CVPR 2019 论文列表

This paper conducts a systematic study on the role of visual attention in Unsupervised Video Object Segmentation (UVOS) tasks. By elaborately annotating three popular video segmentation datasets (DAVIS, Youtube-Objects and SegTrack V2) with dynamic eye-tracking data in the UVOS setting, for the first time, we quantitatively verified the high consistency of visual attention behavior among human observers, and found strong correlation between human attention and explicit primary object judgements during dynamic, task-driven viewing. Such novel observations provide an in-depth insight into the underlying rationale behind UVOS. Inspired by these findings, we decouple UVOS into two sub-tasks: UVOS-driven Dynamic Visual Attention Prediction (DVAP) in spatiotemporal domain, and Attention-Guided Object Segmentation (AGOS) in spatial domain. Our UVOS solution enjoys three major merits: 1) modular training without using expensive video segmentation annotations, instead, using more affordable dynamic fixation data to train the initial video attention module and using existing fixation-segmentation paired static/image data to train the subsequent segmentation module; 2) comprehensive foreground understanding through multi-source learning; and 3) additional interpretability from the biologically-inspired and assessable attention. Experiments on popular benchmarks show that, even without using expensive video object mask annotations, our model achieves compelling performance in comparison with state-of-the-arts.

Reconstructing the 3D model of a physical object typically requires us to align the depth scans obtained from different camera poses into the same coordinate system. Solutions to this global alignment problem usually proceed in two steps. The first step estimates relative transformations between pairs of scans using an off-the-shelf technique. Due to limited information presented between pairs of scans, the resulting relative transformations are generally noisy. The second step then jointly optimizes the relative transformations among all input depth scans. A natural constraint used in this step is the cycle-consistency constraint, which allows us to prune incorrect relative transformations by detecting inconsistent cycles. The performance of such approaches, however, heavily relies on the quality of the input relative transformations. Instead of merely using the relative transformations as the input to perform transformation synchronization, we propose to use a neural network to learn the weights associated with each relative transformation. Our approach alternates between transformation synchronization using weighted relative transformations and predicting new weights of the input relative transformations using a neural network. We demonstrate the usefulness of this approach across a wide range of datasets.

An exact method of correcting the rolling shutter (RS) effect requires recovering the underlying geometry, i.e. the scene structures and the camera motions between scanlines or between views. However, the multiple-view geometry for RS cameras is much more complicated than its global shutter (GS) counterpart, with various degeneracies. In this paper, we first make a theoretical contribution by showing that RS two-view geometry is degenerate in the case of pure translational camera motion. In view of the complex RS geometry, we then propose a Convolutional Neural Network (CNN)-based method which learns the underlying geometry (camera motion and scene structure) from just a single RS image and perform RS image correction. We call our method structure-and-motion-aware RS correction because it reasons about the concealed motions between the scanlines as well as the scene structure. Our method learns from a large-scale dataset synthesized in a geometrically meaningful way where the RS effect is generated in a manner consistent with the camera motion and scene structure. In extensive experiments, our method achieves superior performance compared to other state-of-the-art methods for single image RS correction and subsequent Structure from Motion (SfM) applications.

Convolutional Neural Networks (CNN) have been regarded as a powerful class of models for visual recognition problems. Nevertheless, the convolutional filters in these networks are local operations while ignoring the large-range dependency. Such drawback becomes even worse particularly for video recognition, since video is an information-intensive media with complex temporal variations. In this paper, we present a novel framework to boost the spatio-temporal representation learning by Local and Global Diffusion (LGD). Specifically, we construct a novel neural network architecture that learns the local and global representations in parallel. The architecture is composed of LGD blocks, where each block updates local and global features by modeling the diffusions between these two representations. Diffusions effectively interact two aspects of information, i.e., localized and holistic, for more powerful way of representation learning. Furthermore, a kernelized classifier is introduced to combine the representations from two aspects for video recognition. Our LGD networks achieve clear improvements on the large-scale Kinetics-400 and Kinetics-600 video classification datasets against the best competitors by 3.5% and 0.7%. We further examine the generalization of both the global and local representations produced by our pre-trained LGD networks on four different benchmarks for video action recognition and spatio-temporal action detection tasks. Superior performances over several state-of-the-art techniques on these benchmarks are reported.

We integrate two powerful ideas, geometry and deep visual representation learning, into recurrent network architectures for mobile visual scene understanding. The proposed networks learn to "lift" 2D visual features and integrate them over time into latent 3D feature maps of the scene. They are equipped with differentiable geometric operations, such as projection, unprojection, egomotion estimation and stabilization, in order to compute a geometrically-consistent mapping between the world scene and their 3D latent feature space. We train the proposed architectures to predict novel image views given short frame sequences as input. Their predictions strongly generalize to scenes with a novel number of objects, appearances and configurations, and greatly outperform predictions of previous works that do not consider egomotion stabilization or a space-aware latent feature space. We train the proposed architectures to detect and segment objects in 3D, using the latent 3D feature map as input--as opposed to 2D feature maps computed from video frames. The resulting detections are permanent: they continue to exist even when an object gets occluded or leaves the field of view. Our experiments suggest the proposed space-aware latent feature arrangement and egomotion-stabilized convolutions are essential architectural choices for spatial common sense to emerge in artificial embodied visual agents.

Depth estimation from a single image in the wild remains a challenging problem. One main obstacle is the lack of high-quality training data for images in the wild. In this paper we propose a method to automatically generate such data through Structure-from-Motion (SfM) on Internet videos. The core of this method is a Quality Assessment Network that identifies high-quality reconstructions obtained from SfM. Using this method, we collect single-view depth training data from a large number of YouTube videos and construct a new dataset called YouTube3D. Experiments show that YouTube3D is useful in training depth estimation networks and advances the state of the art of single-view depth estimation in the wild.

We address the problem of detecting scene text in arbitrary shapes, which is a challenging task due to the high variety and complexity of the scene. Specifically, we treat text detection as instance segmentation and propose a segmentation-based framework, which extracts each text instance as an independent connected component. To distinguish different text instances, our method maps pixels onto an embedding space where pixels belonging to the same text are encouraged to appear closer to each other and vise versa. In addition, we introduce a Shape-Aware Loss to make training adaptively accommodate various aspect ratios of text instances and the tiny gaps among them, and a new post-processing pipeline to yield precise bounding box predictions. Experimental results on three challenging datasets (ICDAR15, MSRA-TD500 and CTW1500) demonstrate the effectiveness of our work.

As an alternative to manual pixel-wise annotation, synthetic data has been increasingly used for training semantic segmentation models. Such synthetic images and semantic labels can be easily generated from virtual 3D environments. In this work, we propose an approach to cross-domain semantic segmentation with the auxiliary geometric information, which can also be easily obtained from virtual environments. The geometric information is utilized on two levels for reducing domain shift: on the input level, we augment the standard image translation network with the geometric information to translate synthetic images into realistic style; on the output level, we build a task network which simultaneously performs semantic segmentation and depth estimation. Meanwhile, adversarial training is applied on the joint output space to preserve the correlation between semantics and depth. The proposed approach is validated on two pairs of synthetic to real dataset: from Virtual KITTI to KITTI, and from SYNTHIA to Cityscapes, where we achieve a clear performance gain compared to the baselines and various competing methods, demonstrating the effectiveness of the geometric information for cross-domain semantic segmentation.

Object detection in aerial images is an active yet challenging task in computer vision because of the bird's-eye view perspective, the highly complex backgrounds, and the variant appearances of objects. Especially when detecting densely packed objects in aerial images, methods relying on horizontal proposals for common object detection often introduce mismatches between the Region of Interests (RoIs) and objects. This leads to the common misalignment between the final object classification confidence and localization accuracy. In this paper, we propose a RoI Transformer to address these problems. The core idea of RoI Transformer is to apply spatial transformations on RoIs and learn the transformation parameters under the supervision of oriented bounding box (OBB) annotations. RoI Transformer is with lightweight and can be easily embedded into detectors for oriented object detection. Simply apply the RoI Transformer to light head RCNN has achieved state-of-the-art performances on two common and challenging aerial datasets, i.e., DOTA and HRSC2016, with a neglectable reduction to detection speed. Our RoI Transformer exceeds the deformable Position Sensitive RoI pooling when oriented bounding-box annotations are available. Extensive experiments have also validated the flexibility and effectiveness of our RoI Transformer.

Appearance features have been widely used in video anomaly detection even though they contain complex entangled factors. We propose a new method to model the normal patterns of human movements in surveillance video for anomaly detection using dynamic skeleton features. We decompose the skeletal movements into two sub-components: global body movement and local body posture. We model the dynamics and interaction of the coupled features in our novel Message-Passing Encoder-Decoder Recurrent Network. We observed that the decoupled features collaboratively interact in our spatio-temporal model to accurately identify human-related irregular events from surveillance video sequences. Compared to traditional appearance-based models, our method achieves superior outlier detection performance. Our model also offers "open-box" examination and decision explanation made possible by the semantically understandable features and a network architecture supporting interpretability.

High-quality image inpainting requires filling missing regions in a damaged image with plausible content. Existing works either fill the regions by copying high-resolution patches or generating semantically-coherent patches from region context, while neglecting the fact that both visual and semantic plausibility are highly-demanded. In this paper, we propose a Pyramid-context Encoder Network (denoted as PEN-Net) for image inpainting by deep generative models. The proposed PEN-Net is built upon a U-Net structure with three tailored components, ie., a pyramid-context encoder, a multi-scale decoder, and an adversarial training loss. First, we adopt a U-Net as backbone which can encode the context of an image from high-resolution pixels into high-level semantic features, and decode the features reversely. Second, we propose a pyramid-context encoder, which progressively learns region affinity by attention from a high-level semantic feature map, and transfers the learned attention to its adjacent high-resolution feature map. As the missing content can be filled by attention transfer from deep to shallow in a pyramid fashion, both visual and semantic coherence for image inpainting can be ensured. Third, we further propose a multi-scale decoder with deeply-supervised pyramid losses and an adversarial loss. Such a design not only results in fast convergence in training, but more realistic results in testing. Extensive experiments on a broad range of datasets shows the superior performance of the proposed network.

The dominant deep learning approaches use a "one-size-fits-all" paradigm with the hope that underlying characteristics of diverse inputs can be captured via a fixed structure. They also overlook the importance of explicitly modeling feature hierarchy. However, complex real-world tasks often require discovering diverse reasoning paths for different inputs to achieve satisfying predictions, especially for challenging large-scale recognition tasks with complex label relations. In this paper, we treat the structured commonsense knowledge (e.g. concept hierarchy) as the guidance of customizing more powerful and explainable network structures for distinct inputs, leading to dynamic and individualized inference paths. Give an off-the-shelf large network configuration, the proposed Personalized Modular Network (PMN) is learned by selectively activating a sequence of network modules where each of them is designated to recognize particular levels of structured knowledge. Learning semantic configurations and activation of modules to align well with structured knowledge can be regarded as a decision-making procedure, which is solved by a new graph-based reinforcement learning algorithm. Experiments on three semantic segmentation tasks and classification tasks show our PMN can achieve superior performance with the reduced number of network modules while discovering personalized and explainable module configurations for each input.

Stereo image pairs can be used to improve the performance of super-resolution (SR) since additional information is provided from a second viewpoint. However, it is challenging to incorporate this information for SR since disparities between stereo images vary significantly. In this paper, we propose a parallax-attention stereo superresolution network (PASSRnet) to integrate the information from a stereo image pair for SR. Specifically, we introduce a parallax-attention mechanism with a global receptive field along the epipolar line to handle different stereo images with large disparity variations. We also propose a new and the largest dataset for stereo image SR (namely, Flickr1024). Extensive experiments demonstrate that the parallax-attention mechanism can capture correspondence between stereo images to improve SR performance with a small computational and memory cost. Comparative results show that our PASSRnet achieves the state-of-the-art performance on the Middlebury, KITTI 2012 and KITTI 2015 datasets.

We propose a novel regularization algorithm to train deep neural networks, in which data at training time is severely biased. Since a neural network efficiently learns data distribution, a network is likely to learn the bias information to categorize input data. It leads to poor performance at test time, if the bias is, in fact, irrelevant to the categorization. In this paper, we formulate a regularization loss based on mutual information between feature embedding and bias. Based on the idea of minimizing this mutual information, we propose an iterative algorithm to unlearn the bias information. We employ an additional network to predict the bias distribution and train the network adversarially against the feature embedding network. At the end of learning, the bias prediction network is not able to predict the bias not because it is poorly trained, but because the feature embedding network successfully unlearns the bias information. We also demonstrate quantitative and qualitative experimental results which show that our algorithm effectively removes the bias information from feature embedding.

Given a set of input views, multi-view stereopsis techniques estimate depth maps to represent the 3D reconstruction of the scene; these are fused into a single, consistent, reconstruction -- most often a point cloud. In this work we propose to learn an auto-regressive depth refinement directly from data. While deep learning has improved the accuracy and speed of depth estimation significantly, learned MVS techniques remain limited to the planesweeping paradigm. We refine a set of input depth maps by successively reprojecting information from neighbouring views to leverage multi-view constraints. Compared to learning-based volumetric fusion techniques, an image-based representation allows significantly more detailed reconstructions; compared to traditional point-based techniques, our method learns noise suppression and surface completion in a data-driven fashion. Due to the limited availability of high-quality reconstruction datasets with ground truth, we introduce two novel synthetic datasets to (pre-)train our network. Our approach is able to improve both the output depth maps and the reconstructed point cloud, for both learned and traditional depth estimation front-ends, on both synthetic and real data.

Multi-class segmentation has recently achieved significant performance in natural images and videos. This achievement is due primarily to the public availability of large multi-class datasets. However, there are certain domains, such as biomedical images, where obtaining sufficient multi-class annotations is a laborious and often impossible task and only single-class datasets are available. While existing segmentation research in such domains use private multi-class datasets or focus on single-class segmentations, we propose a unified highly efficient framework for robust simultaneous learning of multi-class segmentations by combining single-class datasets and utilizing a novel way of conditioning a convolutional network for the purpose of segmentation. We demonstrate various ways of incorporating the conditional information, perform an extensive evaluation, and show compelling multi-class segmentation performance on biomedical images, which outperforms current state-of-the-art solutions (up to 2.7%). Unlike current solutions, which are meticulously tailored for particular single-class datasets, we utilize datasets from a variety of sources. Furthermore, we show the applicability of our method also to natural images and evaluate it on the Cityscapes dataset. We further discuss other possible applications of our proposed framework.

Depth estimation from a single image represents a fascinating, yet challenging problem with countless applications. Recent works proved that this task could be learned without direct supervision from ground truth labels leveraging image synthesis on sequences or stereo pairs. Focusing on this second case, in this paper we leverage stereo matching in order to improve monocular depth estimation. To this aim we propose monoResMatch, a novel deep architecture designed to infer depth from a single input image by synthesizing features from a different point of view, horizontally aligned with the input image, performing stereo matching between the two cues. In contrast to previous works sharing this rationale, our network is the first trained end-to-end from scratch. Moreover, we show how obtaining proxy ground truth annotation through traditional stereo algorithms, such as Semi-Global Matching, enables more accurate monocular depth estimation still countering the need for expensive depth labels by keeping a self-supervised approach. Exhaustive experimental results prove how the synergy between i) the proposed monoResMatch architecture and ii) proxy-supervision attains state-of-the-art for self-supervised monocular depth estimation. The code is publicly available at https://github.com/fabiotosi92/monoResMatch-Tensorflow.

Metric learning networks are used to compute image embeddings, which are widely used in many applications such as image retrieval and face recognition. In this paper, we propose to use network distillation to efficiently compute image embeddings with small networks. Network distillation has been successfully applied to improve image classification, but has hardly been explored for metric learning. To do so, we propose two new loss functions that model the communication of a deep teacher network to a small student network. We evaluate our system in several datasets, including CUB-200-2011, Cars-196, Stanford Online Products and show that embeddings computed using small student networks perform significantly better than those computed using standard networks of similar size. Results on a very compact network (MobileNet-0.25), which can be used on mobile devices, show that the proposed method can greatly improve Recall@1 results from 27.5% to 44.6%. Furthermore, we investigate various aspects of distillation for embeddings, including hint and attention layers, semi-supervised learning and cross quality distillation. (Code is available at https://github.com/yulu0724/EmbeddingDistillation).

The performance of deep neural networks improves with more annotated data. The problem is that the budget for annotation is limited. One solution to this is active learning, where a model asks human to annotate data that it perceived as uncertain. A variety of recent methods have been proposed to apply active learning to deep networks but most of them are either designed specific for their target tasks or computationally inefficient for large networks. In this paper, we propose a novel active learning method that is simple but task-agnostic, and works efficiently with the deep networks. We attach a small parametric module, named "loss prediction module," to a target network, and learn it to predict target losses of unlabeled inputs. Then, this module can suggest data that the target model is likely to produce a wrong prediction. This method is task-agnostic as networks are learned from a single loss regardless of target tasks. We rigorously validate our method through image classification, object detection, and human pose estimation, with the recent network architectures. The results demonstrate that our method consistently outperforms the previous methods over the tasks.