ICCV 2021 论文列表

Active learning (AL) is successful based on the assumption that labeled and unlabeled data are obtained from the same class distribution. However, its performance deteriorates under class distribution mismatch, wherein the unlabeled data contain many samples out of the class distribution of labeled data. To effectively handle the problems under class distribution mismatch, we propose a contrastive coding based AL framework named CCAL. Unlike the existing AL methods that focus on selecting the most informative samples for annotating, CCAL extracts both semantic and distinctive features by contrastive learning and combines them in a query strategy to choose the most informative unlabeled samples with matched categories. Theoretically, we prove that the AL error of CCAL has a tight upper bound. Experimentally, we evaluate its performance on CIFAR10, CIFAR100, and an artificial cross-dataset that consists of five datasets; consequently, CCAL achieves state-of-the-art performance by a large margin with remarkably lower annotation cost. To the best of our knowledge, CCAL is the first work related to AL for class distribution mismatch.

The goal of unsupervised co-localization is to locate the object in a scene under the assumptions that 1) the dataset consists of only one superclass, e.g., birds, and 2) there are no human-annotated labels in the dataset. The most recent method achieves impressive co-localization performance by employing self-supervised representation learning approaches such as predicting rotation. In this paper, we introduce a new contrastive objective directly on the attention maps to enhance co-localization performance. Our contrastive loss function exploits rich information of location, which induces the model to activate the extent of the object effectively. In addition, we propose a pixel-wise attention pooling that selectively aggregates the feature map regarding their magnitudes across channels. Our methods are simple and shown effective by extensive qualitative and quantitative evaluation, achieving state-of-the-art co-localization performances by large margins on four datasets: CUB-200-2011, Stanford Cars, FGVC-Aircraft, and Stanford Dogs. Our code will be publicly available online for the research community.

In the past few years, we have witnessed remarkable breakthroughs in self-supervised representation learning. Despite the success and adoption of representations learned through this paradigm, much is yet to be understood about how different training methods and datasets influence performance on downstream tasks. In this paper, we analyze contrastive approaches as one of the most successful and popular variants of self-supervised representation learning. We perform this analysis from the perspective of the training algorithms, pre-training datasets and end tasks. We examine over 700 training experiments including 30 encoders, 4 pre-training datasets and 20 diverse downstream tasks. Our experiments address various questions regarding the performance of self-supervised models compared to their supervised counterparts, current benchmarks used for evaluation, and the effect of the pre-training data on end task performance. We hope the insights and empirical evidence provided by this work will help future research in learning better visual representations.

This paper studies the problem of learning self-supervised representations on videos. In contrast to image modality that only requires appearance information on objects or scenes, video needs to further explore the relations between multiple frames/clips along the temporal dimension. However, the recent proposed contrastive-based self-supervised frameworks do not grasp such relations explicitly since they simply utilize two augmented clips from the same video and compare their distance without referring to their temporal relation. To address this, we present a contrast-and-order representation (CORP) framework for learning self-supervised video representations that can automatically capture both the appearance information within each frame and temporal information across different frames. In particular, given two video clips, our model first predicts whether they come from the same input video, and then predict the temporal ordering of the clips if they come from the same video. We also propose a novel decoupling attention method to learn symmetric similarity (contrast) and anti-symmetric patterns (order). Such design involves neither extra parameters nor computation, but can speed up the learning process and improve accuracy compared to the vanilla multi-head attention. We extensively validate the representation ability of our learned video features for the downstream action recognition task on Kinetics-400 and Something-something V2. Our method outperforms previous state-of-the-arts by a significant margin.

Recent Visual Question Answering (VQA) models have shown impressive performance on the VQA benchmark but remain sensitive to small linguistic variations in input questions. Existing approaches address this by augmenting the dataset with question paraphrases from visual question generation models or adversarial perturbations. These approaches use the combined data to learn an answer classifier by minimizing the standard cross-entropy loss. To more effectively leverage augmented data, we build on the recent success in contrastive learning. We propose a novel training paradigm (ConClaT) that optimizes both cross-entropy and contrastive losses. The contrastive loss encourages representations to be robust to linguistic variations in questions while the cross-entropy loss preserves the discriminative power of representations for answer prediction. We find that optimizing both losses -- either alternately or jointly -- is key to effective training. On the VQA-Rephrasings benchmark, which measures the VQA model's answer consistency across human paraphrases of a question, ConClaT improves Consensus Score by 1.63% over an improved baseline. In addition, on the standard VQA 2.0 benchmark, we improve the VQA accuracy by 0.78% overall. We also show that ConClaT is agnostic to the type of data-augmentation strategy used.

Current one-step multi-object tracking and segmentation (MOTS) methods lag behind recent two-step methods. By separating the instance segmentation stage from the tracking stage, two-step methods can exploit non-video datasets as extra data for training instance segmentation. Moreover, instances belonging to different IDs on different frames, rather than limited numbers of instances in raw consecutive frames, can be gathered to allow more effective hard example mining in the training of trackers. In this paper, we bridge this gap by presenting a novel data augmentation strategy named continuous copy-paste (CCP). Our intuition behind CCP is to fully exploit the pixel-wise annotations provided by MOTS to actively increase the number of instances as well as unique instance IDs in training. Without any modifications to frameworks, current MOTS methods achieve significant performance gains when trained with CCP. Based on CCP, we propose the first effective one-stage online MOTS method named CCPNet, which generates instance masks as well as the tracking results in one shot. Our CCPNet surpasses all state-of-the-art methods by large margins (3.8% higher sMOTSA and 4.1% higher MOTSA for pedestrians on the KITTI MOTS Validation) and ranks 1st on the KITTI MOTS leaderboard. Evaluations across three datasets also demonstrate the effectiveness of both CCP and CCPNet. Our codes are publicly available at: https://github.com/detectRecog/CCP.

Attaining prototypical features to represent class distributions is well established in representation learning. However, learning prototypes online from streaming data proves a challenging endeavor as they rapidly become outdated, caused by an ever-changing parameter space during the learning process. Additionally, continual learning does not assume the data stream to be stationary, typically resulting in catastrophic forgetting of previous knowledge. As a first, we introduce a system addressing both problems, where prototypes evolve continually in a shared latent space, enabling learning and prediction at any point in time. To facilitate learning, a novel objective function synchronizes the latent space with the continually evolving prototypes. In contrast to the major body of work in continual learning, data streams are processed in an online fashion without task information and can be highly imbalanced, for which we propose an efficient memory scheme. As an additional contribution, we propose the learner-evaluator framework that i) generalizes existing paradigms in continual learning, ii) introduces data incremental learning, and iii) models the bridge between continual learning and concept drift. We obtain state-of-the-art performance by a significant margin on eight benchmarks, including three highly imbalanced data streams. Code is publicly available.

Recent advances have enabled a single neural network to serve as an implicit scene representation, establishing the mapping function between spatial coordinates and scene properties. In this paper, we make a further step towards continual learning of the implicit scene representation directly from sequential observations, namely Continual Neural Mapping. The proposed problem setting bridges the gap between batch-trained implicit neural representations and commonly used streaming data in robotics and vision communities. We introduce an experience replay approach to tackle an exemplary task of continual neural mapping: approximating a continuous signed distance function (SDF) from sequential depth images as a scene geometry representation. We show for the first time that a single network can represent scene geometry over time continually without catastrophic forgetting, while achieving promising trade-offs between accuracy and efficiency.

Continually learning in the real world must overcome many challenges, among which noisy labels are a common and inevitable issue. In this work, we present a replay-based continual learning framework that simultaneously addresses both catastrophic forgetting and noisy labels for the first time. Our solution is based on two observations; (i) forgetting can be mitigated even with noisy labels via self-supervised learning, and (ii) the purity of the replay buffer is crucial. Building on this regard, we propose two key components of our method: (i) a self-supervised replay technique named Self-Replay, which can circumvent erroneous training signals arising from noisy labeled data, and (ii) the Self-Centered filter that maintains a purified replay buffer via centrality-based stochastic graph ensembles. The empirical results on MNIST, CIFAR-10, CIFAR-100, and WebVision with real-world noise demonstrate that our framework can maintain a highly pure replay buffer amidst noisy streamed data while greatly outperforming the combinations of the state-of-the-art continual learning and noisy label learning methods.

For several emerging technologies such as augmented reality, autonomous driving and robotics, visual localization is a critical component. Directly regressing camera pose/3D scene coordinates from the input image using deep neural networks has shown great potential. However, such methods assume a stationary data distribution with all scenes simultaneously available during training. In this paper, we approach the problem of visual localization in a continual learning setup -- whereby the model is trained on scenes in an incremental manner. Our results show that similar to the classification domain, non-stationary data induces catastrophic forgetting in deep networks for visual localization. To address this issue, a strong baseline based on storing and replaying images from a fixed buffer is proposed. Furthermore, we propose a new sampling method based on coverage score (Buff-CS) that adapts the existing sampling strategies in the buffering process to the problem of visual localization. Results demonstrate consistent improvements over standard buffering methods on two challenging datasets -- 7Scenes, 12Scenes, and also 19Scenes by combining the former scenes.

We tackle the task of diverse 3D human motion prediction, that is, forecasting multiple plausible future 3D poses given a sequence of observed 3D poses. In this context, a popular approach consists of using a Conditional Variational Autoencoder (CVAE). However, existing approaches that do so either fail to capture the diversity in human motion, or generate diverse but semantically implausible continuations of the observed motion. In this paper, we address both of these problems by developing a new variational framework that accounts for both diversity and context of the generated future motion. To this end, and in contrast to existing approaches, we condition the sampling of the latent variable that acts as source of diversity on the representation of the past observation, thus encouraging it to carry relevant information. Our experiments demonstrate that our approach yields motions not only of higher quality while retaining diversity, but also that preserve the contextual information contained in the observed motion.

Although gait recognition has drawn increasing research attention recently, it remains challenging to learn discriminative temporal representation since the silhouette differences are quite subtle in spatial domain. Inspired by the observation that humans can distinguish gaits of different subjects by adaptively focusing on temporal sequences with different time scales, we propose a context-sensitive temporal feature learning (CSTL) network in this paper, which aggregates temporal features in three scales to obtain motion representation according to the temporal contextual information. Specifically, CSTL introduces relation modeling among multi-scale features to evaluate feature importances, based on which network adaptively enhances more important scale and suppresses less important scale. Besides that, we propose a salient spatial feature learning (SSFL) module to tackle the misalignment problem caused by temporal operation, e.g., temporal convolution. SSFL recombines a frame of salient spatial features by extracting the most discriminative parts across the whole sequence. In this way, we achieve adaptive temporal learning and salient spatial mining simultaneously. Extensive experiments conducted on two datasets demonstrate the state-of-the-art performance. On CASIA-B dataset, we achieve rank-1 accuracies of 98.0%, 95.4% and 87.0% under normal walking, bag-carrying and coat-wearing conditions. On OU-MVLP dataset, we achieve rank-1 accuracy of 90.2%. The source code will be published at https://github.com/OliverHxh/CSTL.

Scene graph generation is an important task in computer vision aimed at improving the semantic understand- ing of the visual world. In this task, the model needs to detect objects and predict visual relationships between them. Most of the existing models predict relationships in parallel assuming their independence. While there are differ- ent ways to capture these dependencies, we explore a conditional approach motivated by the sequence-to-sequence (Seq2Seq) formalism. Different from the previous research, our proposed model predicts visual relationships one at a time in an autoregressive manner by explicitly conditioning on the already predicted relationships. Drawing from translation models in NLP, we propose an encoder- decoder model built using Transformers where the encoder captures global context and long range interactions. The decoder then makes sequential predictions by conditioning on the scene graph constructed so far. In addition, we introduce a novel reinforcement learning-based training strategy tailored to Seq2Seq scene graph generation. By using a self-critical policy gradient training approach with Monte Carlo search we directly optimize for the (mean) recall metrics and bridge the gap between training and evaluation. Experimental results on two public benchmark datasets demonstrate that our Seq2Seq learning approach achieves strong empirical performance, out- performing previous state-of-the-art, while remaining efficient in terms of training and inference time. Full code for this work is available here: https://github.com/ layer6ai-labs/SGG-Seq2Seq.

Deep convolutional neural networks (CNNs) are achieving great successes for image super-resolution (SR), where global context is crucial for accurate restoration. However, the basic convolutional layer in CNNs is designed to extract local patterns, lacking the ability to model global context. Many efforts have been devoted to augmenting SR networks with the global context information, especially by global feature interaction methods. These works incorporate the global context into local feature representation. However, recent advances in neuroscience show that it is necessary for the neurons to dynamically modulate their functions according to context, which is neglected in most CNN based SR methods. Motivated by those observations and analyses, we propose context reasoning attention network (CRAN) to adaptively modulate the convolution kernel according to the global context. Specifically, we extract global context descriptors, which are further enhanced with semantic reasoning. Channel and spatial interactions are then proposed to generate context reasoning attention mask, which is applied to modify the convolution kernel adaptively. Such a modulated convolution layer is utilized as basic component to build the network blocks and itself. Extensive experiments on benchmark datasets with multiple degradation models show that our CRAN achieves superior SR results and favourable efficiency trade-off.

Data augmentation is vital for deep learning neural networks. By providing massive training samples, it helps to improve the generalization ability of the model. Weakly supervised semantic segmentation (WSSS) is a challenging problem that has been deeply studied in recent years, conventional data augmentation approaches for WSSS usually employ geometrical transformations, random cropping, and color jittering. However, merely increasing the same contextual semantic data does not bring much gain to the networks to distinguish the objects, e.g., the correct image-level classification of "aeroplane" may be not only due to the recognition of the object itself but also its co-occurrence context like "sky", which will cause the model to focus less on the object features. To this end, we present a Context Decoupling Augmentation (CDA) method, to change the inherent context in which the objects appear and thus drive the network to remove the dependence between object instances and contextual information. To validate the effectiveness of the proposed method, extensive experiments on PASCAL VOC 2012 dataset with several alternative network architectures demonstrate that CDA can boost various popular WSSS methods to the new state-of-the-art by a large margin.

This paper introduces a motion retargeting method that preserves self-contacts and prevents inter-penetration. Self-contacts, such as when hands touch each other or the torso or the head, are important attributes of human body language and dynamics, yet existing methods do not model or preserve these contacts. Likewise, self-penetrations, such as a hand passing into the torso, are a typical artifact of motion estimation methods. The input to our method is a human motion sequence and a target skeleton and character geometry. The method identifies self-contacts and ground contacts in the input motion, and optimizes the motion to apply to the output skeleton, while preserving these contacts and reducing self-penetrations. We introduce a novel geometry-conditioned recurrent network with an encoder-space optimization strategy that achieves efficient retargeting while satisfying contact constraints. In experiments, our results quantitatively outperform previous methods and in the user study our retargeted motions are rated as higher-quality than those produced by recent works. We also show our method generalizes to motion estimated from human videos where we improve over previous works that produce noticeable interpenetration.

Black-box attacks aim to generate adversarial noise tofail the victim deep neural network in the black box. Thecentral task in black-box attack method design is to estimateand characterize the victim model in the high-dimensionalmodel space based on feedback results of queries submittedto the victim network. The central performance goal is tominimize the number of queries needed for successful at-tack. Existing attack methods directly search and refine theadversarial noise in an extremely high-dimensional space,requiring hundreds or even thousands queries to the victimnetwork. To address this challenge, we propose to explore aconsistency and sensitivity guided ensemble attack (CSEA)method in a low-dimensional space. Specifically, we esti-mate the victim model in the black box using a learned lin-ear composition of an ensemble of surrogate models withdiversified network structures. Using random block maskson the input image, these surrogate models jointly constructand submit randomized and sparsified queries to the victimmodel. Based on these query results and guided by a con-sistency constraint, the surrogate models can be trained us-ing a very small number of queries such that their learnedcomposition is able to accurately approximate the victimmodel in the high-dimensional space. The randomized andsparsified queries also provide important information for usto construct an attack sensitivity map for the input image,with which the adversarial attack can be locally refined tofurther increase its success rate. Our extensive experimen-tal results demonstrate that our proposed approach signifi-cantly reduces the number of queries to the victim networkwhile maintaining very high success rates, outperformingexisting black-box attack methods by large margins.

Compared with the progress made on human activity classification, much less success has been achieved on human interaction understanding (HIU). Apart from the latter task is much more challenging, the main cause is that recent approaches learn human interactive relations via shallow graphical models, which is inadequate to model complicated human interactions. In this paper, we propose a consistency-aware graph network, which combines the representative ability of graph network and the consistency-aware reasoning to facilitate the HIU task. Our network consists of three components, a backbone CNN to extract image features, a factor graph network to learn third-order interactive relations among participants, and a consistency-aware reasoning module to enforce labeling and grouping consistencies. Our key observation is that the consistency-aware-reasoning bias for HIU can be embedded into an energy function, minimizing which delivers consistent predictions. An efficient mean-field inference algorithm is proposed, such that all modules of our network could be trained jointly in an end-to-end manner. Experimental results show that our approach achieves leading performance on three benchmarks. Code will be publicly available.

Within Convolutional Neural Network (CNN), the convolution operations are good at extracting local features but experience difficulty to capture global representations. Within visual transformer, the cascaded self-attention modules can capture long-distance feature dependencies but unfortunately deteriorate local feature details. In this paper, we propose a hybrid network structure, termed Conformer, to take advantage of convolutional operations and self-attention mechanisms for enhanced representation learning. Conformer roots in the Feature Coupling Unit (FCU), which fuses local features and global representations under different resolutions in an interactive fashion. Conformer adopts a concurrent structure so that local features and global representations are retained to the maximum extent. Experiments show that Conformer, under the comparable parameter complexity, outperforms the visual transformer (DeiT-B) by 2.3% on ImageNet. On MSCOCO, it outperforms ResNet-101 by 3.7% and 3.6% mAPs for object detection and instance segmentation, respectively, demonstrating the great potential to be a general backbone network. Code is available at github.com/pengzhiliang/Conformer.