CVPR 2021 论文列表

Occluded person re-identification (Re-ID) is a challenging task as persons are frequently occluded by various obstacles or other persons, especially in the crowd scenario. To address these issues, we propose a novel end-to-end Part-Aware Transformer (PAT) for occluded person Re-ID through diverse part discovery via a transformer encoder-decoder architecture, including a pixel context based transformer encoder and a part prototype based transformer decoder. The proposed PAT model enjoys several merits. First, to the best of our knowledge, this is the first work to exploit the transformer encoder-decoder architecture for occluded person Re-ID in a unified deep model. Second, to learn part prototypes well with only identity labels, we design two effective mechanisms including part diversity and part discriminability. Consequently, we can achieve diverse part discovery for occluded person Re-ID in a weakly supervised manner. Extensive experimental results on six challenging benchmarks for three tasks (occluded, partial and holistic Re-ID) demonstrate that our proposed PAT performs favorably against stat-of-the-art methods.

We propose a universal building block of Convolutional Neural Network (ConvNet) to improve the performance without any inference-time costs. The block is named Diverse Branch Block (DBB), which enhances the representational capacity of a single convolution by combining diverse branches of different scales and complexities to enrich the feature space, including sequences of convolutions, multi-scale convolutions, and average pooling. After training, a DBB can be equivalently converted into a single conv layer for deployment. Unlike the advancements of novel ConvNet architectures, DBB complicates the training-time microstructure while maintaining the macro architecture, so that it can be used as a drop-in replacement for regular conv layers of any architecture. In this way, the model can be trained to reach a higher level of performance and then transformed into the original inference-time structure for inference. DBB improves ConvNets on image classification (up to 1.9% higher top-1 accuracy on ImageNet), object detection and semantic segmentation. The PyTorch code and models are released at https://github.com/DingXiaoH/DiverseBranchBlock.

Universal domain adaptation (UniDA) has been proposed to transfer knowledge learned from a label-rich source domain to a label-scarce target domain without any constraints on the label sets. In practice, however, it is difficult to obtain a large amount of perfectly clean labeled data in a source domain with limited resources. Existing UniDA methods rely on source samples with correct annotations, which greatly limits their application in the real world. Hence, we consider a new realistic setting called Noisy UniDA, in which classifiers are trained with noisy labeled data from the source domain and unlabeled data with an unknown class distribution from the target domain. This paper introduces a two-head convolutional neural network framework to solve all problems simultaneously. Our network consists of one common feature generator and two classifiers with different decision boundaries. By optimizing the divergence between the two classifiers' outputs, we can detect noisy source samples, find "unknown" classes in the target domain, and align the distribution of the source and target domains. In an extensive evaluation of different domain adaptation settings, the proposed method outperformed existing methods by a large margin in most settings.

Due to the subjective annotation and the inherent inter-class similarity of facial expressions, one of key challenges in Facial Expression Recognition (FER) is the annotation ambiguity. In this paper, we proposes a solution, named DMUE, to address the problem of annotation ambiguity from two perspectives: the latent Distribution Mining and the pairwise Uncertainty Estimation. For the former, an auxiliary multi-branch learning framework is introduced to better mine and describe the latent distribution in the label space. For the latter, the pairwise relationship of semantic feature between instances are fully exploited to estimate the ambiguity extent in the instance space. The proposed method is independent to the backbone architectures, and brings no extra burden for inference. The experiments are conducted on the popular real-world benchmarks and the synthetic noisy datasets. Either way, the proposed DMUE stably achieves leading performance.

In this paper, we explore the compression of deep neural networks by quantizing the weights and activations into multi-bit binary networks (MBNs). A distribution-aware multi-bit quantization (DMBQ) method that incorporates the distribution prior into the optimization of quantization is proposed. Instead of solving the optimization in each iteration, DMBQ search the optimal quantization scheme over the distribution space beforehand, and select the quantization scheme during training using a fast lookup table based strategy. Based upon DMBQ, we further propose loss-guided bit-width allocation (LBA) to adaptively quantize and even prune the neural network. The first-order Taylor expansion is applied to build a metric for evaluating the loss sensitivity of the quantization of each channel, and automatically adjust the bit-width of weights and activations channel-wisely. We extend our method to image classification tasks and experimental results show that our method not only outperforms state-of-the-art quantized networks in terms of accuracy but also is more efficient in terms of training time compared with state-of-the-art MBNs, even for the extremely low bit width (below 1-bit) quantization cases.

Despite the success of the deep neural networks, it remains challenging to effectively build a system for long-tail visual recognition tasks. To address this problem, we first investigate the performance bottleneck of the two-stage learning framework via ablative study. Motivated by our discovery, we develop a unified distribution alignment strategy for long-tail visual recognition. Particularly, we first propose an adaptive calibration strategy for each data point to calibrate its classification scores. Then we introduce a generalized re-weight method to incorporate the class prior, which provides a flexible and unified solution to copy with diverse scenarios of various visual recognition tasks. We validate our method by extensive experiments on four tasks, including image classification, semantic segmentation, object detection, and instance segmentation. Our approach achieves the state-of-the-art results across all four recognition tasks with a simple and unified framework.

A practical long-term tracker typically contains three key properties, i.e., an efficient model design, an effective global re-detection strategy and a robust distractor awareness mechanism. However, most state-of-the-art long-term trackers (e.g., Pseudo and re-detecting based ones) do not take all three key properties into account and therefore may either be time-consuming or drift to distractors. To address the issues, we propose a two-task tracking framework (named DMTrack), which utilizes two core components (i.e., one-shot detection and re-identification (re-id) association) to achieve distractor-aware fast tracking via Dynamic convolutions (d-convs) and Multiple object tracking (MOT) philosophy. To achieve precise and fast global detection, we construct a lightweight one-shot detector using a novel dynamic convolutions generation method, which provides a unified and more flexible way for fusing target information into the search field. To distinguish the target from distractors, we resort to the philosophy of MOT to reason distractors explicitly by maintaining all potential similarities' tracklets. Benefited from the strength of high recall detection and explicit object association, our tracker achieves state-of-the-art performance on the LaSOT, OxUvA, TLP, VOT2018LT and VOT2019LT benchmarks and runs in real-time (3x faster than comparisons).

Knowledge distillation is a widely used paradigm for inheriting information from a complicated teacher network to a compact student network and maintaining the strong performance. Different from image classification, object detectors are much more sophisticated with multiple loss functions in which features that semantic information rely on are tangled. In this paper, we point out that the information of features derived from regions excluding objects are also essential for distilling the student detector, which is usually ignored in existing approaches. In addition, we elucidate that features from different regions should be assigned with different importance during distillation. To this end, we present a novel distillation algorithm via decoupled features (DeFeat) for learning a better student detector. Specifically, two levels of decoupled features will be processed for embedding useful information into the student, i.e., decoupled features from neck and decoupled proposals from classification head. Extensive experiments on various detectors with different backbones show that the proposed DeFeat is able to surpass the state-of-the-art distillation methods for object detection. For example, DeFeat improves ResNet50 based Faster R-CNN from 37.4% to 40.9% mAP, and improves ResNet50 based RetinaNet from 36.5% to 39.7% mAP on COCO benchmark. Code will be released.

Knowledge distillation transfers knowledge from the teacher network to the student one, with the goal of greatly improving the performance of the student network. Previous methods mostly focus on proposing feature transformation and loss functions between the same level's features to improve the effectiveness. We differently study the factor of connection path cross levels between teacher and student networks, and reveal its great importance. For the first time in knowledge distillation, cross-stage connection paths are proposed. A new review mechanism becomes vastly effective and structurally simple. Our finally designed nested and compact framework requires negligible computation overhead, and outperforms other methods on a variety of tasks. We apply our method to classification, object detection, and instance segmentation tasks. All of them witness significant student network performance improvement.

We propose a causal framework to explain the catastrophic forgetting in Class-Incremental Learning (CIL) and then derive a novel distillation method that is orthogonal to the existing anti-forgetting techniques, such as data replay and feature/label distillation. We first 1) place CIL into the framework, 2) answer why the forgetting happens: the causal effect of the old data is lost in new training, and then 3) explain how the existing techniques mitigate it: they bring the causal effect back. Based on the causal framework, we propose to distill the Colliding Effect between the old and the new data, which is fundamentally equivalent to the causal effect of data replay, but without any cost of replay storage. Thanks to the causal effect analysis, we can further capture the Incremental Momentum Effect of the data stream, removing which can help to retain the old effect overwhelmed by the new data effect, and thus alleviate the forgetting of the old class in testing. Extensive experiments on three CIL benchmarks: CIFAR-100, ImageNet-Sub&Full, show that the proposed causal effect distillation can improve various state-of-the-art CIL methods by a large margin (0.72%-9.06%)

Having access to multi-modal cues (e.g. vision and audio) empowers some cognitive tasks to be done faster compared to learning from a single modality. In this work, we propose to transfer knowledge across heterogeneous modalities, even though these data modalities may not be semantically correlated. Rather than directly aligning the representations of different modalities, we compose audio, image, and video representations across modalities to uncover the richer multi-modal knowledge. Our main idea is to learn a compositional embedding that closes the cross-modal semantic gap and captures the task-relevant semantics, which facilitates pulling together representations across modalities by compositional contrastive learning. We establish a new, comprehensive multi-modal distillation benchmark on three video datasets: UCF101, ActivityNet, and VGGSound. Moreover, we demonstrate that our model significantly outperforms a variety of existing knowledge distillation methods in transferring audio-visual knowledge to improve video representation learning.

The current evaluation protocol of long-tailed visual recognition trains the classification model on the long-tailed source label distribution and evaluates its performance on the uniform target label distribution. Such protocol has questionable practicality since the target may also be long-tailed. Therefore, we formulate long-tailed visual recognition as a label shift problem where the target and source label distributions are different. One of the significant hurdles in dealing with the label shift problem is the entanglement between the source label distribution and the model prediction. In this paper, we focus on disentangling the source label distribution from the model prediction. We first introduce a simple but overlooked baseline method that matches the target label distribution by post-processing the model prediction trained by the cross-entropy loss and the Softmax function. Although this method surpasses state-of-the-art methods on benchmark datasets, it can be further improved by directly disentangling the source label distribution from the model prediction in the training phase. Thus, we propose a novel method, LAbel distribution DisEntangling (LADE) loss based on the optimal bound of Donsker-Varadhan representation. LADE achieves state-of-the-art performance on benchmark datasets such as CIFAR-100-LT, Places-LT, ImageNet-LT, and iNaturalist 2018. Moreover, LADE outperforms existing methods on various shifted target label distributions, showing the general adaptability of our proposed method.

Image virtual try-on replaces the clothes on a person image with a desired in-shop clothes image. It is challenging because the person and the in-shop clothes are unpaired. Existing methods formulate virtual try-on as either in-painting or cycle consistency. Both of these two formulations encourage the generation networks to reconstruct the input image in a self-supervised manner. However, existing methods do not differentiate clothing and non-clothing regions. A straightforward generation impedes the virtual try-on quality because of the heavily coupled image contents. In this paper, we propose a Disentangled Cycle-consistency Try-On Network (DCTON). The DCTON is able to produce highly-realistic try-on images by disentangling important components of virtual try-on including clothes warping, skin synthesis, and image composition. Moreover, DCTON can be naturally trained in a self-supervised manner following cycle consistency learning. Extensive experiments on challenging benchmarks show that DCTON outperforms state-of-the-art approaches favorably.

In multi-object tracking, the tracker maintains in its memory the appearance and motion information for each object in the scene. This memory is utilized for finding matches between tracks and detections, and is updated based on the matching. Many approaches model each target in isolation and lack the ability to use all the targets in the scene to jointly update the memory. This can be problematic when there are similarly looking objects in the scene. In this paper, we solve the problem of simultaneously considering all tracks during memory updating, with only a small spatial overhead, via a novel multi-track pooling module. We additionally propose a training strategy adapted to multi-track pooling which generates hard tracking episodes online. We show that the combination of these innovations results in a strong discriminative appearance model under the bilinear LSTM tracking framework, enabling the use of greedy data association to achieve online tracking performance. Our experiments demonstrate real-time, state-of-the-art online tracking performance on public multi-object tracking (MOT) datasets.

Recently, with the emergence of retrieval requirements for certain individual in the same superclass, e.g., birds, persons, cars, fine-grained recognition task has attracted a significant amount of attention from academia and industry. In fine-grained recognition scenario, the inter-class differences are quite diverse and subtle, which makes it challenging to extract all the discriminative cues. Traditional training mechanism optimizes the overall discriminativeness of the whole feature. It may stop early when some feature elements has been trained to distinguish training samples well, leaving other elements insufficiently trained for a feature. This would result in a less generalizable feature extractor that only captures major discriminative cues and ignores subtle ones. Therefore, there is a need for a training mechanism that enforces the discriminativeness of all the elements in the feature to capture more the subtle visual cues. In this paper, we propose a Discrimination-Aware Mechanism (DAM) that iteratively identifies insufficiently trained elements and improves them. DAM is able to increase the number of well learned elements, which captures more visual cues by the feature extractor. In this way, a more informative representation is learned, which brings better generalization performance. We show that DAM can be easily applied to both proxy-based and pair-based loss functions, and thus can be used in most existing fine-grained recognition paradigms. Comprehensive experiments on CUB-200-2011, Cars196, Market-1501, and MSMT17 datasets demonstrate the advantages of our DAM based loss over the related state-of-the-art approaches.

Image-text matching is an important multi-modal task with massive applications. It tries to match the image and the text with similar semantic information. Existing approaches do not explicitly transform the different modalities into a common space. Meanwhile, the attention mechanism which is widely used in image-text matching models does not have supervision. We propose a novel attention scheme which projects the image and text embedding into a common space and optimises the attention weights directly towards the evaluation metrics. The proposed attention scheme can be considered as a kind of supervised attention and requiring no additional annotations. It is trained via a novel Discrete-continuous action space policy gradient algorithm, which is more effective in modelling complex action space than previous continuous action space policy gradient. We evaluate the proposed methods on two widely-used benchmark datasets: Flickr30k and MS-COCO, outperforming the previous approaches by a large margin.

We tackle the problem of learning the geometry of multiple categories of deformable objects jointly. Recent work has shown that it is possible to learn a unified dense pose predictor for several categories of related objects. However, training such models requires to initialize inter-category correspondences by hand. This is suboptimal and the resulting models fail to maintain correct correspondences as individual categories are learned. In this paper, we show that improved correspondences can be learned automatically as a natural byproduct of learning category-specific dense pose predictors. To do this, we express correspondences between different categories and between images and categories using a unified embedding. Then, we use the latter to enforce two constraints: symmetric inter-category cycle consistency and a new asymmetric image-to-category cycle consistency. Without any manual annotations for the inter-category correspondences, we obtain state-of-the-art alignment results, outperforming dedicated methods for matching 3D shapes. Moreover, the new model is also better at the task of dense pose prediction than prior work.

Generative adversarial networks (GANs) learn to map noise latent vectors to high-fidelity image outputs. It is found that the input latent space shows semantic correlations with the output image space. Recent works aim to interpret the latent space and discover meaningful directions that correspond to human interpretable image transformations. However, these methods either rely on explicit scores of attributes (e.g., memorability) or are restricted to binary ones (e.g., gender), which largely limits the applicability of editing tasks, especially for free-form artistic tasks like style/anime editing. In this paper, we propose an adversarial method, AdvStyle, for discovering interpretable directions in the absence of well-labeled scores or binary attributes. In particular, the proposed adversarial method simultaneously optimizes the discovered directions and the attribute assessor using the target attribute data as positive samples, while the generated ones being negative. In this way, arbitrary attributes can be edited by collecting positive data only, and the proposed method learns a controllable representation enabling manipulation of non-binary attributes like anime styles and facial characteristics. Moreover, the proposed learning strategy attenuates the entanglement between attributes, such that multiattribute manipulation can be easily achieved without any additional constraint. Furthermore, we reveal several interesting semantics with the involuntarily learned negative directions. Extensive experiments on 9 anime attributes and 7 human attributes demonstrate the effectiveness of our adversarial approach qualitatively and quantitatively. Code is available at https://github.com/BERYLSHEEP/AdvStyle.