitative and quantitative tasks including identity and dynamics transfer, sequence generation, Frechet Inception Distance, and factor classification. VDSM achieves state-of-the-art performance and exceeds adversarial methods, even when the methods use additional supervision.

It is widely accepted that reasoning about object shape is important for object recognition. However, the most powerful object recognition methods today do not explicitly make use of object shape during learning. In this work, motivated by recent developments in low-shot learning, findings in developmental psychology, and the increased use of synthetic data in computer vision research, we investigate how reasoning about 3D shape can be used to improve low-shot learning methods' generalization performance. We propose a new way to improve existing low-shot learning approaches by learning a discriminative embedding space using 3D object shape, and using this embedding by learning how to map images into it. Our new approach improves the performance of image-only low-shot learning approaches on multiple datasets. We also introduce Toys4K, a 3D object dataset with the largest number of object categories currently available, which supports low-shot learning.

Flat filling is a critical step in digital artistic content creation with the objective of filling line arts with flat colors. We present a deep learning framework for user-guided line art flat filling that can compute the "influence areas" of the user color scribbles, i.e., the areas where the user scribbles should propagate and influence. This framework explicitly controls such scribble influence areas for artists to manipulate the colors of image details and avoid color leakage/contamination between scribbles, and simultaneously, leverages data-driven color generation to facilitate content creation. This framework is based on a Split Filling Mechanism (SFM), which first splits the user scribbles into individual groups and then independently processes the colors and influence areas of each group with a Convolutional Neural Network (CNN). Learned from more than a million illustrations, the framework can estimate the scribble influence areas in a content-aware manner, and can smartly generate visually pleasing colors to assist the daily works of artists. We show that our proposed framework is easy to use, allowing even amateurs to obtain professional-quality results on a wide variety of line arts.

Weakly supervised object localization (WSOL) remains an open problem due to the deficiency of finding object extent information using a classification network. While prior works struggle to localize objects by various spatial regularization strategies, we argue that how to extract object structural information from the trained classification network is neglected. In this paper, we propose a two-stage approach, termed structure-preserving activation (SPA), towards fully leveraging the structure information incorporated in convolutional features for WSOL. In the first stage, a restricted activation module (RAM) is designed to alleviate the structure-missing issue caused by the classification network, based on the observation that the unbounded classification map and global average pooling layer drive the network to focus only on object parts. In the second stage, we propose a post-process approach, termed the self-correlation map generating (SCG) module to obtain structure-preserving localization maps on the basis of the activation maps acquired from the first stage. Specifically, we utilize the high-order self-correlation (HSC) to extract the inherent structural information retained in the learned model and then aggregate HSC of multiple points for precise object localization. Extensive experiments on two publicly available benchmarks including CUB-200-2011 and ILSVRC show that the proposed SPA achieves substantial and consistent performance gains compared with baseline approaches.

Aligning partial views of a scene into a single whole is essential to understanding one's environment and is a key component of numerous robotics tasks such as SLAM and SfM. Recent approaches have proposed end-to-end systems that can outperform traditional methods by leveraging pose supervision. However, with the rising prevalence of cameras with depth sensors, we can expect a new stream of raw RGB-D data without the annotations needed for supervision. We propose UnsupervisedR&R: an end-to-end unsupervised approach to learning point cloud registration from raw RGB-D video. The key idea is to leverage differentiable alignment and rendering to enforce photometric and geometric consistency between frames. We evaluate our approach on indoor scene datasets and find that we outperform existing traditional approaches with classical and learned descriptors while being competitive with supervised geometric point cloud registration approaches.

Inspired by the fact that human eyes continue to develop tracking ability in early and middle childhood, we propose to use tracking as a proxy task for a computer vision system to learn the visual representations. Modelled on the Catch game played by the children, we design a Catch-the-Patch (CtP) game for a 3D-CNN model to learn visual representations that would help with video-related tasks. In the proposed pretraining framework, we cut an image patch from a given video and let it scale and move according to a pre-set trajectory. The proxy task is to estimate the position and size of the image patch in a sequence of video frames, given only the target bounding box in the first frame. We discover that using multiple image patches simultaneously brings clear benefits. We further increase the difficulty of the game by randomly making patches invisible. Extensive experiments on mainstream benchmarks demonstrate the superior performance of CtP against other video pretraining methods. In addition, CtP-pretrained features are less sensitive to domain gaps than those trained by a supervised action recognition task. When both trained on Kinetics-400, we are pleasantly surprised to find that CtP-pretrained representation achieves much higher action classification accuracy than its fully supervised counterpart on Something-Something dataset.

The vision-based reinforcement learning (RL) has achieved tremendous success. However, generalizing vision-based RL policy to unknown test environments still remains as a challenging problem. Unlike previous works that focus on training a universal RL policy that is invariant to discrepancies between test and training environment, we focus on developing an independent module to disperse interference factors irrelevant to the task, thereby providing ""clean"" observations for the RL policy. The proposed unsupervised visual attention and invariance method (VAI) contains three key components: 1) an unsupervised keypoint detection model which captures semantically meaningful keypoints in observations; 2) an unsupervised visual attention module which automatically generates the distraction-invariant attention mask for each observation; 3) a self-supervised adapter for visual distraction invariance which reconstructs distraction-invariant attention mask from observations with artificial disturbances generated by a series of foreground and background augmentations. All components are optimized in an unsupervised way, without manual annotation or access to environment internals, and only the adapter is used during inference time to provide distraction-free observations to RL policy. VAI empirically shows powerful generalization capabilities and significantly outperforms current state-of-the-art (SOTA) method by 15% 49% in DeepMind Control suite benchmark and 61% 229% in our proposed robot manipulation benchmark, in term of cumulative rewards per episode.

These days, unsupervised super-resolution (SR) is soaring due to its practical and promising potential in real scenarios. The philosophy of off-the-shelf approaches lies in the augmentation of unpaired data, i.e. first generating synthetic low-resolution (LR) images Y^g corresponding to real-world high-resolution (HR) images X^r in the real-world LR domain Y^r, and then utilizing the pseudo pairs Y^g, X^r for training in a supervised manner. Unfortunately, since image translation itself is an extremely challenging task, the SR performance of these approaches is severely limited by the domain gap between generated synthetic LR images and real LR images. In this paper, we propose a novel domain-distance aware super-resolution (DASR) approach for unsupervised real-world image SR. The domain gap between training data (e.g. Y^g) and testing data (e.g. Y^r) is addressed with our domain-gap aware training and domain-distance weighted supervision strategies. Domain-gap aware training takes additional benefit from real data in the target domain while domain-distance weighted supervision brings forward the more rational use of labeled source domain data. The proposed method is validated on synthetic and real datasets and the experimental results show that DASR consistently outperforms state-of-the-art unsupervised SR approaches in generating SR outputs with more realistic and natural textures. Codes are available at https://github.com/ShuhangGu/DASR.

In this paper, we present a large scale unlabeled person re-identification (Re-ID) dataset "LUPerson" and make the first attempt of performing unsupervised pre-training for improving the generalization ability of the learned person Re-ID feature representation. This is to address the problem that all existing person Re-ID datasets are all of limited scale due to the costly effort required for data annotation. Previous research tries to leverage models pre-trained on ImageNet to mitigate the shortage of person Re-ID data but suffers from the large domain gap between ImageNet and person Re-ID data. LUPerson is an unlabeled dataset of 4M images of over 200K identities, which is 30xlarger than the largest existing Re-ID dataset. It also covers a much diverse range of capturing environments (e.g., camera settings, scenes, etc.). Based on this dataset, we systematically study the key factors for learning Re-ID features from two perspectives: data augmentation and contrastive loss. Unsupervised pre-training performed on this large-scale dataset effectively leads to a generic Re-ID feature that can benefit all existing person Re-ID methods. Using our pre-trained model in some basic frameworks, our methods achieve state-of-the-art results without bells and whistles on four widely used Re-ID datasets: CUHK03, Market1501, DukeMTMC, and MSMT17. Our results also show that the performance improvement is more significant on small-scale target datasets or under few-shot setting.

We study the problem of unsupervised discovery and segmentation of object parts, which, as an intermediate local representation, are capable of finding intrinsic object structure and providing more explainable recognition results. Recent unsupervised methods have greatly relaxed the dependency on annotated data which are costly to obtain, but still rely on additional information such as object segmentation masks or saliency map. To remove such a dependency and further improve the part segmentation performance, we develop a novel approach by disentangling the appearance and shape representations of object parts followed with reconstruction losses without using additional object mask information. To avoid degenerated solutions, a bottleneck block is designed to squeeze and expand the appearance representation, leading to a more effective disentanglement between geometry and appearance. Combined with a self-supervised part classification loss and an improved geometry concentration constraint, we can segment more consistent parts with semantic meanings. Comprehensive experiments on a wide variety of objects such as face, bird, and PASCAL VOC objects demonstrate the effectiveness of the proposed method.

Despite the importance of unsupervised object detection, to the best of our knowledge, there is no previous work addressing this problem. One main issue, widely known to the community, is that object boundaries derived only from 2D image appearance are ambiguous and unreliable. To address this, we exploit LiDAR clues to aid unsupervised object detection. By exploiting the 3D scene structure, the issue of localization can be considerably mitigated. We further identify another major issue, seldom noticed by the community, that the long-tailed and open-ended (sub-)category distribution should be accommodated. In this paper, we present the first practical method for unsupervised object detection with the aid of LiDAR clues. In our approach, candidate object segments based on 3D point clouds are firstly generated. Then, an iterative segment labeling process is conducted to assign segment labels and to train a segment labeling network, which is based on features from both 2D images and 3D point clouds. The labeling process is carefully designed so as to mitigate the issue of long-tailed and open-ended distribution. The final segment labels are set as pseudo annotations for object detection network training. Extensive experiments on the large-scale Waymo Open dataset suggest that the derived unsupervised object detection method achieves reasonable accuracy compared with that of strong supervision within the LiDAR visible range.