CVPR 2021 论文列表

Multi-pedestrian trackers perform well when targets are clearly visible making the association task quite easy. However, when heavy occlusions are present, a mechanism to reidentify persons is needed. The common approach is to extract visual features from new detections and compare them with the features of previously found tracks. Since those detections can have substantial overlaps with nearby targets - especially in crowded scenarios - the extracted features are insufficient for a reliable re-identification. In contrast, we propose a novel occlusion handling strategy that explicitly models the relation between occluding and occluded tracks outperforming the feature-based approach, while not depending on a separate re-identification network. Furthermore, we improve the track management of a regression-based method in order to bypass missing detections and to deal with tracks leaving the scene at the border of the image. Finally, we apply our tracker in both temporal directions and merge tracklets belonging to the same target, which further enhances the performance. We demonstrate the effectiveness of our tracking components with ablative experiments and surpass the state-of-the-art methods on the three popular pedestrian tracking benchmarks MOT16, MOT17, and MOT20.

Despite considerable similarities between multiple object tracking (MOT) and single object tracking (SOT) tasks, modern MOT methods have not benefited from the development of SOT ones to achieve satisfactory performance. The major reason for this situation is that it is inappropriate and inefficient to apply multiple SOT models directly to the MOT task, although advanced SOT methods are of the strong discriminative power and can run at fast speeds. In this paper, we propose a novel and end-to-end trainable MOT architecture that extends CenterNet by adding an SOT branch for tracking objects in parallel with the existing branch for object detection, allowing the MOT task to benefit from the strong discriminative power of SOT methods in an effective and efficient way. Unlike most existing SOT methods which learn to distinguish the target object from its local backgrounds, the added SOT branch trains a separate SOT model per target online to distinguish the target from its surrounding targets, assigning SOT models the novel discrimination. Moreover, similar to the detection branch, the SOT branch treats objects as points, making its online learning efficient even if multiple targets are processed simultaneously. Without tricks, the proposed tracker achieves MOTAs of 0.710 and 0.686, IDF1s of 0.719 and 0.714, on MOT17 and MOT20 benchmarks, respectively, while running at 16 FPS on MOT17.

Deep neural networks may perform poorly when training datasets are heavily class-imbalanced. Recently, two-stage methods decouple representation learning and classifier learning to improve performance. But there is still the vital issue of miscalibration. To address it, we design two methods to improve calibration and performance in such scenarios. Motivated by the fact that predicted probability distributions of classes are highly related to the numbers of class instances, we propose label-aware smoothing to deal with different degrees of over-confidence for classes and improve classifier learning. For dataset bias between these two stages due to different samplers, we further propose shifted batch normalization in the decoupling framework. Our proposed methods set new records on multiple popular long-tailed recognition benchmark datasets, including CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, Places-LT, and iNaturalist 2018.

Binarization of neural network models is considered as one of the promising methods to deploy deep neural network models on resource-constrained environments such as mobile devices. However, Binary Neural Networks (BNNs) tend to suffer from severe accuracy degradation compared to the full-precision counterpart model. Several techniques were proposed to improve the accuracy of BNNs. One of the approaches is to balance the distribution of binary activations so that the amount of information in the binary activations becomes maximum. Based on extensive analysis, in stark contrast to previous work, we argue that unbalanced activation distribution can actually improve the accuracy of BNNs. We also show that adjusting the threshold values of binary activation functions results in the unbalanced distribution of the binary activation, which increases the accuracy of BNN models. Experimental results show that the accuracy of previous BNN models (e.g. XNOR-Net and Bi-Real-Net) can be improved by simply shifting the threshold values of binary activation functions without requiring any other modification.

Image matting is a fundamental and challenging problem in computer vision and graphics. Most existing matting methods leverage a user-supplied trimap as an auxiliary input to produce good alpha matte. However, obtaining high-quality trimap itself is arduous, thus restricting the application of these methods. Recently, some trimap-free methods have emerged, however, the matting quality is still far behind the trimap-based methods. The main reason is that, without the trimap guidance in some cases, the target network is ambiguous about which is the foreground target. In fact, choosing the foreground is a subjective procedure and depends on the user's intention. To this end, this paper proposes an improved deep image matting framework which is trimap-free and only needs several user click interactions to eliminate the ambiguity. Moreover, we introduce a new uncertainty estimation module that can predict which parts need polishing and a following local refinement module. Based on the computation budget, users can choose how many local parts to improve with the uncertainty guidance. Quantitative and qualitative results show that our method performs better than existing trimap-free methods and comparably to state-of-the-art trimap-based methods with minimal user effort.

We wish to detect specific categories of objects, for online vision systems that will run in the real world. Object detection is already very challenging. It is even harder when the images are blurred, from the camera being in a car or a hand-held phone. Most existing efforts either focused on sharp images, with easy to label ground truth, or they have treated motion blur as one of many generic corruptions. Instead, we focus especially on the details of egomotion induced blur. We explore five classes of remedies, where each targets different potential causes for the performance gap between sharp and blurred images. For example, first deblurring an image changes its human interpretability, but at present, only partly improves object detection. The other four classes of remedies address multi-scale texture, out-of-distribution testing, label generation, and conditioning by blur-type. Surprisingly, we discover that custom label generation aimed at resolving spatial ambiguity, ahead of all others, markedly improves object detection. Also, in contrast to findings from classification, we see a noteworthy boost by conditioning our model on bespoke categories of motion blur. We validate and cross-breed the different remedies experimentally on blurred COCO images and real-world blur datasets, producing an easy and practical favorite model with superior detection rates.

Text recognition is a popular research subject with many associated challenges. Despite the considerable progress made in recent years, the text recognition task itself is still constrained to solve the problem of reading cropped line text images and serves as a subtask of optical character recognition (OCR) systems. As a result, the final text recognition result is limited by the performance of the text detector. In this paper, we propose a simple, elegant and effective paradigm called Implicit Feature Alignment (IFA), which can be easily integrated into current text recognizers, resulting in a novel inference mechanism called IFA-inference. This enables an ordinary text recognizer to process multi-line text such that text detection can be completely freed. Specifically, we integrate IFA into the two most prevailing text recognition streams (attention-based and CTC-based) and propose attention-guided dense prediction (ADP) and Extended CTC (ExCTC). Furthermore, the Wasserstein-based Hollow Aggregation Cross-Entropy (WH-ACE) is proposed to suppress negative predictions to assist in training ADP and ExCTC. We experimentally demonstrate that IFA achieves state-of-the-art performance on end-to-end document recognition tasks while maintaining the fastest speed, and ADP and ExCTC complement each other on the perspective of different application scenarios. Code will be available at https://github.com/Wang-Tianwei/Implicit-feature-alignment.

Recently, image-to-image translation has made significant progress in achieving both multi-label (i.e., translation conditioned on different labels) and multi-style (i.e., generation with diverse styles) tasks. However, due to the unexplored independence and exclusiveness in the labels, existing endeavors are defeated by involving uncontrolled manipulations to the translation results. In this paper, we propose Hierarchical Style Disentanglement (HiSD) to address this issue. Specifically, we organize the labels into a hierarchical tree structure, in which independent tags, exclusive attributes, and disentangled styles are allocated from top to bottom. Correspondingly, a new translation process is designed to adapt the above structure, in which the styles are identified for controllable translations. Both qualitative and quantitative results on the CelebA-HQ dataset verify the ability of the proposed HiSD. We hope our method will serve as a solid baseline and provide fresh insights with the hierarchically organized annotations for future research in image-to-image translation. The code will be released.

Both non-local (NL) operation and sparse representation are crucial for Single Image Super-Resolution (SISR). In this paper, we investigate their combinations and propose a novel Non-Local Sparse Attention (NLSA) with dynamic sparse attention pattern. NLSA is designed to retain long-range modeling capability from NL operation while enjoying robustness and high-efficiency of sparse representation. Specifically, NLSA rectifies NL attention with spherical locality sensitive hashing (LSH) that partitions the input space into hash buckets of related features. For every query signal, NLSA assigns a bucket to it and only computes attention within the bucket. The resulting sparse attention prevents the model from attending to locations that are noisy and less-informative, while reducing the computational cost from quadratic to asymptotic linear with respect to the spatial size. Extensive experiments validate the effectiveness and efficiency of NLSA. With a few non-local sparse attention modules, our architecture, called non-local sparse network (NLSN), reaches state-of-the-art performance for SISR quantitatively and qualitatively.

The new trend of full-screen devices encourages us to position a camera behind a screen. Removing the bezel and centralizing the camera under the screen brings larger display-to-body ratio and enhances eye contact in video chat, but also causes image degradation. In this paper, we focus on a newly-defined Under-Display Camera (UDC), as a novel real-world single image restoration problem. First, we take a 4k Transparent OLED (T-OLED) and a phone Pentile OLED (P-OLED) and analyze their optical systems to understand the degradation. Second, we design a Monitor-Camera Imaging System (MCIS) for easier real pair data acquisition, and a model-based data synthesizing pipeline to generate Point Spread Function (PSF) and UDC data only from display pattern and camera measurements. Finally, we resolve the complicated degradation using deconvolution-based pipeline and learning-based methods. Our model demonstrates a real-time high-quality restoration. The presented methods and results reveal the promising research values and directions of UDC.

Although recent inpainting approaches have demonstrated significant improvement with deep neural networks, they still suffer from artifacts such as blunt structures and abrupt colors when filling in the missing regions. To address these issues, we propose an external-internal inpainting scheme with a monochromic bottleneck that helps image inpainting models remove these artifacts. In the external learning stage, we reconstruct missing structures and details in the monochromic space to reduce the learning dimension. In the internal learning stage, we propose a novel internal color propagation method with progressive learning strategies for consistent color restoration. Extensive experiments demonstrate that our proposed scheme helps image inpainting models produce more structure-preserved and visually compelling results.

Existing inpainting methods have achieved promising performance in recovering defected images of specific scenes. However, filling holes involving multiple semantic categories remains challenging due to the obscure semantic boundaries and the mixture of different semantic textures. In this paper, we introduce coherence priors between the semantics and textures which make it possible to concentrate on completing separate textures in a semantic-wise manner. Specifically, we adopt a multi-scale joint optimization framework to first model the coherence priors and then accordingly interleavingly optimize image inpainting and semantic segmentation in a coarse-to-fine manner. A Semantic-Wise Attention Propagation (SWAP) module is devised to refine completed image textures across scales by exploring non-local semantic coherence, which effectively mitigates mix-up of textures. We also propose two coherence losses to constrain the consistency between the semantics and the inpainted image in terms of the overall structure and detailed textures. Experimental results demonstrate the superiority of our proposed method for challenging cases with complex holes.

Existing image generator networks rely heavily on spatial convolutions and, optionally, self-attention blocks in order to gradually synthesize images in a coarse-to-fine manner. Here, we present a new architecture for image generators, where the color value at each pixel is computed independently given the value of a random latent vector and the coordinate of that pixel. No spatial convolutions or similar operations that propagate information across pixels are involved during the synthesis. We analyze the modeling capabilities of such generators when trained in an adversarial fashion, and observe the new generators to achieve similar generation quality to state-of-the-art convolutional generators. We also investigate several interesting properties unique to the new architecture.

While deep convolutional neural networks (CNNs) have achieved great success on image de-raining task, most existing methods can only learn fixed mapping rules between paired rainy/clean images on a single dataset. This limits their applications in practical situations with multiple and incremental datasets where the mapping rules may change for different types of rain streaks. However, the catastrophic forgetting of traditional deep CNN model challenges the design of generalized framework for multiple and incremental datasets. A strategy of sharing the network structure but independently updating and storing the network parameters on each dataset has been developed as a potential solution. Nevertheless, this strategy is not applicable to compact systems as it dramatically increases the overall training time and parameter space. To alleviate such limitation, in this study, we propose a parameter importance guided weights modification approach, named PIGWM. Specifically, with new dataset (e.g. new rain dataset), the well-trained network weights are updated according to their importance evaluated on previous training dataset. With extensive experimental validation, we demonstrate that a single network with a single parameter set of our proposed method can process multiple rain datasets almost without performance degradation. The proposed model is capable of achieving superior performance on both inhomogeneous and incremental datasets, and is promising for highly compact systems to gradually learn myriad regularities of the different types of rain streaks. The results indicate that our proposed method has great potential for other computer vision tasks with dynamic learning environments.

We tackle the challenging task of image change captioning. The goal is to describe the subtle difference between two very similar images by generating a sentence caption. While the recent methods mainly focus on proposing new model architectures for this problem, we instead focus on an alternative training scheme. Inspired by the success of multi-task learning, we formulate a training scheme that uses an auxiliary task to improve the training of the change captioning network. We argue that the task of composed query image retrieval is a natural choice as the auxiliary task. Given two almost similar images as the input, the primary network generates a caption describing the fine change between those two images. Next, the auxiliary network is provided with the generated caption and one of those two images. It then tries to pick the second image among a set of candidates. This forces the primary network to generate detailed and precise captions via having an extra supervision loss by the auxiliary network. Furthermore, we propose a new scheme for selecting a negative set of candidates for the retrieval task that can effectively improve the performance. We show that the proposed training strategy performs well on the task of change captioning on benchmark datasets.

Vector graphics are widely used to represent fonts, logos, digital artworks, and graphic designs. But, while a vast body of work has focused on generative algorithms for raster images, only a handful of options exists for vector graphics. One can always rasterize the input graphic and resort to image-based generative approaches, but this negates the advantages of the vector representation. The current alternative is to use specialized models that require explicit supervision on the vector graphics representation at training time. This is not ideal because large-scale high-quality vector-graphics datasets are difficult to obtain. Furthermore, the vector representation for a given design is not unique, so models that supervise on the vector representation are unnecessarily constrained. Instead, we propose a new neural network that can generate complex vector graphics with varying topologies, and only requires in-direct supervision from readily-available raster training images (i.e., with no vector counterparts). To enable this, we use a differentiable rasterization pipeline that renders the generated vector shapes and composites them together onto a raster canvas. We demonstrate our method on a range of datasets, and provide comparison with state-of-the-art SVG-VAE and DeepSVG, both of which require explicit vector graphics supervision. Finally, we also demonstrate our approach on the MNIST dataset, for which no groundtruth vector representation is available.

We propose a dense object detector with an instance-wise sampling strategy, named IQDet. Instead of using human prior sampling strategies, we first extract the regional feature of each ground-truth to estimate the instance-wise quality distribution. According to a mixture model in spatial dimensions, the distribution is more noise-robust and adapted to the semantic pattern of each instance. Based on the distribution, we propose a quality sampling strategy, which automatically selects training samples in a probabilistic manner and trains with more high-quality samples. Extensive experiments on MS COCO show that our method steadily improves baseline by nearly 2.4 AP without bells and whistles. Moreover, our best model achieves 51.6 AP, outperforming all existing state-of-the-art one-stage detectors and it is completely cost-free in inference time.

A natural way to model the evolution of an object (growth of a leaf for instance) is to estimate a plausible deforming path between two observations. This interpolation process can generate deceiving results when the set of considered deformations is not relevant to the observed data. To overcome this issue, the framework of deformation modules allows to incorporate in the model structured deformation patterns coming from prior knowledge on the data. The goal of this article is twofold. First defining new deformation modules incorporating structures coming from the elastic properties of the objects. Second, presenting the IMODAL library allowing to perform registration through structured deformations. This library is modular: adapted priors can be easily defined by the user, several priors can be combined into a global one and various types of data can be considered such as curves, meshes or images.

Synthesizing variations of a specific reference image with semantically valid content is an important task in terms of personalized generation as well as for data augmentation. In this work, we propose an inversion based method, denoted as IMAge-Guided model INvErsion (IMAGINE), to generate high-quality and diverse images only from one single training sample. We mainly leverage the knowledge of image semantics from a pre-trained classifier and achieve plausible generations via matching multi-level feature representations in the classifier, associated with adversarial training with an external discriminator. IMAGINE enables the synthesis procedure to be able to simultaneously 1) enforce semantic specificity constraints during the synthesis, 2) produce realistic images without the introduction of generator training, 3) allow fine controls over the synthesized image, and 4) be model-compact. With extensive experimental results, we demonstrate qualitatively and quantitatively that IMAGINE performs favorably against state-of-the-art GAN-based and inversion-based methods, across three different image domains, i.e., the object, scene and texture.