CVPR 2019 Papers

In this paper, we are interested in the human pose estimation problem with a focus on learning reliable high-resolution representations. Most existing methods recover high-resolution representations from low-resolution representations produced by a high-to-low resolution network. Instead, our proposed network maintains high-resolution representations through the whole process. We start from a high-resolution subnetwork as the first stage, gradually add high-to-low resolution subnetworks one by one to form more stages, and connect the mutli-resolution subnetworks in parallel. We conduct repeated multi-scale fusions such that each of the high-to-low resolution representations receives information from other parallel representations over and over, leading to rich high-resolution representations. As a result, the predicted keypoint heatmap is potentially more accurate and spatially more precise. We empirically demonstrate the effectiveness of our network through the superior pose estimation results over two benchmark datasets: the COCO keypoint detection dataset and the MPII Human Pose dataset. In addition, we show the superiority of our network in pose tracking on the PoseTrack dataset. The code and models have been publicly available at https://github.com/leoxiaobin/deep-high-resolution-net.pytorch.

Recently, global covariance pooling (GCP) has shown great advance in improving classification performance of deep convolutional neural networks (CNNs). However, existing deep GCP networks compute covariance pooling of convolutional activations with assumption that activations are sampled from Gaussian distributions, which may not hold in practice and fails to fully characterize the statistics of activations. To handle this issue, this paper proposes a novel deep global generalized Gaussian network (3G-Net), whose core is to estimate a global covariance of generalized Gaussian for modeling the last convolutional activations. Compared with GCP in Gaussian setting, our 3G-Net assumes the distribution of activations follows a generalized Gaussian, which can capture more precise characteristics of activations. However, there exists no analytic solution for parameter estimation of generalized Gaussian, making our 3G-Net challenging. To this end, we first present a novel regularized maximum likelihood estimator for robust estimating covariance of generalized Gaussian, which can be optimized by a modified iterative re-weighted method. Then, to efficiently estimate the covariance of generaized Gaussian under deep CNN architectures, we approximate this re-weighted method by developing an unrolling re-weighted module and a square root covariance layer. In this way, 3GNet can be flexibly trained in an end-to-end manner. The experiments are conducted on large-scale ImageNet-1K and Places365 datasets, and the results demonstrate our 3G-Net outperforms its counterparts while achieving very competitive performance to state-of-the-arts.

The reconstruction of a discrete surface from a point cloud is a fundamental geometry processing problem that has been studied for decades, with many methods developed. We propose the use of a deep neural network as a geometric prior for surface reconstruction. Specifically, we overfit a neural network representing a local chart parameterization to part of an input point cloud using the Wasserstein distance as a measure of approximation. By jointly fitting many such networks to overlapping parts of the point cloud, while enforcing a consistency condition, we compute a manifold atlas. By sampling this atlas, we can produce a dense reconstruction of the surface approximating the input cloud. The entire procedure does not require any training data or explicit regularization, yet, we show that it is able to perform remarkably well: not introducing typical overfitting artifacts, and approximating sharp features closely at the same time. We experimentally show that this geometric prior produces good results for both man-made objects containing sharp features and smoother organic objects, as well as noisy inputs. We compare our method with a number of well-known reconstruction methods on a standard surface reconstruction benchmark.

Video inpainting, which aims at filling in missing regions in a video, remains challenging due to the difficulty of preserving the precise spatial and temporal coherence of video contents. In this work we propose a novel flow-guided video inpainting approach. Rather than filling in the RGB pixels of each frame directly, we consider the video inpainting as a pixel propagation problem. We first synthesize a spatially and temporally coherent optical flow field across video frames using a newly designed Deep Flow Completion network, then use the synthesized flow fields to guide the propagation of pixels to fill up the missing regions in the video. Specifically, the Deep Flow Competion network follows a coarse-to-fine refinement strategy to complete the flow fields, while their quality is further improved by hard flow example mining. Following the guide of the completed flow fields, the missing video regions can be filled up precisely. Our method is evaluated on DAVIS and YouTubeVOS datasets qualitatively and quantitatively, achieving the state-of-the-art performance in terms of inpainting quality and speed.

In this paper, we propose to learn a deep fitting degree scoring network for monocular 3D object detection, which aims to score fitting degree between proposals and object conclusively. Different from most existing monocular frameworks which use tight constraint to get 3D location, our approach achieves high-precision localization through measuring the visual fitting degree between the projected 3D proposals and the object. We first regress the dimension and orientation of the object using an anchor-based method so that a suitable 3D proposal can be constructed. We propose FQNet, which can infer the 3D IoU between the 3D proposals and the object solely based on 2D cues. Therefore, during the detection process, we sample a large number of candidates in the 3D space and project these 3D bounding boxes on 2D image individually. The best candidate can be picked out by simply exploring the spatial overlap between proposals and the object, in the form of the output 3D IoU score of FQNet. Experiments on the KITTI dataset demonstrate the effectiveness of our framework.

This paper presents the first end-to-end network for exemplar-based video colorization. The main challenge is to achieve temporal consistency while remaining faithful to the reference style. To address this issue, we introduce a recurrent framework that unifies the semantic correspondence and color propagation steps. Both steps allow a provided reference image to guide the colorization of every frame, thus reducing accumulated propagation errors. Video frames are colorized in sequence based on the colorization history, and its coherency is further enforced by the temporal consistency loss. All of these components, learned end-to-end, help produce realistic videos with good temporal stability. Experiments show our result is superior to the state-of-the-art methods both quantitatively and qualitatively.

Deep embedding learning aims to learn a distance metric for effective similarity measurement, which has achieved promising performance in various tasks. As the vast majority of training samples produce gradients with magnitudes close to zero, hard example mining is usually employed to improve the effectiveness and efficiency of the training procedure. However, most existing sampling methods are designed by hand, which ignores the dependence between examples and suffer from exhaustive searching. In this paper, we propose a deep embedding with discriminative sampling policy (DE-DSP) learning framework by simultaneously training two models: a deep sampler network that learns effective sampling strategies, and a feature embedding that maps samples to the feature space. Rather than exhaustively calculating the hardness of all the examples for mining through forward-propagation, the deep sampler network exploits the strong prior of relations among samples to learn discriminative sampling policy in an more efficient manner. Experimental results demonstrate faster convergence and stronger discriminative power of our DE-DSP framework under different embedding objectives.

Egocentric interaction recognition aims to recognize the camera wearer's interactions with the interactor who faces the camera wearer in egocentric videos. In such a human-human interaction analysis problem, it is crucial to explore the relations between the camera wearer and the interactor. However, most existing works directly model the interactions as a whole and lack modeling the relations between the two interacting persons. To exploit the strong relations for egocentric interaction recognition, we introduce a dual relation modeling framework which learns to model the relations between the camera wearer and the interactor based on the individual action representations of the two persons. Specifically, we develop a novel interactive LSTM module, the key component of our framework, to explicitly model the relations between the two interacting persons based on their individual action representations, which are collaboratively learned with an interactor attention module and a global-local motion module. Experimental results on three egocentric interaction datasets show the effectiveness of our method and advantage over state-of-the-arts.

In this paper, we propose the first end-to-end convolutional neural network (CNN) architecture, Defocus Map Estimation Network (DMENet), for spatially varying defocus map estimation. To train the network, we produce a novel depth-of-field (DOF) dataset, SYNDOF, where each image is synthetically blurred with a ground-truth depth map. Due to the synthetic nature of SYNDOF, the feature characteristics of images in SYNDOF can differ from those of real defocused photos. To address this gap, we use domain adaptation that transfers the features of real defocused photos into those of synthetically blurred ones. Our DMENet consists of four subnetworks: blur estimation, domain adaptation, content preservation, and sharpness calibration networks. The subnetworks are connected to each other and jointly trained with their corresponding supervisions in an end-to-end manner. Our method is evaluated on publicly available blur detection and blur estimation datasets and the results show the state-of-the-art performance.In this paper, we propose the first end-to-end convolutional neural network (CNN) architecture, Defocus Map Estimation Network (DMENet), for spatially varying defocus map estimation. To train the network, we produce a novel depth-of-field (DOF) dataset, SYNDOF, where each image is synthetically blurred with a ground-truth depth map. Due to the synthetic nature of SYNDOF, the feature characteristics of images in SYNDOF can differ from those of real defocused photos. To address this gap, we use domain adaptation that transfers the features of real defocused photos into those of synthetically blurred ones. Our DMENet consists of four subnetworks: blur estimation, domain adaptation, content preservation, and sharpness calibration networks. The subnetworks are connected to each other and jointly trained with their corresponding supervisions in an end-to-end manner. Our method is evaluated on publicly available blur detection and blur estimation datasets and the results show the state-of-the-art performance.

ChArUco boards are used for camera calibration, monocular pose estimation, and pose verification in both robotics and augmented reality. Such fiducials are detectable via traditional computer vision methods (as found in OpenCV) in well-lit environments, but classical methods fail when the lighting is poor or when the image undergoes extreme motion blur. We present Deep ChArUco, a real-time pose estimation system which combines two custom deep networks, ChArUcoNet and RefineNet, with the Perspective-n-Point (PnP) algorithm to estimate the marker's 6DoF pose. ChArUcoNet is a two-headed marker-specific convolutional neural network (CNN) which jointly outputs ID-specific classifiers and 2D point locations. The 2D point locations are further refined into subpixel coordinates using RefineNet. Our networks are trained using a combination of auto-labeled videos of the target marker, synthetic subpixel corner data, and extreme data augmentation. We evaluate Deep ChArUco in challenging low-light, high-motion, high-blur scenarios and demonstrate that our approach is superior to a traditional OpenCV-based method for ChArUco marker detection and pose estimation.

Blind video decaptioning is a problem of automatically removing text overlays and inpainting the occluded parts in videos without any input masks. While recent deep learning based inpainting methods deal with a single image and mostly assume that the positions of the corrupted pixels are known, we aim at automatic text removal in video sequences without mask information. In this paper, we propose a simple yet effective framework for fast blind video decaptioning. We construct an encoder-decoder model, where the encoder takes multiple source frames that can provide visible pixels revealed from the scene dynamics. These hints are aggregated and fed into the decoder. We apply a residual connection from the input frame to the decoder output to enforce our network to focus on the corrupted regions only. Our proposed model was ranked in the first place in the ECCV Chalearn 2018 LAP Inpainting Competition Track2: Video decaptioning. In addition, we further improve this strong model by applying a recurrent feedback. The recurrent feedback not only enforces temporal coherence but also provides strong clues on where the corrupted pixels are. Both qualitative and quantitative experiments demonstrate that our full model produces accurate and temporally consistent video results in real time (50+ fps).

Learning effective distance metric between data has gained increasing popularity, for its promising performance on various tasks, such as face verification, zero-shot learning, and image retrieval. A major line of researches employs hard data mining, which makes efforts on searching a subset of significant data. However, hard data mining based approaches only rely on a small percentage of data, which is apt to overfitting. This motivates us to propose a novel framework, named deep asymmetric metric learning via rich relationship mining (DAMLRRM), to mine rich relationship under satisfying sampling size. DAMLRRM constructs two asymmetric data streams that are differently structured and of unequal length. The asymmetric structure enables the two data streams to interlace each other, which allows for the informative comparison between new data pairs over iterations. To improve the generalization ability, we further relax the constraint on the intra-class relationship. Rather than greedily connecting all possible positive pairs, DAMLRRM builds a minimum-cost spanning tree within each category to ensure the formation of a connected region. As such there exists at least one direct or indirect path between arbitrary positive pairs to bridge intra-class relevance. Extensive experimental results on three benchmark datasets including CUB-200-2011, Cars196, and Stanford Online Products show that DAMLRRM effectively boosts the performance of existing deep metric learning approaches.

There has been an increasing research interest in age-invariant face recognition. However, matching faces with big age gaps remains a challenging problem, primarily due to the significant discrepancy of face appearance caused by aging. To reduce such discrepancy, in this paper we present a novel algorithm to remove age-related components from features mixed with both identity and age information. Specifically, we factorize a mixed face feature into two uncorrelated components: identity-dependent component and age-dependent component, where the identity-dependent component contains information that is useful for face recognition. To implement this idea, we propose the Decorrelated Adversarial Learning (DAL) algorithm, where a Canonical Mapping Module (CMM) is introduced to find maximum correlation of the paired features generated by the backbone network, while the backbone network and the factorization module are trained to generate features reducing the correlation. Thus, the proposed model learns the decomposed features of age and identity whose correlation is significantly reduced. Simultaneously, the identity-dependent feature and the age-dependent feature are supervised by ID and age preserving signals respectively to ensure they contain the correct information. Extensive experiments have been conducted on the popular public-domain face aging datasets (FG-NET, MORPH Album 2, and CACD-VS) to demonstrate the effectiveness of the proposed approach.

Recent semantic segmentation methods exploit encoder-decoder architectures to produce the desired pixel-wise segmentation prediction. The last layer of the decoders is typically a bilinear upsampling procedure to recover the final pixel-wise prediction. We empirically show that this oversimple and data-independent bilinear upsampling may lead to sub-optimal results. In this work, we propose a data-dependent upsampling (DUpsampling) to replace bilinear, which takes advantages of the redundancy in the label space of semantic segmentation and is able to recover the pixel-wise prediction from low-resolution outputs of CNNs. The main advantage of the new upsampling layer lies in that with a relatively lower-resolution feature map such as 1/16 or 1/32 of the input size, we can achieve even better segmentation accuracy, significantly reducing computation complexity. This is made possible by 1) the new upsampling layer's much improved reconstruction capability; and more importantly 2) the DUpsampling based decoder's flexibility in leveraging almost arbitrary combinations of the CNN encoders' features. Experiments on PASCAL VOC demonstrate that with much less computation complexity, our decoder outperforms the state-of-the-art decoder. Finally, without any post-processing, the framework equipped with our proposed decoder achieves new state-of-the-art performance on two datasets: 88.1% mIOU on PASCAL VOC with 30% computation of the previously best model; and 52.5% mIOU on PASCAL Context.

Defocus blur detection aims to detect out-of-focus regions from an image. Although attracting more and more attention due to its widespread applications, defocus blur detection still confronts several challenges such as the interference of background clutter, sensitivity to scales and missing boundary details of defocus blur regions. To deal with these issues, we propose a deep neural network which recurrently fuses and refines multi-scale deep features (DeFusionNet) for defocus blur detection. We firstly utilize a fully convolutional network to extract multi-scale deep features. The features from bottom layers are able to capture rich low-level features for details preservation, while the features from top layers can characterize the semantic information to locate blur regions. These features from different layers are fused as shallow features and semantic features, respectively. After that, the fused shallow features are propagated to top layers for refining the fine details of detected defocus blur regions, and the fused semantic features are propagated to bottom layers to assist in better locating the defocus regions. The feature fusing and refining are carried out in a recurrent manner. Also, we finally fuse the output of each layer at the last recurrent step to obtain the final defocus blur map by considering the sensitivity to scales of the defocus degree. Experiments on two commonly used defocus blur detection benchmark datasets are conducted to demonstrate the superority of DeFusionNet when compared with other 10 competitors. Code and more results can be found at: http://tangchang.net

Successful visual recognition networks benefit from aggregating information spanning from a wide range of scales. Previous research has investigated information fusion of connected layers or multiple branches in a block, seeking to strengthen the power of multi-scale representations. Despite their great successes, existing practices often allocate the neurons for each scale manually, and keep the same ratio in all aggregation blocks of an entire network, rendering suboptimal performance. In this paper, we propose to learn the neuron allocation for aggregating multi-scale information in different building blocks of a deep network. The most informative output neurons in each block are preserved while others are discarded, and thus neurons for multiple scales are competitively and adaptively allocated. Our scale aggregation network (ScaleNet) is constructed by repeating a scale aggregation (SA) block that concatenates feature maps at a wide range of scales. Feature maps for each scale are generated by a stack of downsampling, convolution and upsampling operations. The data-driven neuron allocation and SA block achieve strong representational power at the cost of considerably low computational complexity. The proposed ScaleNet, by replacing all 3x3 convolutions in ResNet with our SA blocks, achieves better performance than ResNet and its outstanding variants like ResNeXt and SE-ResNet, in the same computational complexity. On ImageNet classification, ScaleNets absolutely reduce the top-1 error rate of ResNets by 1.12 (101 layers) and 1.82 (50 layers). On COCO object detection, ScaleNets absolutely improve the mAP with backbone of ResNets by 3.6 and 4.6 on Faster-RCNN, respectively. Code and models are released on https://github.com/Eli-YiLi/ScaleNet.

Recently, graph convolutional neural networks have been widely studied for graph-structured data representation and learning. In this paper, we present Graph Diffusion-Embedding networks (GDENs), a new model for graph-structured data representation and learning. GDENs are motivated by our development of graph based feature diffusion. GDENs integrate both feature diffusion and graph node (low-dimensional) embedding simultaneously into a unified network by employing a novel diffusion-embedding architecture. GDENs have two main advantages. First, the equilibrium representation of the diffusion-embedding operation in GDENs can be obtained via a simple closed-form solution, which thus guarantees the compactivity and efficiency of GDENs. Second, the proposed GDENs can be naturally extended to address the data with multiple graph structures. Experiments on various semi-supervised learning tasks on several benchmark datasets demonstrate that the proposed GDENs significantly outperform traditional graph convolutional networks.

The goal of this paper is to detect the spatio-temporal extent of an action. The two-stream detection network based on RGB and flow provides state-of-the-art accuracy at the expense of a large model-size and heavy computation. We propose to embed RGB and optical-flow into a single two-in-one stream network with new layers. A motion condition layer extracts motion information from flow images, which is leveraged by the motion modulation layer to generate transformation parameters for modulating the low-level RGB features. The method is easily embedded in existing appearance- or two-stream action detection networks, and trained end-to-end. Experiments demonstrate that leveraging the motion condition to modulate RGB features improves detection accuracy. With only half the computation and parameters of the state-of-the-art two-stream methods, our two-in-one stream still achieves impressive results on UCF101-24, UCFSports and J-HMDB.

Recently, Convolutional Neural Network (CNN) has achieved great success in face detection. However, it remains a challenging problem for the current face detection methods owing to high degree of variability in scale, pose, occlusion, expression, appearance and illumination. In this Paper, we propose a novel detection network named Dual Shot face Detector(DSFD). which inherits the architecture of SSD and introduces a Feature Enhance Module (FEM) for transferring the original feature maps to extend the single shot detector to dual shot detector. Specially, progressive anchor loss (PAL) computed by using two set of anchors is adopted to effectively facilitate the features. Additionally, we propose an improved anchor matching (IAM) method by integrating novel data augmentation techniques and anchor design strategy in our DSFD to provide better initialization for the regressor. Extensive experiments on popular benchmarks: WIDER FACE (easy: 0.966, medium: 0.957, hard: 0.904) and FDDB ( discontinuous: 0.991, continuous: 0.862 ) demonstrate the superiority of DSFD over the state-of-the-art face detection methods (e.g., PyramidBox and SRN). Code will be made available upon publication.