CVPR 2020 Papers

Attributes and objects can compose diverse compositions. To model the compositional nature of these general concepts, it is a good choice to learn them through transformations, such as coupling and decoupling. However, complex transformations need to satisfy specific principles to guarantee the rationality. In this paper, we first propose a previously ignored principle of attribute-object transformation: Symmetry. For example, coupling peeled-apple with attribute peeled should result in peeled-apple, and decoupling peeled from apple should still output apple. Incorporating the symmetry principle, a transformation framework inspired by group theory is built, i.e. SymNet. SymNet consists of two modules, Coupling Network and Decoupling Network. With the group axioms and symmetry property as objectives, we adopt Deep Neural Networks to implement SymNet and train it in an end-to-end paradigm. Moreover, we propose a Relative Moving Distance (RMD) based recognition method to utilize the attribute change instead of the attribute pattern itself to classify attributes. Our symmetry learning can be utilized for the Compositional Zero-Shot Learning task and outperforms the state-of-the-art on widely-used benchmarks. Code is available at https://github.com/DirtyHarryLYL/SymNet.

Swapping text in scene images while preserving original fonts, colors, sizes and background textures is a challenging task due to the complex interplay between different factors. In this work, we present SwapText, a three-stage framework to transfer texts across scene images. First, a novel text swapping network is proposed to replace text labels only in the foreground image. Second, a background completion network is learned to reconstruct background images. Finally, the generated foreground image and background image are used to generate the word image by the fusion network. Using the proposing framework, we can manipulate the texts of the input images even with severe geometric distortion. Qualitative and quantitative results are presented on several scene text datasets, including regular and irregular text datasets. We conducted extensive experiments to prove the usefulness of our method such as image based text translation, text image synthesis.

Autonomous driving system development is critically dependent on the ability to replay complex and diverse traffic scenarios in simulation. In such scenarios, the ability to accurately simulate the vehicle sensors such as cameras, lidar or radar is hugely helpful. However, current sensor simulators leverage gaming engines such as Unreal or Unity, requiring manual creation of environments, objects, and material properties. Such approaches have limited scalability and fail to produce realistic approximations of camera, lidar, and radar data without significant additional work. In this paper, we present a simple yet effective approach to generate realistic scenario sensor data, based only on a limited amount of lidar and camera data collected by an autonomous vehicle. Our approach uses texture-mapped surfels to efficiently reconstruct the scene from an initial vehicle pass or set of passes, preserving rich information about object 3D geometry and appearance, as well as the scene conditions. We then leverage a SurfelGAN network to reconstruct realistic camera images for novel positions and orientations of the self-driving vehicle and moving objects in the scene. We demonstrate our approach on the Waymo Open Dataset and show that it can synthesize realistic camera data for simulated scenarios. We also create a novel dataset that contains cases in which two self-driving vehicles observe the same scene at the same time. We use this dataset to provide additional evaluation and demonstrate the usefulness of our SurfelGAN model.

Annotating a qualitative large-scale facial expression dataset is extremely difficult due to the uncertainties caused by ambiguous facial expressions, low-quality facial images, and the subjectiveness of annotators. These uncertainties suspend the progress of large-scale Facial Expression Recognition (FER) in data-driven deep learning era. To address this problelm, this paper proposes to suppress the uncertainties by a simple yet efficient Self-Cure Network (SCN). Specifically, SCN suppresses the uncertainty from two different aspects: 1) a self-attention mechanism over FER dataset to weight each sample in training with a ranking regularization, and 2) a careful relabeling mechanism to modify the labels of these samples in the lowest-ranked group. Experiments on synthetic FER datasets and our collected WebEmotion dataset validate the effectiveness of our method. Results on public benchmarks demonstrate that our SCN outperforms current state-of-the-art methods with 88.14% on RAF-DB, 60.23% on AffectNet, and 89.35% on FERPlus.

In recent years, the supervised learning strategy for real noisy image denoising has been emerging and has achieved promising results. In contrast, realistic noise removal for raw noisy videos is rarely studied due to the lack of noisy-clean pairs for dynamic scenes. Clean video frames for dynamic scenes cannot be captured with a long-exposure shutter or averaging multi-shots as was done for static images. In this paper, we solve this problem by creating motions for controllable objects, such as toys, and capturing each static moment for multiple times to generate clean video frames. In this way, we construct a dataset with 55 groups of noisy-clean videos with ISO values ranging from 1600 to 25600. To our knowledge, this is the first dynamic video dataset with noisy-clean pairs. Correspondingly, we propose a raw video denoising network (RViDeNet) by exploring the temporal, spatial, and channel correlations of video frames. Since the raw video has Bayer patterns, we pack it into four sub-sequences, i.e RGBG sequences, which are denoised by the proposed RViDeNet separately and finally fused into a clean video. In addition, our network not only outputs a raw denoising result, but also the sRGB result by going through an image signal processing (ISP) module, which enables users to generate the sRGB result with their favourite ISPs. Experimental results demonstrate that our method outperforms state-of-the-art video and raw image denoising algorithms on both indoor and outdoor videos.

In computer vision, superpixels have been widely used as an effective way to reduce the number of image primitives for subsequent processing. But only a few attempts have been made to incorporate them into deep neural networks. One main reason is that the standard convolution operation is defined on regular grids and becomes inefficient when applied to superpixels. Inspired by an initialization strategy commonly adopted by traditional superpixel algorithms, we present a novel method that employs a simple fully convolutional network to predict superpixels on a regular image grid. Experimental results on benchmark datasets show that our method achieves state-of-the-art superpixel segmentation performance while running at about 50fps. Based on the predicted superpixels, we further develop a downsampling/upsampling scheme for deep networks with the goal of generating high-resolution outputs for dense prediction tasks. Specifically, we modify a popular network architecture for stereo matching to simultaneously predict superpixels and disparities. We show that improved disparity estimation accuracy can be obtained on public datasets.

This paper introduces SuperGlue, a neural network that matches two sets of local features by jointly finding correspondences and rejecting non-matchable points. Assignments are estimated by solving a differentiable optimal transport problem, whose costs are predicted by a graph neural network. We introduce a flexible context aggregation mechanism based on attention, enabling SuperGlue to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from image pairs. SuperGlue outperforms other learned approaches and achieves state-of-the-art results on the task of pose estimation in challenging real-world indoor and outdoor environments. The proposed method performs matching in real-time on a modern GPU and can be readily integrated into modern SfM or SLAM systems. The code and trained weights are publicly available at github.com/magicleap/SuperGluePretrainedNetwork.

Image segmentation is a fundamental vision task and still remains a crucial step for many applications. In this paper, we propose a fast image segmentation method based on a novel super boundary-to-pixel direction (super-BPD) and a customized segmentation algorithm with super-BPD. Precisely, we define BPD on each pixel as a two-dimensional unit vector pointing from its nearest boundary to the pixel. In the BPD, nearby pixels from different regions have opposite directions departing from each other, and nearby pixels in the same region have directions pointing to the other or each other (i.e., around medial points). We make use of such property to partition image into super-BPDs, which are novel informative superpixels with robust direction similarity for fast grouping into segmentation regions. Extensive experimental results on BSDS500 and Pascal Context demonstrate the accuracy and efficiency of the proposed super-BPD in segmenting images. Specifically, we achieve comparable or superior performance with MCG while running at 25fps vs 0.07fps. Super-BPD also exhibits a noteworthy transferability to unseen scenes.

We propose a novel method that tracks fast moving objects, mainly non-uniform spherical, in full 6 degrees of freedom, estimating simultaneously their 3D motion trajectory, 3D pose and object appearance changes with a time step that is a fraction of the video frame exposure time. The sub-frame object localization and appearance estimation allows realistic temporal super-resolution and precise shape estimation. The method, called TbD-3D (Tracking by Deblatting in 3D) relies on a novel reconstruction algorithm which solves a piece-wise deblurring and matting problem. The 3D rotation is estimated by minimizing the reprojection error. As a second contribution, we present a new challenging dataset with fast moving objects that change their appearance and distance to the camera. High-speed camera recordings with zero lag between frame exposures were used to generate videos with different frame rates annotated with ground-truth trajectory and pose.

Exemplar-based colorization aims to add colors to a grayscale image guided by a content related reference im- age. Existing methods are either sensitive to the selection of reference images (content, position) or extremely time and resource consuming, which limits their practical applica- tion. To tackle these problems, we propose a deep exemplar colorization architecture inspired by the characteristics of stylization in feature extracting and blending. Our coarse- to-fine architecture consists of two parts: a fast transfer sub-net and a robust colorization sub-net. The transfer sub- net obtains a coarse chrominance map via matching basic feature statistics of the input pairs in a progressive way. The colorization sub-net refines the map to generate the final re- sults. The proposed end-to-end network can jointly learn faithful colorization with a related reference and plausible color prediction with unrelated reference. Extensive exper- imental validation demonstrates that our approach outper- forms the state-of-the-art methods in less time whether in exemplar-based colorization or image stylization tasks.

StyleGAN generates photorealistic portrait images of faces with eyes, teeth, hair and context (neck, shoulders, background), but lacks a rig-like control over semantic face parameters that are interpretable in 3D, such as face pose, expressions, and scene illumination. Three-dimensional morphable face models (3DMMs) on the other hand offer control over the semantic parameters, but lack photorealism when rendered and only model the face interior, not other parts of a portrait image (hair, mouth interior, background). We present the first method to provide a face rig-like control over a pretrained and fixed StyleGAN via a 3DMM. A new rigging network, RigNet is trained between the 3DMM's semantic parameters and StyleGAN's input. The network is trained in a self-supervised manner, without the need for manual annotations. At test time, our method generates portrait images with the photorealism of StyleGAN and provides explicit control over the 3D semantic parameters of the face.

Existing fully-supervised person re-identification (ReID) methods usually suffer from poor generalization capability caused by domain gaps. The key to solving this problem lies in filtering out identity-irrelevant interference and learning domain-invariant person representations. In this paper, we aim to design a generalizable person ReID framework which trains a model on source domains yet is able to generalize/perform well on target domains. To achieve this goal, we propose a simple yet effective Style Normalization and Restitution (SNR) module. Specifically, we filter out style variations (e.g., illumination, color contrast) by Instance Normalization (IN). However, such a process inevitably removes discriminative information. We propose to distill identity-relevant feature from the removed information and restitute it to the network to ensure high discrimination. For better disentanglement, we enforce a dual causal loss constraint in SNR to encourage the separation of identity-relevant features and identity-irrelevant features. Extensive experiments demonstrate the strong generalization capability of our framework. Our models empowered by the SNR modules significantly outperform the state-of-the-art domain generalization approaches on multiple widely-used person ReID benchmarks, and also show superiority on unsupervised domain adaptation.

Despite the success of deep neural networks (DNNs), state-of-the-art models are too large to deploy on low-resource devices or common server configurations in which multiple models are held in memory. Model compression methods address this limitation by reducing the memory footprint, latency, or energy consumption of a model with minimal impact on accuracy. We focus on the task of reducing the number of learnable variables in the model. In this work we combine ideas from weight hashing and dimensionality reductions resulting in a simple and powerful structured multi-hashing method based on matrix products that allows direct control of model size of any deep network and is trained end-to-end. We demonstrate the strength of our approach by compressing models from the ResNet, EfficientNet, and MobileNet architecture families. Our method allows us to drastically decrease the number of variables while maintaining high accuracy. For instance, by applying our approach to EfficentNet-B4 (16M parameters) we reduce it to the size of B0 (5M parameters), while gaining over 3% in accuracy over B0 baseline. On the commonly used benchmark CIFAR10 we reduce the ResNet32 model by 75% with no loss in quality, and are able to do a 10x compression while still achieving above 90% accuracy.

Model compression techniques, such as pruning and quantization, are becoming increasingly important to reduce the memory footprints and the amount of computations. Despite model size reduction, achieving performance enhancement on devices is, however, still challenging mainly due to the irregular representations of sparse matrix formats. This paper proposes a new weight representation scheme for Sparse Quantized Neural Networks, specifically achieved by fine-grained and unstructured pruning method. The representation is encrypted in a structured regular format, which can be efficiently decoded through XOR-gate network during inference in a parallel manner. We demonstrate various deep learning models that can be compressed and represented by our proposed format with fixed and high compression ratio. For example, for fully-connected layers of AlexNet on ImageNet dataset, we can represent the sparse weights by only 0.28 bits/weight for 1-bit quantization and 91% pruning rate with a fixed decoding rate and full memory bandwidth usage. Decoding through XOR-gate network can be performed without any model accuracy degradation with additional patch data associated with small overhead.

Structures matter in single image super resolution (SISR). Recent studies benefiting from generative adversarial network (GAN) have promoted the development of SISR by recovering photo-realistic images. However, there are always undesired structural distortions in the recovered images. In this paper, we propose a structure-preserving super resolution method to alleviate the above issue while maintaining the merits of GAN-based methods to generate perceptual-pleasant details. Specifically, we exploit gradient maps of images to guide the recovery in two aspects. On the one hand, we restore high-resolution gradient maps by a gradient branch to provide additional structure priors for the SR process. On the other hand, we propose a gradient loss which imposes a second-order restriction on the super-resolved images. Along with the previous image-space loss functions, the gradient-space objectives help generative networks concentrate more on geometric structures. Moreover, our method is model-agnostic, which can be potentially used for off-the-shelf SR networks. Experimental results show that we achieve the best PI and LPIPS performance and meanwhile comparable PSNR and SSIM compared with state-of-the-art perceptual-driven SR methods. Visual results demonstrate our superiority in restoring structures while generating natural SR images.

Single image depth prediction is a challenging task due to its ill-posed nature and challenges with capturing ground truth for supervision. Large-scale disparity data generated from stereo photos and 3D videos is a promising source of supervision, however, such disparity data can only approximate the inverse ground truth depth up to an affine transformation. To more effectively learn from such pseudo-depth data, we propose to use a simple pair-wise ranking loss with a novel sampling strategy. Instead of randomly sampling point pairs, we guide the sampling to better characterize structure of important regions based on the low-level edge maps and high-level object instance masks. We show that the pair-wise ranking loss, combined with our structure-guided sampling strategies, can significantly improve the quality of depth map prediction. In addition, we introduce a new relative depth dataset of about 21K diverse high-resolution web stereo photos to enhance the generalization ability of our model. In experiments, we conduct cross-dataset evaluation on six benchmark datasets and show that our method consistently improves over the baselines, leading to superior quantitative and qualitative results.

Unsupervised domain adaptation (UDA) casts a light when dealing with insufficient or no labeled data in the target domain by exploring the well-annotated source knowledge in different distributions. Most research efforts on UDA explore to seek a domain-invariant classifier over source supervision. However, due to the scarcity of label information in the target domain, such a classifier has a lack of ground-truth target supervision, which dramatically obstructs the robustness and discrimination of the classifier. To this end, we develop a novel Generative cross-domain learning via Structure-Preserving (GSP), which attempts to transform target data into the source domain in order to take advantage of source supervision. Specifically, a novel cross-domain graph alignment is developed to capture the intrinsic relationship across two domains during target-source translation. Simultaneously, two distinct classifiers are trained to trigger the domain-invariant feature learning both guided with source supervision, one is a traditional source classifier and the other is a source-supervised target classifier. Extensive experimental results on several cross-domain visual benchmarks have demonstrated the effectiveness of our model by comparing with other state-of-the-art UDA algorithms.

In this paper, we propose a novel image segmentation method to tackle two critical problems of medical image, which are (i) ambiguity of structure boundary in the medical image domain and (ii) uncertainty of the segmented region without specialized domain knowledge. To solve those two problems in automatic medical segmentation, we propose a novel structure boundary preserving segmentation framework. To this end, the boundary key point selection algorithm is proposed. In the proposed algorithm, the key points on the structural boundary of the target object are estimated. Then, a boundary preserving block (BPB) with the boundary key point map is applied for predicting the structure boundary of the target object. Further, for embedding experts' knowledge in the fully automatic segmentation, we propose a novel shape boundary-aware evaluator (SBE) with the ground-truth structure information indicated by experts. The proposed SBE could give feedback to the segmentation network based on the structure boundary key point. The proposed method is general and flexible enough to be built on top of any deep learning-based segmentation network. We demonstrate that the proposed method could surpass the state-of-the-art segmentation network and improve the accuracy of three different segmentation network models on different types of medical image datasets.

3D object detection from point cloud data plays an essential role in autonomous driving. Current single-stage detectors are efficient by progressively downscaling the 3D point clouds in a fully convolutional manner. However, the downscaled features inevitably lose spatial information and cannot make full use of the structure information of 3D point cloud, degrading their localization precision. In this work, we propose to improve the localization precision of single-stage detectors by explicitly leveraging the structure information of 3D point cloud. Specifically, we design an auxiliary network which converts the convolutional features in the backbone network back to point-level representations. The auxiliary network is jointly optimized, by two point-level supervisions, to guide the convolutional features in the backbone network to be aware of the object structure. The auxiliary network can be detached after training and therefore introduces no extra computation in the inference stage. Besides, considering that single-stage detectors suffer from the discordance between the predicted bounding boxes and corresponding classification confidences, we develop an efficient part-sensitive warping operation to align the confidences to the predicted bounding boxes. Our proposed detector ranks at the top of KITTI 3D/BEV detection leaderboards and runs at 25 FPS for inference.