CVPR 2021 Papers

Given a set of putative 2D-2D line correspondences, we aim to identify correct matches. Existing methods exploit the geometric constraints. They are only applicable to structured scenes with orthogonality, parallelism and coplanarity. In contrast, we propose the first approach suitable for both structured and unstructured scenes. Instead of geometric constraint, we leverage the spatial regularity on sphere. Specifically, we propose to map line correspondences into vectors tangent to sphere. We use these vectors to encode both angular and positional variations of image lines, which is more reliable and concise than directly using inclinations, midpoints or endpoints of image lines. Neighboring vectors mapped from correct matches exhibit a spatial regularity called local trend consistency, regardless of the type of scenes. To encode this regularity, we design a neural network and also propose a novel loss function that enforces the smoothness constraint of vector field. In addition, we establish a large real-world dataset for image line matching. Experiments showed that our approach outperforms state-of-the-art ones in terms of accuracy, efficiency and robustness, and also leads to high generalization.

Recently, Siamese-based trackers have achieved promising performance in visual tracking. Most recent Siamese-based trackers typically employ a depth-wise cross-correlation (DW-XCorr) to obtain multi-channel correlation information from the two feature maps (target and search region). However, DW-XCorr has several limitations within Siamese-based tracking: it can easily be fooled by distractors, has fewer activated channels, and provides weak discrimination of object boundaries. Further, DW-XCorr is a handcrafted parameter-free module and cannot fully benefit from offline learning on large-scale data. We propose a learnable module, called the asymmetric convolution (ACM), which learns to better capture the semantic correlation information in offline training on large-scale data. Different from DW-XCorr and its predecessor(XCorr), which regard a single feature map as the convolution kernel, our ACM decomposes the convolution operation on a concatenated feature map into two mathematically equivalent operations, thereby avoiding the need for the feature maps to be of the same size (width and height)during concatenation. Our ACM can incorporate useful prior information, such as bounding-box size, with standard visual features. Furthermore, ACM can easily be integrated into existing Siamese trackers based on DW-XCorror XCorr. To demonstrate its generalization ability, we integrate ACM into three representative trackers: SiamFC, SiamRPN++, and SiamBAN. Our experiments reveal the benefits of the proposed ACM, which outperforms existing methods on six tracking benchmarks. On the LaSOT test set, our ACM-based tracker obtains a significant improvement of 5.8% in terms of success (AUC), over the baseline.

Despite the great success of Siamese-based trackers, their performance under complicated scenarios is still not satisfying, especially when there are distractors. To this end, we propose a novel Siamese relation network, which introduces two efficient modules, i.e. Relation Detector (RD) and Refinement Module (RM). RD performs in a meta-learning way to obtain a learning ability to filter the distractors from the background while RM aims to effectively integrate the proposed RD into the Siamese framework to generate accurate tracking result.Moreover, to further improve the discriminability and robustness of the tracker, we introduce a contrastive training strategy that attempts not only to learn matching the same target but also to learn how to distinguish the different objects. Therefore, our tracker can achieve accurate tracking results when facing background clutters, fast motion, and occlusion. Experimental results on five popular benchmarks, including VOT2018, VOT2019, OTB100, LaSOT, and UAV123, show that the proposed method is effective and can achieve state-of-the-art results. The code will be available at https://github.com/hqucv/siamrn

Existing works on visual counting primarily focus on one specific category at a time, such as people, animals, and cells. In this paper, we are interested in counting everything, that is to count objects from any category given only a few annotated instances from that category. To this end, we pose counting as a few-shot regression task. To tackle this task, we present a novel method that takes a query image together with a few exemplar objects from the query image and predicts a density map for the presence of all objects of interest in the query image. We also present a novel adaptation strategy to adapt our network to any novel visual category at test time, using only a few exemplar objects from the novel category. We also introduce a dataset of 147 object categories containing over 6000 images that are suitable for the few-shot counting task. The images are annotated with two types of annotation, dots and bounding boxes, and they can be used for developing few-shot counting models. Experiments on this dataset shows that our method outperforms several state-of-the-art object detectors and few-shot counting approaches. Our code and dataset can be found at https://github.com/cvlab-stonybrook/LearningToCountEverything.

Temporal correspondence -- linking pixels or objects across frames -- is a fundamental supervisory signal for the video models. For the panoptic understanding of dynamic scenes, we further extend this concept to every segment. Specifically, we aim to learn coarse segment-level matching and fine pixel-level matching together. We implement this idea by designing two novel learning objectives. To validate our proposals, we adopt a deep siamese model and train the model to learn the temporal correspondence on two different levels (i.e., segment and pixel) along with the target task. At inference time, the model processes each frame independently without any extra computation and post-processing. We show that our per-frame inference model can achieve new state-of-the-art results on Cityscapes-VPS and VIPER datasets. Moreover, due to its high efficiency, the model runs in a fraction of time (3x) compared to the previous state-of-the-art approach. The codes and models will be released.

Non-prior face modeling aims to reconstruct 3D face only from images without shape assumptions. While plausible facial details are predicted, the models tend to over-depend on local color appearance and suffer from ambiguous noise. To address such problem, this paper presents a novel Learning to Aggregate and Personalize (LAP) framework for unsupervised robust 3D face modeling. Instead of using controlled environment, the proposed method implicitly disentangles ID-consistent and scene-specific face from unconstrained photo set. Specifically, to learn ID-consistent face, LAP adaptively aggregates intrinsic face factors of an identity based on a novel curriculum learning approach with relaxed consistency loss. To adapt the face for a personalized scene, we propose a novel attribute-refining network to modify ID-consistent face with target attribute and details. Based on the proposed method, we make unsupervised 3D face modeling benefit from meaningful image facial structure and possibly higher resolutions. Extensive experiments on benchmarks show LAP recovers superior or competitive face shape and texture, compared with state-of-the-art (SOTA) methods with or without prior and supervision.

Snapshot hyperspectral imaging has been developed to capture the spectral information of dynamic scenes. In this paper, we propose a deep neural network by learning the tensor low-rank prior of hyperspectral images (HSI) in the feature domain to promote the reconstruction quality. Our method is inspired by the canonical-polyadic (CP) decomposition theory, where a low-rank tensor can be expressed as a weight summation of several rank-1 component tensors. Specifically, we first learn the tensor low-rank prior of the image features with two steps: (a) we generate rank-1 tensors with discriminative components to collect the contextual information from both spatial and channel dimensions of the image features; (b) we aggregate those rank-1 tensors into a low-rank tensor as a 3D attention map to exploit the global correlation and refine the image features. Then, we integrate the learned tensor low-rank prior into an iterative optimization algorithm to obtain an end-to-end HSI reconstruction. Experiments on both synthetic and real data demonstrate the superiority of our method.

Single image low light enhancement is an important task and it has many practical applications. Most existing methods adopt a single image approach. Although their performance is satisfying on a static single image, we found, however, they suffer serious temporal instability when handling low light videos. We notice the problem is because existing data-driven methods are trained from single image pairs where no temporal information is available. Unfortunately, training from real temporally consistent data is also problematic because it is impossible to collect pixel-wisely paired low and normal light videos under controlled environments in large scale and diversities with noise of identical statistics. In this paper, we propose a novel method to enforce the temporal stability in low light video enhancement with only static images. The key idea is to learn and infer motion field (optical flow) from a single image and synthesize short range video sequences. Our strategy is general and can extend to large scale datasets directly. Based on this idea, we propose our method which can infer motion prior for single image low light video enhancement and enforce temporal consistency. Rigorous experiments and user study demonstrate the state-of-the-art performance of our proposed method. Our code and model will be publicly available at https://github.com/zkawfanx/StableLLVE.

Data-free learning for student networks is a new paradigm for solving users' anxiety caused by the privacy problem of using original training data. Since the architectures of modern convolutional neural networks (CNNs) are compact and sophisticated, the alternative images or meta-data generated from the teacher network are often broken. Thus, the student network cannot achieve the comparable performance to that of the pre-trained teacher network especially on the large-scale image dataset. Different to previous works, we present to maximally utilize the massive available unlabeled data in the wild. Specifically, we first thoroughly analyze the output differences between teacher and student network on the original data and develop a data collection method. Then, a noisy knowledge distillation algorithm is proposed for achieving the performance of the student network. In practice, an adaptation matrix is learned with the student network for correcting the label noise produced by the teacher network on the collected unlabeled images. The effectiveness of our DFND (Data-Free Noisy Distillation) method is then verified on several benchmarks to demonstrate its superiority over state-of-the-art data-free distillation methods. Experiments on various datasets demonstrate that the student networks learned by the proposed method can achieve comparable performance with those using the original dataset. Code is available at https://github.com/huawei-noah/Data-Efficient-Model-Compression

Existing semantic segmentation works mainly focus on learning the contextual information in high-level semantic features with CNNs. In order to maintain a precise boundary, low-level texture features are directly skip-connected into the deeper layers. Nevertheless, texture features are not only about local structure, but also include global statistical knowledge of the input image. In this paper, we fully take advantages of the low-level texture features and propose a novel Statistical Texture Learning Network (STLNet) for semantic segmentation. For the first time, STLNet analyzes the distribution of low level information and efficiently utilizes them for the task. Specifically, a novel Quantization and Counting Operator (QCO) is designed to describe the texture information in a statistical manner. Based on QCO, two modules are introduced: (1) Texture Enhance Module (TEM), to capture texture-related information and enhance the texture details; (2) Pyramid Texture Feature Extraction Module (PTFEM), to effectively extract the statistical texture features from multiple scales. Through extensive experiments, we show that the proposed STLNet achieves state-of-the-art performance on three semantic segmentation benchmarks: Cityscapes, PASCAL Context and ADE20K.

The classical maximum a-posteriori (MAP) framework for non-blind image deblurring requires defining suitable data and regularization terms, whose interplay yields the desired clear image through optimization. The vast majority of prior work focuses on advancing one of these two crucial ingredients, while keeping the other one standard. Considering the indispensable roles and interplay of both data and regularization terms, we propose a simple and effective approach to jointly learn these two terms, embedding deep neural networks within the constraints of the MAP framework, trained in an end-to-end manner. The neural networks not only yield suitable image-adaptive features for both terms, but actually predict per-pixel spatially-variant features instead of the commonly used spatially-uniform ones. The resulting spatially-variant data and regularization terms particularly improve the restoration of fine-scale structures and detail. Quantitative and qualitative results underline the effectiveness of our approach, substantially outperforming the current state of the art.

Recent advances in deep learning have demonstrated excellent results for Facial Attribute Recognition (FAR), typically trained with large-scale labeled data. However, in many real-world FAR applications, only limited labeled data are available, leading to remarkable deterioration in performance for most existing deep learning-based FAR methods. To address this problem, here we propose a method termed Spatial-Semantic Patch Learning (SSPL). The training of SSPL involves two stages. First, three auxiliary tasks, consisting of a Patch Rotation Task (PRT), a Patch Segmentation Task (PST), and a Patch Classification Task (PCT), are jointly developed to learn the spatial-semantic relationship from large-scale unlabeled facial data. We thus obtain a powerful pre-trained model. In particular, PRT exploits the spatial information of facial images in a self-supervised learning manner. PST and PCT respectively capture the pixel-level and image-level semantic information of facial images based on a facial parsing model. Second, the spatial-semantic knowledge learned from auxiliary tasks is transferred to the FAR task. By doing so, it enables that only a limited number of labeled data are required to fine-tune the pre-trained model. We achieve superior performance compared with state-of-the-art methods, as substantiated by extensive experiments and studies.

We propose an architecture and training scheme to predict video frames by explicitly modeling dis-occlusions and capturing the evolution of semantically consistent regions in the video. The scene layout (semantic map) and motion (optical flow) are decomposed into layers, which are predicted and fused with their context to generate future layouts and motions. The appearance of the scene is warped from past frames using the predicted motion in co-visible regions; dis-occluded regions are synthesized with content-aware inpainting utilizing the predicted scene layout. The result is a predictive model that explicitly represents objects and learns their class-specific motion, which we evaluate on video prediction benchmarks.

Human pose transfer has received great attention due to its wide applications, yet is still a challenging task that is not well solved. Recent works have achieved great success to transfer the person image from the source to the target pose. However, most of them cannot well capture the semantic appearance, resulting in inconsistent and less realistic textures on the reconstructed results. To address this issue, we propose a new two-stage framework to handle the pose and appearance translation. In the first stage, we predict the target semantic parsing maps to eliminate the difficulties of pose transfer and further benefit the latter translation of per-region appearance style. In the second one, with the predicted target semantic maps, we suggest a new person image generation method by incorporating the region-adaptive normalization, in which it takes the per-region styles to guide the target appearance generation. Extensive experiments show that our proposed SPGNet can generate more semantic, consistent, and photo-realistic results and perform favorably against the state of the art methods in terms of quantitative and qualitative evaluation. The source code and model are available at https://github.com/cszy98/SPGNet.git.

Existing color-guided depth super-resolution (DSR) approaches require paired RGB-D data as training examples where the RGB image is used as structural guidance to recover the degraded depth map due to their geometrical similarity. However, the paired data may be limited or expensive to be collected in actual testing environment. Therefore, we explore for the first time to learn the cross-modal knowledge at training stage, where both RGB and depth modalities are available, but test on the target dataset, where only single depth modality exists. Our key idea is to distill the knowledge of scene structural guidance from color modality to the single DSR task without changing its network architecture. Specifically, we propose an auxiliary depth estimation (DE) task that takes color image as input to estimate a depth map, and train both DSR task and DE task collaboratively to boost the performance of DSR. A cross-task distillation module is designed to realize bilateral cross-task knowledge transfer. Moreover, to address the problem of RGB-D structure inconsistency and boost the structure perception, we advance a structure prediction (SP) task that provides extra structure regularization to help both DSR and DE networks learn more informative structure representations for depth recovery. Extensive experiments demonstrate that our scheme achieves superior performance in comparison with other DSR methods.

We propose a novel joint lossy image and residual compression framework for learning l_infinity-constrained near-lossless image compression. Specifically, we obtain a lossy reconstruction of the raw image through lossy image compression and uniformly quantize the corresponding residual to satisfy a given tight l_infinity error bound. Suppose that the error bound is zero, i.e., lossless image compression, we formulate the joint optimization problem of compressing both the lossy image and the original residual in terms of variational auto-encoders and solve it with end-to-end training. To achieve scalable compression with the error bound larger than zero, we derive the probability model of the quantized residual by quantizing the learned probability model of the original residual, instead of training multiple networks. We further correct the bias of the derived probability model caused by the context mismatch between training and inference. Finally, the quantized residual is encoded according to the bias-corrected probability model and is concatenated with the bitstream of the compressed lossy image. Experimental results demonstrate that our near-lossless codec achieves the state-of-the-art performance for lossless and near-lossless image compression, and achieves competitive PSNR while much smaller l_infinity error compared with lossy image codecs at high bit rates.

Temporal action localization is an important yet challenging task in video understanding. Typically, such a task aims at inferring both the action category and localization of the start and end frame for each action instance in a long, untrimmed video. While most current models achieve good results by using pre-defined anchors and numerous actionness, such methods could be bothered with both large number of outputs and heavy tuning of locations and sizes corresponding to different anchors. Instead, anchor-free methods is lighter, getting rid of redundant hyper-parameters, but gains few attention. In this paper, we propose the first purely anchor-free temporal localization method, which is both efficient and effective. Our model includes (i) an end-to-end trainable basic predictor, (ii) a saliency-based refinement module to gather more valuable boundary features for each proposal with a novel boundary pooling, and (iii) several consistency constraints to make sure our model can find the accurate boundary given arbitrary proposals. Extensive experiments show that our method beats all anchor-based and actionness-guided methods with a remarkable margin on THUMOS14, achieving state-of-the-art results, and comparable ones on ActivityNet v1.3. Our code will be made available upon publication.

Generative models for 3D point clouds are extremely important for scene/object reconstruction applications in autonomous driving and robotics. Despite recent success of deep learning-based representation learning, it remains a great challenge for deep neural networks to synthesize or reconstruct high-fidelity point clouds, because of the difficulties in 1) learning effective pointwise representations; and 2) generating realistic point clouds from complex distributions. In this paper, we devise a dual-generators framework for point cloud generation, which generalizes vanilla generative adversarial learning framework in a progressive manner. Specifically, the first generator aims to learn effective point embeddings in a breadth-first manner, while the second generator is used to refine the generated point cloud based on a depth-first point embedding to generate a robust and uniform point cloud. The proposed dual-generators framework thus is able to progressively learn effective point embeddings for accurate point cloud generation. Experimental results on a variety of object categories from the most popular point cloud generation dataset, ShapeNet, demonstrate the state-of-the-art performance of the proposed method for accurate point cloud generation.

Uncertainty is the only certainty there is. Modeling data uncertainty is essential for regression, especially in unconstrained settings. Traditionally the direct regression formulation is considered and the uncertainty is modeled by modifying the output space to a certain family of probabilistic distributions. On the other hand, classification based regression and ranking based solutions are more popular in practice while the direct regression methods suffer from the limited performance. How to model the uncertainty within the present-day technologies for regression remains an open issue. In this paper, we propose to learn probabilistic ordinal embeddings which represent each data as a multivariate Gaussian distribution rather than a deterministic point in the latent space. An ordinal distribution constraint is proposed to exploit the ordinal nature of regression. Our probabilistic ordinal embeddings can be integrated into popular regression approaches and empower them with the ability of uncertainty estimation. Experimental results show that our approach achieves competitive performance. Code is available at https://github.com/Li-Wanhua/POEs.