CVPR 2021 论文列表

Knowledge distillation pursues a diminutive yet well-behaved student network by harnessing the knowledge learned by a cumbersome teacher model. Prior methods achieve this by making the student imitate shallow behaviors, such as soft targets, features, or attention, of the teacher. In this paper, we argue that what really matters for distillation is the intrinsic problem-solving process captured by the teacher. By dissecting the decision process in a layer-wise manner, we found that the decision-making procedure in the teacher model is conducted in a coarse-to-fine manner, where coarse-grained discrimination (e.g., animal vs vehicle) is attained in early layers, and fine-grained discrimination (e.g., dog vs cat, car vs truck) in latter layers. Motivated by this observation, we propose a new distillation method, dubbed as Tree-like Decision Distillation (TDD), to endow the student with the same problem-solving mechanism as that of the teacher. Extensive experiments demonstrated that TDD yields competitive performance compared to state of the arts. More importantly, it enjoys better interpretability due to its interpretable decision distillation instead of dark knowledge distillation.

Previous models for vision-to-language generation tasks usually pretrain a visual encoder and a language generator in the respective domains and jointly finetune them with the target task. However, this direct transfer practice may suffer from the discord between visual specificity and language fluency since they are often separately trained from large corpora of visual and text data with no common ground. In this work, we claim that a transitional adaptation task is required between pretraining and finetuning to harmonize the visual encoder and the language model for challenging downstream target tasks like visual storytelling. We propose a novel approach named Transitional Adaptation of Pretrained Model (TAPM) that adapts the multi-modal modules to each other with a simpler alignment task between visual inputs only with no need for text labels. Through extensive experiments, we show that the adaptation step significantly improves the performance of multiple language models for sequential video and image captioning tasks. We achieve new state-of-the-art performance on both language metrics and human evaluation in the multi-sentence description task of LSMDC 2019 and the image storytelling task of VIST. Our experiments reveal that this improvement in caption quality does not depend on the specific choice of language models.

Correlation acts as a critical role in the tracking field, especially in recent popular Siamese-based trackers. The correlation operation is a simple fusion manner to consider the similarity between the template and the search region. However, the correlation operation itself is a local linear matching process, leading to lose semantic information and fall into local optimum easily, which may be the bottleneck of designing high-accuracy tracking algorithms. Is there any better feature fusion method than correlation? To address this issue, inspired by Transformer, this work presents a novel attention-based feature fusion network, which effectively combines the template and search region features solely using attention. Specifically, the proposed method includes an ego-context augment module based on self-attention and a cross-feature augment module based on cross-attention. Finally, we present a Transformer tracking (named TransT) method based on the Siamese-like feature extraction backbone, the designed attention-based fusion mechanism, and the classification and regression head. Experiments show that our TransT achieves very promising results on six challenging datasets, especially on large-scale LaSOT, TrackingNet, and GOT-10k benchmarks. Our tracker runs at approximatively 50 fps on GPU. Code and models are available at https://github.com/chenxin-dlut/TransT.

In video object tracking, there exist rich temporal contexts among successive frames, which have been largely overlooked in existing trackers. In this work, we bridge the individual video frames and explore the temporal contexts across them via a transformer architecture for robust object tracking. Different from classic usage of the transformer in natural language processing tasks, we separate its encoder and decoder into two parallel branches and carefully design them within the Siamese-like tracking pipelines. The transformer encoder promotes the target templates via attention-based feature reinforcement, which benefits the high-quality tracking model generation. The transformer decoder propagates the tracking cues from previous templates to the current frame, which facilitates the object searching process. Our transformer-assisted tracking framework is neat and trained in an end-to-end manner. With the proposed transformer, a simple Siamese matching approach is able to outperform the current top-performing trackers. By combining our transformer with the recent discriminative tracking pipeline, our method sets several new state-of-the-art records on prevalent tracking benchmarks.

Self-attention techniques, and specifically Transformers, are dominating the field of text processing and are becoming increasingly popular in computer vision classification tasks. In order to visualize the parts of the image that led to a certain classification, existing methods either rely on the obtained attention maps or employ heuristic propagation along the attention graph. In this work, we propose a novel way to compute relevancy for Transformer networks. The method assigns local relevance based on the Deep Taylor Decomposition principle and then propagates these relevancy scores through the layers. This propagation involves attention layers and skip connections, which challenge existing methods. Our solution is based on a specific formulation that is shown to maintain the total relevancy across layers. We benchmark our method on very recent visual Transformer networks, as well as on a text classification problem, and demonstrate a clear advantage over the existing explainability methods.

Few-shot object detection (FSOD) aims to learn detectors that can be generalized to novel classes with only a few instances. Unlike previous attempts that exploit meta-learning techniques to facilitate FSOD, this work tackles the problem from the perspective of sample expansion. To this end, we propose a simple yet effective Transformation Invariant Principle (TIP) that can be flexibly applied to various meta-learning models for boosting the detection performance on novel class objects. Specifically, by introducing consistency regularization on predictions from various transformed images, we augment vanilla FSOD models with the generalization ability to objects perturbed by various transformation, such as occlusion and noise. Importantly, our approach can extend supervised FSOD models to naturally cope with unlabeled data, thus addressing a more practical and challenging semi-supervised FSOD problem. Extensive experiments on PASCAL VOC and MSCOCO datasets demonstrate the effectiveness of our TIP under both of the two FSOD settings.

This paper defines a new visual reasoning paradigm by introducing an important factor, i.e. transformation. The motivation comes from the fact that most existing visual reasoning tasks, such as CLEVR in VQA, are solely defined to test how well the machine understands the concepts and relations within static settings, like one image. We argue that this kind of state driven visual reasoning approach has limitations in reflecting whether the machine has the ability to infer the dynamics between different states, which has been shown as important as state-level reasoning for human cognition in Piaget's theory. To tackle this problem, we propose a novel transformation driven visual reasoning task. Given both the initial and final states, the target is to infer the corresponding single-step or multi-step transformation, represented as a triplet (object, attribute, value) or a sequence of triplets, respectively. Following this definition, a new dataset namely TRANCE is constructed on the basis of CLEVR, including three levels of settings, i.e. Basic (single-step transformation), Event (multi-step transformation), and View (multi-step transformation with variant views). Experimental results show that the state-of-the-art visual reasoning models perform well on Basic, but are still far from human-level intelligence on Event and View. We believe the proposed new paradigm will boost the development of machine visual reasoning. More advanced methods and real data need to be investigated in this direction. The resource of TVR is available at https://hongxin2019.github.io/TVR.

Domain adaptation has been widely explored by transferring the knowledge from a label-rich source domain to a related but unlabeled target domain. Most existing domain adaptation algorithms attend to adapting feature representations across two domains with the guidance of a shared source-supervised classifier. However, such classifier limits the generalization ability towards unlabeled target recognition. To remedy this, we propose a Transferable Semantic Augmentation (TSA) approach to enhance the classifier adaptation ability through implicitly generating source features towards target semantics. Specifically, TSA is inspired by the fact that deep feature transformation towards a certain direction can be represented as meaningful semantic altering in the original input space. Thus, source features can be augmented to effectively equip with target semantics to train a more transferable classifier. To achieve this, for each class, we first use the inter-domain feature mean difference and target intra-class feature covariance to construct a multivariate normal distribution. Then we augment source features with random directions sampled from the distribution class-wisely. Interestingly, such source augmentation is implicitly implemented through an expected transferable cross-entropy loss over the augmented source distribution, where an upper bound of the expected loss is derived and minimized, introducing negligible computational overhead. As a light-weight and general technique, TSA can be easily plugged into various domain adaptation methods, bringing remarkable improvements. Comprehensive experiments on cross-domain benchmarks validate the efficacy of TSA.

Unsupervised domain adaptation (UDA) enables transferring knowledge from a related source domain to a fully unlabeled target domain. Despite the significant advances in UDA, the performance gap remains quite large between UDA and supervised learning with fully labeled target data. Active domain adaptation (ADA) mitigates the gap under minimal annotation cost by selecting a small quota of target samples to annotate and incorporating them into training. Due to the domain shift, the query selection criteria of prior active learning methods may be ineffective to select the most informative target samples for annotation. In this paper, we propose Transferable Query Selection (TQS), which selects the most informative samples under domain shift by an ensemble of three new criteria: transferable committee, transferable uncertainty, and transferable domainness. We further develop a randomized selection algorithm to enhance the diversity of the selected samples. Experiments show that TQS remarkably outperforms previous UDA and ADA methods on several domain adaptation datasets. Deeper analyses demonstrate that TQS can select the most informative target samples under the domain shift.

Recent breakthroughs of Neural Architecture Search (NAS) extend the field's research scope towards a broader range of vision tasks and more diversified search spaces. While existing NAS methods mostly design architectures on a single task, algorithms that look beyond single-task search are surging to pursue a more efficient and universal solution across various tasks. Many of them leverage transfer learning and seek to preserve, reuse, and refine network design knowledge to achieve higher efficiency in future tasks. However, the enormous computational cost and experiment complexity of cross-task NAS are imposing barriers for valuable research in this direction. Existing NAS benchmarks all focus on one type of vision task, i.e., classification. In this work, we propose TransNAS-Bench-101, a benchmark dataset containing network performance across seven tasks, covering classification, regression, pixel-level prediction, and self-supervised tasks. This diversity provides opportunities to transfer NAS methods among tasks and allows for more complex transfer schemes to evolve. We explore two fundamentally different types of search space: cell-level search space and macro-level search space. With 7,352 backbones evaluated on seven tasks, 51,464 trained models with detailed training information are provided. With TransNAS-Bench-101, we hope to encourage the advent of exceptional NAS algorithms that raise cross-task search efficiency and generalizability to the next level. Our dataset and code will be available at Mindspore and VEGA.

Image inpainting is the task of plausibly restoring missing pixels within a hole region that is to be removed from a target image. Most existing technologies exploit patch similarities within the image, or leverage large-scale training data to fill the hole using learned semantic and texture information. However, due to the ill-posed nature of the inpainting task, such methods struggle to complete larger holes containing complicated scenes. In this paper, we propose TransFill, a multi-homography transformed fusion method to fill the hole by referring to another source image that shares scene contents with the target image. We first align the source image to the target image by estimating multiple homographies guided by different depth levels. We then learn to adjust the color and apply a pixel-level warping to each homography-warped source image to make it more consistent with the target. Finally, a pixel-level fusion module is learned to selectively merge the different proposals. Our method achieves state-of-the-art performance on pairs of images across a variety of wide baselines and color differences, and generalizes to user-provided image pairs.

Human trajectory prediction is critical for autonomous platforms like self-driving cars or social robots. We present a latent belief energy-based model (LB-EBM) for diverse human trajectory forecast. LB-EBM is a probabilistic model with cost function defined in the latent space to account for the movement history and social context. The low-dimensionality of the latent space and the high expressivity of the EBM make it easy for the model to capture the multimodality of pedestrian trajectory distributions. LB-EBM is learned from expert demonstrations (i.e., human trajectories) projected into the latent space. Sampling from or optimizing the learned LB-EBM yields a belief vector which is used to make a path plan, which then in turn helps to predict a long-range trajectory. The effectiveness of LB-EBM and the two-step approach are supported by strong empirical results. Our model is able to make accurate, multi-modal, and social compliant trajectory predictions and improves over prior state-of-the-arts performance on the Stanford Drone trajectory prediction benchmark by 10.9% and on the ETH-UCY benchmark by 27.6%.

In the setting of continual learning, a network is trained on a sequence of tasks, and suffers from catastrophic forgetting. To balance plasticity and stability of network in continual learning, in this paper, we propose a novel network training algorithm called Adam-NSCL, which sequentially optimizes network parameters in the null space of previous tasks. We first propose two mathematical conditions respectively for achieving network stability and plasticity in continual learning. Based on them, the network training for sequential tasks can be simply achieved by projecting the candidate parameter update into the approximate null space of all previous tasks in the network training process, where the candidate parameter update can be generated by Adam. The approximate null space can be derived by applying singular value decomposition to the uncentered covariance matrix of all input features of previous tasks for each linear layer. For efficiency, the uncentered covariance matrix can be incrementally computed after learning each task. We also empirically verify the rationality of the approximate null space at each linear layer. We apply our approach to training networks for continual learning on benchmark datasets of CIFAR-100 and TinyImageNet, and the results suggest that the proposed approach outperforms or matches the state-ot-the-art continual learning approaches.

Generative Adversarial Networks (GANs) have demonstrated unprecedented success in various image generation tasks. The encouraging results, however, come at the price of a cumbersome training process, during which the generator and discriminator are alternately updated in two stages. In this paper, we investigate a general training scheme that enables training GANs efficiently in only one stage. Based on the adversarial losses of the generator and discriminator, we categorize GANs into two classes, Symmetric GANs and Asymmetric GANs, and introduce a novel gradient decomposition method to unify the two, allowing us to train both classes in one stage and hence alleviate the training effort. We also computationally analyze the efficiency of the proposed method, and empirically demonstrate that, the proposed method yields a solid: 1.5x acceleration across various datasets and network architectures. Furthermore, we show that the proposed method is readily applicable to other adversarial-training scenarios, such as data-free knowledge distillation. The code is available at https://github.com/zju-vipa/OSGAN.

Simulation has the potential to massively scale evaluation of self-driving systems, enabling rapid development as well as safe deployment. Bridging the gap between simulation and the real world requires realistic multi-agent behaviors. Existing simulation environments rely on heuristic-based models that directly encode traffic rules, which cannot capture irregular maneuvers (e.g., nudging, U-turns) and complex interactions (e.g., yielding, merging). In contrast, we leverage real-world data to learn directly from human demonstration, and thus capture more naturalistic driving behaviors. To this end, we propose TrafficSim, a multi-agent behavior model for realistic traffic simulation. In particular, we parameterize the policy with an implicit latent variable model that generates socially-consistent plans for all actors in the scene jointly. To learn a robust policy amenable for long horizon simulation, we unroll the policy in training and optimize through the fully differentiable simulation across time. Our learning objective incorporates both human demonstrations as well as common sense. We show TrafficSim generates significantly more realistic traffic scenarios as compared to a diverse set of baselines. Notably, we can exploit trajectories generated by TrafficSim as effective data augmentation for training better motion planner.

Tracking humans in crowded video sequences is an important constituent of visual scene understanding. Increasing crowd density challenges visibility of humans, limiting the scalability of existing pedestrian trackers to higher crowd densities. For that reason, we propose to revitalize head tracking with Crowd of Heads Dataset (CroHD), consisting of 9 sequences of 11,463 frames with over 2,276,838 heads and 5,230 tracks annotated in diverse scenes. For evaluation, we proposed a new metric, IDEucl, to measure an algorithm's efficacy in preserving a unique identity for the longest stretch in image coordinate space, thus building a correspondence between pedestrian crowd motion and the performance of a tracking algorithm. Moreover, we also propose a new head detector, HeadHunter, which is designed for small head detection in crowded scenes. We extend HeadHunter with a Particle Filter and a color histogram based re-identification module for head tracking. To establish this as a strong baseline, we compare our tracker with existing state-of-the-art pedestrian trackers on CroHD and demonstrate superiority, especially in identity preserving tracking metrics. With a light-weight head detector and a tracker which is efficient at identity preservation, we believe our contributions will serve useful in advancement of pedestrian tracking in dense crowds. We make our dataset, code and models publicly available at https://project.inria.fr/crowdscience/project/dense-crowd-head-tracking/.

We propose an unsupervised method for detecting and tracking moving objects in 3D, in unlabelled RGB-D videos. The method begins with classic handcrafted techniques for segmenting objects using motion cues: we estimate optical flow and camera motion, and conservatively segment regions that appear to be moving independently of the background. Treating these initial segments as pseudo-labels, we learn an ensemble of appearance-based 2D and 3D detectors, under heavy data augmentation. We use this ensemble to detect new instances of the "moving" type, even if they are not moving, and add these as new pseudo-labels. Our method is an expectation-maximization algorithm, where in the expectation step we fire all modules and look for agreement among them, and in the maximization step we re-train the modules to improve this agreement. The constraint of ensemble agreement helps combat contamination of the generated pseudo-labels (during the E step), and data augmentation helps the modules generalize to yet-unlabelled data (during the M step). We compare against existing unsupervised object discovery and tracking methods, using challenging videos from CATER and KITTI, and show strong improvements over the state-of-the-art.

Most online multi-object trackers perform object detection stand-alone in a neural net without any input from tracking. In this paper, we present a new online joint detection and tracking model, TraDeS (TRAck to DEtect and Segment), exploiting tracking clues to assist detection end-to-end. TraDeS infers object tracking offset by a cost volume, which is used to propagate previous object features for improving current object detection and segmentation. Effectiveness and superiority of TraDeS are shown on 4 datasets, including MOT (2D tracking), nuScenes (3D tracking), MOTS and Youtube-VIS (instance segmentation tracking). Project page: https://jialianwu.com/projects/TraDeS.html.

Surgical skills have a great influence on surgical safety and patients' well-being. Traditional assessment of surgical skills involves strenuous manual efforts, which lacks efficiency and repeatability. Therefore, we attempt to automatically predict how well the surgery is performed using the surgical video. In this paper, a unified multi-path framework for automatic surgical skill assessment is proposed, which takes care of multiple composing aspects of surgical skills, including surgical tool usage, intraoperative event pattern, and other skill proxies. The dependency relationships among these different aspects are specially modeled by a path dependency module in the framework. We conduct extensive experiments on the JIGSAWS dataset of simulated surgical tasks, and a new clinical dataset of real laparoscopic surgeries. The proposed framework achieves promising results on both datasets, with the state-of-the-art on the simulated dataset advanced from 0.71 Spearman's correlation to 0.80. It is also shown that combining multiple skill aspects yields better performance than relying on a single aspect.