CVPR 2019 Papers

We propose sphere generative adversarial network (GAN), a novel integral probability metric (IPM)-based GAN. Sphere GAN uses the hypersphere to bound IPMs in the objective function. Thus, it can be trained stably. On the hypersphere, sphere GAN exploits the information of higher-order statistics of data using geometric moment matching, thereby providing more accurate results. In the paper, we mathematically prove the good properties of sphere GAN. In experiments, sphere GAN quantitatively and qualitatively surpasses recent state-of-the-art GANs for unsupervised image generation problems with the CIFAR-10, STL-10, and LSUN bedroom datasets. Source code is available at https://github.com/pswkiki/SphereGAN.

As intelligent systems increasingly blend into our everyday life, artificial social intelligence becomes a prominent area of research. Intelligent systems must be socially intelligent in order to comprehend human intents and maintain a rich level of interaction with humans. Human language offers a unique unconstrained approach to probe through questions and reason through answers about social situations. This unconstrained approach extends previous attempts to model social intelligence through numeric supervision (e.g. sentiment and emotions labels). In this paper, we introduce Social-IQ, a unconstrained benchmark specifically designed to train and evaluate socially intelligent technologies. By providing a rich source of open-ended questions and answers, Social-IQ opens the door to explainable social intelligence. The dataset contains rigorously annotated and validated videos, questions and answers, as well as annotations for the complexity level of each question and answer. Social-IQ contains 1,250 natural in-the-wild social situations, 7,500 questions and 52,500 correct and incorrect answers. Although humans can reason about social situations with very high accuracy (95.08%), existing state-of-the-art computational models struggle on this task. As a result, Social-IQ brings novel challenges that will spark future research in social intelligence modeling, visual reasoning, and multimodal question answering (QA).

Several tasks in machine learning are evaluated using non-differentiable metrics such as mean average precision or Spearman correlation. However, their non-differentiability prevents from using them as objective functions in a learning framework. Surrogate and relaxation methods exist but tend to be specific to a given metric. In the present work, we introduce a new method to learn approximations of such non-differentiable objective functions. Our approach is based on a deep architecture that approximates the sorting of arbitrary sets of scores. It is trained virtually for free using synthetic data. This sorting deep (SoDeep) net can then be combined in a plug-and-play manner with existing deep architectures. We demonstrate the interest of our approach in three different tasks that require ranking: Cross-modal text-image retrieval, multi-label image classification and visual memorability ranking. Our approach yields very competitive results on these three tasks, which validates the merit and the flexibility of SoDeep as a proxy for sorting operation in ranking-based losses.

DensePose supersedes traditional landmark detectors by densely mapping image pixels to body surface coordinates. This power, however, comes at a greatly increased annotation cost, as supervising the model requires to manually label hundreds of points per pose instance. In this work, we thus seek methods to significantly slim down the DensePose annotations, proposing more efficient data collection strategies. In particular, we demonstrate that if annotations are collected in video frames, their efficacy can be multiplied for free by using motion cues. To explore this idea, we introduce DensePose-Track, a dataset of videos where selected frames are annotated in the traditional DensePose manner. Then, building on geometric properties of the DensePose mapping, we use the video dynamic to propagate ground-truth annotations in time as well as to learn from Siamese equivariance constraints. Having performed exhaustive empirical evaluation of various data annotation and learning strategies, we demonstrate that doing so can deliver significantly improved pose estimation results over strong baselines. However, despite what is suggested by some recent works, we show that merely synthesizing motion patterns by applying geometric transformations to isolated frames is significantly less effective, and that motion cues help much more when they are extracted from videos.

Local windows are routinely used in computer vision and almost without exception the center of the window is aligned with the pixels being processed. We show that this conventional wisdom is not universally applicable. When a pixel is on an edge, placing the center of the window on the pixel is one of the fundamental reasons that cause many filtering algorithms to blur the edges. Based on this insight, we propose a new Side Window Filtering (SWF) technique which aligns the window's side or corner with the pixel being processed. The SWF technique is surprisingly simple yet theoretically rooted and very effective in practice. We show that many traditional linear and nonlinear filters can be easily implemented under the SWF framework. Extensive analysis and experiments show that implementing the SWF principle can significantly improve their edge preserving capabilities and achieve state of the art performances in applications such as image smoothing, denoising, enhancement, structure-preserving texture-removing, mutual-structure extraction, and HDR tone mapping. In addition to image filtering, we further show that the SWF principle can be extended to other applications involving the use of a local window. Using colorization by optimization as an example, we demonstrate that implementing the SWF principle can effectively prevent artifacts such as color leakage associated with the conventional implementation. Given the ubiquity of window based operations in computer vision, the new SWF technique is likely to benefit many more applications.

Siamese network based trackers formulate tracking as convolutional feature cross-correlation between target template and searching region. However, Siamese trackers still have accuracy gap compared with state-of-the-art algorithms and they cannot take advantage of feature from deep networks, such as ResNet-50 or deeper. In this work we prove the core reason comes from the lack of strict translation invariance. By comprehensive theoretical analysis and experimental validations, we break this restriction through a simple yet effective spatial aware sampling strategy and successfully train a ResNet-driven Siamese tracker with significant performance gain. Moreover, we propose a new model architecture to perform depth-wise and layer-wise aggregations, which not only further improves the accuracy but also reduces the model size. We conduct extensive ablation studies to demonstrate the effectiveness of the proposed tracker, which obtains currently the best results on four large tracking benchmarks, including OTB2015, VOT2018, UAV123, and LaSOT. Our model will be released to facilitate further studies based on this problem.

We introduce a new silhouette-based representation for modeling clothed human bodies using deep generative models. Our method can reconstruct a complete and textured 3D model of a person wearing clothes from a single input picture. Inspired by the visual hull algorithm, our implicit representation uses 2D silhouettes and 3D joints of a body pose to describe the immense shape complexity and variations of clothed people. Given a segmented 2D silhouette of a person and its inferred 3D joints from the input picture, we first synthesize consistent silhouettes from novel view points around the subject. The synthesized silhouettes which are the most consistent with the input segmentation are fed into a deep visual hull algorithm for robust 3D shape prediction. We then infer the texture of the subject's back view using the frontal image and segmentation mask as input to a conditional generative adversarial network. Our experiments demonstrate that our silhouette-based model is an effective representation and the appearance of the back view can be predicted reliably using an image-to-image translation network. While classic methods based on parametric models often fail for single-view images of subjects with challenging clothing, our approach can still produce successful results, which are comparable to those obtained from multi-view input.

The last decade has witnessed a growing interest in video salient object detection (VSOD). However, the research community long-term lacked a well-established VSOD dataset representative of real dynamic scenes with high-quality annotations. To address this issue, we elaborately collected a visual-attention-consistent Densely Annotated VSOD (DAVSOD) dataset, which contains 226 videos with 23,938 frames that cover diverse realistic-scenes, objects, instances and motions. With corresponding real human eye-fixation data, we obtain precise ground-truths. This is the first work that explicitly emphasizes the challenge of saliency shift, i.e., the video salient object(s) may dynamically change. To further contribute the community a complete benchmark, we systematically assess 17 representative VSOD algorithms over seven existing VSOD datasets and our DAVSOD with totally 84K frames (largest-scale). Utilizing three famous metrics, we then present a comprehensive and insightful performance analysis. Furthermore, we propose a baseline model. It is equipped with a saliency shift- aware convLSTM, which can efficiently capture video saliency dynamics through learning human attention-shift behavior. Extensive experiments open up promising future directions for model development and comparison.

We introduce a data-driven model for interactively synthesizing in-the-wild images from semantic label input masks. Our approach is dramatically different from recent work in this space, in that we make use of no learning. Instead, our approach uses simple but classic tools for matching scene context, shapes, and parts to a stored library of exemplars. Though simple, this approach has several notable advantages over recent work: (1) because nothing is learned, it is not limited to specific training data distributions (such as cityscapes, facades, or faces); (2) it can synthesize arbitrarily high-resolution images, limited only by the resolution of the exemplar library; (3) by appropriately composing shapes and parts, it can generate an exponentially large set of viable candidate output images (that can say, be interactively searched by a user). We present results on the diverse COCO dataset, significantly outperforming learning-based approaches on standard image synthesis metrics. Finally, we explore user-interaction and user-controllability, demonstrating that our system can be used as a platform for user-driven content creation.

For the task of mobility analysis of 3D shapes, we propose joint analysis for simultaneous motion part segmentation and motion attribute estimation, taking a single 3D model as input. The problem is significantly different from those tackled in the existing works which assume the availability of either a pre-existing shape segmentation or multiple 3D models in different motion states. To that end, we develop Shape2Motion which takes a single 3D point cloud as input, and jointly computes a mobility-oriented segmentation and the associated motion attributes. Shape2Motion is comprised of two deep neural networks designed for mobility proposal generation and mobility optimization, respectively. The key contribution of these networks is the novel motion-driven features and losses used in both motion part segmentation and motion attribute estimation. This is based on the observation that the movement of a functional part preserves the shape structure. We evaluate Shape2Motion with a newly proposed benchmark for mobility analysis of 3D shapes. Results demonstrate that our method achieves the state-of-the-art performance both in terms of motion part segmentation and motion attribute estimation.

Synthesizing photo-realistic images from text descriptions is a challenging problem. Previous studies have shown remarkable progresses on visual quality of the generated images. In this paper, we consider semantics from the input text descriptions in helping render photo-realistic images. However, diverse linguistic expressions pose challenges in extracting consistent semantics even they depict the same thing. To this end, we propose a novel photo-realistic text-to-image generation model that implicitly disentangles semantics to both fulfill the high-level semantic consistency and low-level semantic diversity. To be specific, we design (1) a Siamese mechanism in the discriminator to learn consistent high-level semantics, and (2) a visual-semantic embedding strategy by semantic-conditioned batch normalization to find diverse low-level semantics. Extensive experiments and ablation studies on CUB and MS-COCO datasets demonstrate the superiority of the proposed method in comparison to state-of-the-art methods.

We propose spatially-adaptive normalization, a simple but effective layer for synthesizing photorealistic images given an input semantic layout. Previous methods directly feed the semantic layout as input to the network, forcing the network to memorize the information throughout all the layers. Instead, we propose using the input layout for modulating the activations in normalization layers through a spatially-adaptive, learned affine transformation. Experiments on several challenging datasets demonstrate the superiority of our method compared to existing approaches, regarding both visual fidelity and alignment with input layouts. Finally, our model allows users to easily control the style and content of image synthesis results as well as create multi-modal results. Code is available upon publication.

Context is essential for semantic segmentation. Due to the diverse shapes of objects and their complex layout in various scene images, the spatial scales and shapes of contexts for different objects have very large variation. It is thus ineffective or inefficient to aggregate various context information from a predefined fixed region. In this work, we propose to generate a scale- and shape-variant semantic mask for each pixel to confine its contextual region. To this end, we first propose a novel paired convolution to infer the semantic correlation of the pair and based on that to generate a shape mask. Using the inferred spatial scope of the contextual region, we propose a shape-variant convolution, of which the receptive field is controlled by the shape mask that varies with the appearance of input. In this way, the proposed network aggregates the context information of a pixel from its semantic-correlated region instead of a predefined fixed region. Furthermore, this work also proposes a labeling denoising model to reduce wrong predictions caused by the noisy low-level features. Without bells and whistles, the proposed segmentation network achieves new state-of-the-arts consistently on the six public segmentation datasets.

In this paper, we aim to learn discriminative representation for facial action unit (AU) detection from large amount of videos without manual annotations. Inspired by the fact that facial actions are the movements of facial muscles, we depict the movements as the transformation between two face images in different frames and use it as the self-supervisory signal to learn the representations. However, under the uncontrolled condition, the transformation is caused by both facial actions and head motions. To remove the influence by head motions, we propose a Twin-Cycle Autoencoder (TCAE) that can disentangle the facial action related movements and the head motion related ones. Specifically, TCAE is trained to respectively change the facial actions and head poses of the source face to those of the target face. Our experiments validate TCAE's capability of decoupling the movements. Experimental results also demonstrate that the learned representation is discriminative for AU detection, where TCAE outperforms or is comparable with the state-of-the-art self-supervised learning methods and supervised AU detection methods.

Improvements in data-capture and face modeling techniques have enabled us to create high-fidelity realistic face models. However, driving these realistic face models requires special input data, e.g., 3D meshes and unwrapped textures. Also, these face models expect clean input data taken under controlled lab environments, which is very different from data collected in the wild. All these constraints make it challenging to use the high-fidelity models in tracking for commodity cameras. In this paper, we propose a self-supervised domain adaptation approach to enable the animation of high-fidelity face models from a commodity camera. Our approach first circumvents the requirement for special input data by training a new network that can directly drive a face model just from a single 2D image. Then, we overcome the domain mismatch between lab and uncontrolled environments by performing self-supervised domain adaptation based on "consecutive frame texture consistency" based on the assumption that the appearance of the face is consistent over consecutive frames, avoiding the necessity of modeling the new environment such as lighting or background. Experiments show that we are able to drive a high-fidelity face model to perform complex facial motion from a cellphone camera without requiring any labeled data from the new domain.

We present a self-supervision method for 3D hand pose estimation from depth maps. We begin with a neural network initialized with synthesized data and fine-tune it on real but unlabelled depth maps by minimizing a set of data-fitting terms. By approximating the hand surface with a set of spheres, we design a differentiable hand renderer to align estimates by comparing the rendered and input depth maps. In addition, we place a set of priors including a data-driven term to further regulate the estimate's kinematic feasibility. Our method makes highly accurate estimates comparable to current supervised methods which require large amounts of labelled training samples, thereby advancing state-of-the-art in unsupervised learning for hand pose estimation.

This paper proposes an uncalibrated photometric stereo method for non-Lambertian scenes based on deep learning. Unlike previous approaches that heavily rely on assumptions of specific reflectances and light source distributions, our method is able to determine both shape and light directions of a scene with unknown arbitrary reflectances observed under unknown varying light directions. To achieve this goal, we propose a two-stage deep learning architecture, called SDPS-Net, which can effectively take advantage of intermediate supervision, resulting in reduced learning difficulty compared to a single-stage model. Experiments on both synthetic and real datasets show that our proposed approach significantly outperforms previous uncalibrated photometric stereo methods.

We present a self-supervised learning approach for optical flow. Our method distills reliable flow estimations from non-occluded pixels, and uses these predictions as ground truth to learn optical flow for hallucinated occlusions. We further design a simple CNN to utilize temporal information from multiple frames for better flow estimation. These two principles lead to an approach that yields the best performance for unsupervised optical flow learning on the challenging benchmarks including MPI Sintel, KITTI 2012 and 2015. More notably, our self-supervised pre-trained model provides an excellent initialization for supervised fine-tuning. Our fine-tuned models achieve state-of-the-art results on all three datasets. At the time of writing, we achieve EPE=4.26 on the Sintel benchmark, outperforming all submitted methods.

Recently, deep convolutional neural networks (CNNs) have been widely explored in single image super-resolution (SISR) and obtained remarkable performance. However, most of the existing CNN-based SISR methods mainly focus on wider or deeper architecture design, neglecting to explore the feature correlations of intermediate layers, hence hindering the representational power of CNNs. To address this issue, in this paper, we propose a second-order attention network (SAN) for more powerful feature expression and feature correlation learning. Specifically, a novel train- able second-order channel attention (SOCA) module is developed to adaptively rescale the channel-wise features by using second-order feature statistics for more discriminative representations. Furthermore, we present a non-locally enhanced residual group (NLRG) structure, which not only incorporates non-local operations to capture long-distance spatial contextual information, but also contains repeated local-source residual attention groups (LSRAG) to learn increasingly abstract feature representations. Experimental results demonstrate the superiority of our SAN network over state-of-the-art SISR methods in terms of both quantitative metrics and visual quality.