CVPR 2020 Papers

We propose a new method for video object segmentation (VOS) that addresses object pattern learning from unlabeled videos, unlike most existing methods which rely heavily on extensive annotated data. We introduce a unified unsupervised/weakly supervised learning framework, called MuG, that comprehensively captures intrinsic properties of VOS at multiple granularities. Our approach can help advance understanding of visual patterns in VOS and significantly reduce annotation burden. With a carefully-designed architecture and strong representation learning ability, our learned model can be applied to diverse VOS settings, including object-level zero-shot VOS, instance-level zero-shot VOS, and one-shot VOS. Experiments demonstrate promising performance in these settings, as well as the potential of MuG in leveraging unlabeled data to further improve the segmentation accuracy.

In this paper, we introduce an open vocabulary model for image hashtag prediction - the task of mapping an image to its accompanying hashtags. Recent work shows that to build an accurate hashtag prediction model, it is necessary to model the user because of the self-expression problem, in which similar image content may be labeled with different tags. To take into account the user behaviour, we propose a new model that extracts a representation of a user based on his/her image history. Our model allows to improve a user representation with new images or add a new user without retraining the model. Because new hashtags appear all the time on social networks, we design an open vocabulary model which can deal with new hashtags without retraining the model. Our model learns a cross-modal embedding between user conditional visual representations and hashtag word representations. Experiments on a subset of the YFCC100M dataset demonstrate the efficacy of our user representation in user conditional hashtag prediction and user retrieval. We further validate the open vocabulary prediction ability of our model.

Humans perceive the 3D world as a set of distinct objects that are characterized by various low-level (geometry, reflectance) and high-level (connectivity, adjacency, symmetry) properties. Recent methods based on convolutional neural networks (CNNs) demonstrated impressive progress in 3D reconstruction, even when using a single 2D image as input. However, the majority of these methods focuses on recovering the local 3D geometry of an object without considering its part-based decomposition or relations between parts. We address this challenging problem by proposing a novel formulation that allows to jointly recover the geometry of a 3D object as a set of primitives as well as their latent hierarchical structure without part-level supervision. Our model recovers the higher level structural decomposition of various objects in the form of a binary tree of primitives, where simple parts are represented with fewer primitives and more complex parts are modeled with more components. Our experiments on the ShapeNet and D-FAUST datasets demonstrate that considering the organization of parts indeed facilitates reasoning about 3D geometry.

Composing and recognizing new concepts from known sub-concepts has been a fundamental and challenging vision task, mainly due to 1) the diversity of sub-concepts and 2) the intricate contextuality between sub-concepts and their corresponding visual features. However, most of the current methods simply treat the contextuality as rigid semantic relationships and fail to capture fine-grained contextual correlations. We propose to learn unseen concepts in a hierarchical decomposition-and-composition manner. Considering the diversity of sub-concepts, our method decomposes each seen image into visual elements according to its labels, and learns corresponding sub-concepts in their individual subspaces. To model intricate contextuality between sub-concepts and their visual features, compositions are generated from these subspaces in three hierarchical forms, and the composed concepts are learned in a unified composition space. To further refine the captured contextual relationships, adaptively semi-positive concepts are defined and then learned with pseudo supervision exploited from the generated compositions. We validate the proposed approach on two challenging benchmarks, and demonstrate its superiority over state-of-the-art approaches.

We study on image super-resolution (SR), which aims to recover realistic textures from a low-resolution (LR) image. Recent progress has been made by taking high-resolution images as references (Ref), so that relevant textures can be transferred to LR images. However, existing SR approaches neglect to use attention mechanisms to transfer high-resolution (HR) textures from Ref images, which limits these approaches in challenging cases. In this paper, we propose a novel Texture Transformer Network for Image Super-Resolution (TTSR), in which the LR and Ref images are formulated as queries and keys in a transformer, respectively. TTSR consists of four closely-related modules optimized for image generation tasks, including a learnable texture extractor by DNN, a relevance embedding module, a hard-attention module for texture transfer, and a soft-attention module for texture synthesis. Such a design encourages joint feature learning across LR and Ref images, in which deep feature correspondences can be discovered by attention, and thus accurate texture features can be transferred. The proposed texture transformer can be further stacked in a cross-scale way, which enables texture recovery from different levels (e.g., from 1x to 4x magnification). Extensive experiments show that TTSR achieves significant improvements over state-of-the-art approaches on both quantitative and qualitative evaluations.

Since annotating pixel-level labels for semantic segmentation is laborious, leveraging synthetic data is an attractive solution. However, due to the domain gap between synthetic domain and real domain, it is challenging for a model trained with synthetic data to generalize to real data. In this paper, considering the fundamental difference between the two domains as the texture, we propose a method to adapt to the target domain's texture. First, we diversity the texture of synthetic images using a style transfer algorithm. The various textures of generated images prevent a segmentation model from overfitting to one specific (synthetic) texture. Then, we fine-tune the model with self-training to get direct supervision of the target texture. Our results achieve state-of-the-art performance and we analyze the properties of the model trained on the stylized dataset with extensive experiments.

Temporal action localization (TAL) in untrimmed videos recently receives tremendous research enthusiasm. To our best knowledge, this is the first attempt in the literature to explore this task under an unsupervised setting, hereafter referred to as action co-localization (ACL), where only the total count of unique actions that appear in the video set is known. To solve ACL, we propose a two-step "clustering + localization" iterative procedure. The clustering step provides noisy pseudo-labels for the localization step, and the localization step provides temporal co-attention models that in turn improve the clustering performance. Using such two-step procedure, weakly-supervised TAL can be regarded as a direct extension of our ACL model. Technically, our contributions are two-folds: 1) temporal co-attention models, either class-specific or class-agnostic, learned from video-level labels or pseudo-labels in an iterative reinforced fashion; 2) new losses specially designed for ACL, including action-background separation loss and cluster-based triplet loss. Comprehensive evaluations are conducted on 20-action THUMOS14 and 100-action ActivityNet-1.2. On both benchmarks, the proposed model for ACL exhibits strong performances, even surprisingly comparable with state-of-the-art weakly-supervised methods. For example, previous best weakly-supervised model achieves 26.8% under mAP@0.5 on THUMOS14, our new records are 30.1% (weakly-supervised) and 25.0% (unsupervised).

Human drivers have a remarkable ability to drive in diverse visual conditions and situations, e.g., from maneuvering in rainy, limited visibility conditions with no lane markings to turning in a busy intersection while yielding to pedestrians. In contrast, we find that state-of-the-art sensorimotor driving models struggle when encountering diverse settings with varying relationships between observation and action. To generalize when making decisions across diverse conditions, humans leverage multiple types of situation-specific reasoning and learning strategies. Motivated by this observation, we develop a framework for learning a situational driving policy that effectively captures reasoning under varying types of scenarios. Our key idea is to learn a mixture model with a set of policies that can capture multiple driving modes. We first optimize the mixture model through behavior cloning and show it to result in significant gains in terms of driving performance in diverse conditions. We then refine the model by directly optimizing for the driving task itself, i.e., supervised with the navigation task reward. Our method is more scalable than methods assuming access to privileged information, e.g., perception labels, as it only assumes demonstration and reward-based supervision. We achieve over 98% success rate on the CARLA driving benchmark as well as state-of-the-art performance on a newly introduced generalization benchmark.

Saliency detection on RGB-D images is receiving more and more research interests recently. Previous models adopt the early fusion or the result fusion scheme to fuse the input RGB and depth data or their saliency maps, which incur the problem of distribution gap or information loss. Some other models use the feature fusion scheme but are limited by the linear feature fusion methods. In this paper, we propose to fuse attention learned in both modalities. Inspired by the Non-local model, we integrate the self-attention and each other's attention to propagate long-range contextual dependencies, thus incorporating multi-modal information to learn attention and propagate contexts more accurately. Considering the reliability of the other modality's attention, we further propose a selection attention to weight the newly added attention term. We embed the proposed attention module in a two-stream CNN for RGB-D saliency detection. Furthermore, we also propose a residual fusion module to fuse the depth decoder features into the RGB stream. Experimental results on seven benchmark datasets demonstrate the effectiveness of the proposed model components and our final saliency model. Our code and saliency maps are available at https://github.com/nnizhang/S2MA.

Instance segmentation is a challenging task for both modeling and annotation. Due to the high annotation cost, modeling becomes more difficult because of the limited amount of supervision. We aim to improve the accuracy of the existing instance segmentation models by utilizing a large amount of detection supervision. We propose ShapeProp, which learns to activate the salient regions within the object detection and propagate the areas to the whole instance through an iterative learnable message passing module. ShapeProp can benefit from more bounding box supervision to locate the instances more accurately and utilize the feature activations from the larger number of instances to achieve more accurate segmentation. We extensively evaluate ShapeProp on three datasets (MS COCO, PASCAL VOC, and BDD100k) with different supervision setups based on both two-stage (Mask R-CNN) and single-stage (RetinaMask) models. The results show our method establishes new states of the art for semi-supervised instance segmentation.

Self-supervised representation learning targets to learn convnet-based image representations from unlabeled data. Inspired by the success of NLP methods in this area, in this work we propose a self-supervised approach based on spatially dense image descriptions that encode discrete visual concepts, here called visual words. To build such discrete representations, we quantize the feature maps of a first pre-trained self-supervised convnet, over a k-means based vocabulary. Then, as a self-supervised task, we train another convnet to predict the histogram of visual words of an image (i.e., its Bag-of-Words representation) given as input a perturbed version of that image. The proposed task forces the convnet to learn perturbation-invariant and context-aware image features, useful for downstream image understanding tasks. We extensively evaluate our method and demonstrate very strong empirical results, e.g., our pre-trained self-supervised representations transfer better on detection task and similarly on classification over classes "unseen" during pre-training, when compared to the supervised case. This also shows that the process of image discretization into visual words can provide the basis for very powerful self-supervised approaches in the image domain, thus allowing further connections to be made to related methods from the NLP domain that have been extremely successful so far.

High-dynamic range (HDR) imaging is an essential imaging modality for a wide range of applications in uncontrolled environments, including autonomous driving, robotics, and mobile phone cameras. However, existing HDR techniques in commodity devices struggle with dynamic scenes due to multi-shot acquisition and post-processing time, e.g. mobile phone burst photography, making such approaches unsuitable for real-time applications. In this work, we propose a method for snapshot HDR imaging by learning an optical HDR encoding in a single image which maps saturated highlights into neighboring unsaturated areas using a diffractive optical element (DOE). We propose a novel rank-1 parameterization of the proposed DOE which avoids vast trainable parameters and keeps high frequencies' encoding compared with conventional end-to-end design methods. We further propose a reconstruction network tailored to this rank-1 parametrization for recovery of clipped information from the encoded measurements. The proposed end-to-end framework is validated through simulation and real-world experiments and improves the PSNR by more than 7 dB over state-of-the-art end-to-end designs.

Relighting is an essential step in realistically transferring objects from a captured image into another environment. For example, authentic telepresence in Augmented Reality requires faces to be displayed and relit consistent with the observer's scene lighting. We investigate end-to-end deep learning architectures that both de-light and relight an image of a human face. Our model decomposes the input image into intrinsic components according to a diffuse physics-based image formation model. We enable non-diffuse effects including cast shadows and specular highlights by predicting a residual correction to the diffuse render. To train and evaluate our model, we collected a portrait database of 21 subjects with various expressions and poses. Each sample is captured in a controlled light stage setup with 32 individual light sources. Our method creates precise and believable relighting results and generalizes to complex illumination conditions and challenging poses, including when the subject is not looking straight at the camera.

High-fidelity face completion is a challenging task due to the rich and subtle facial textures involved. What makes it more complicated is the correlations between different facial components, for example, the symmetry in texture and structure between both eyes. While recent works adopted the attention mechanism to learn the contextual relations among elements of the face, they have largely overlooked the disastrous impacts of inaccurate attention scores; in addition, they fail to pay sufficient attention to key facial components, the completion results of which largely determine the authenticity of a face image. Accordingly, in this paper, we design a comprehensive framework for face completion based on the U-Net structure. Specifically, we propose a dual spatial attention module to efficiently learn the correlations between facial textures at multiple scales; moreover, we provide an oracle supervision signal to the attention module to ensure that the obtained attention scores are reasonable. Furthermore, we take the location of the facial components as prior knowledge and impose a multi-discriminator on these regions, with which the fidelity of facial components is significantly promoted. Extensive experiments on two high-resolution face datasets including CelebA-HQ and Flickr-Faces-HQ demonstrate that the proposed approach outperforms state-of-the-art methods by large margins.

Detecting and analyzing nanoscale motion patterns of vesicles, smaller than the microscope resolution ( 250 nm), inside living biological cells is a challenging problem. State-of-the-art CV approaches based on detection, tracking, optical flow or deep learning perform poorly for this problem. We propose an integrative approach, built upon physics based simulations, nanoscopy algorithms, and shallow residual attention network to make it possible for the first time to analysis sub-resolution motion patterns in vesicles that may also be of sub-resolution diameter. Our results show state-of-the-art performance, 89% validation accuracy on simulated dataset and 82% testing accuracy on an experimental dataset of living heart muscle cells imaged under three different pathological conditions. We demonstrate automated analysis of the motion states and changed in them for over 9000 vesicles. Such analysis will enable large scale biological studies of vesicle transport and interaction in living cells in the future.

We present a novel, end-to-end learnable, multiview 3D point cloud registration algorithm. Registration of multiple scans typically follows a two-stage pipeline: the initial pairwise alignment and the globally consistent refinement. The former is often ambiguous due to the low overlap of neighboring point clouds, symmetries and repetitive scene parts. Therefore, the latter global refinement aims at establishing the cyclic consistency across multiple scans and helps in resolving the ambiguous cases. In this paper we propose, to the best of our knowledge, the first end-to-end algorithm for joint learning of both parts of this two-stage problem. Experimental evaluation on well accepted benchmark datasets shows that our approach outperforms the state-of-the-art by a significant margin, while being end-to-end trainable and computationally less costly. Moreover, we present detailed analysis and an ablation study that validate the novel components of our approach. The source code and pretrained models are publicly available under https://github.com/zgojcic/3D_multiview_reg.

We propose to construct a view graph to excavate the information of the whole given sequence for absolute camera pose estimation. Specifically, we harness GNNs to model the graph, allowing even non-consecutive frames to exchange information with each other. Rather than adopting the regular GNNs directly, we redefine the nodes, edges, and embedded functions to fit the relocalization task. Redesigned GNNs cooperate with CNNs in guiding knowledge propagation and feature extraction respectively to process multi-view high-dimension image features iteratively at different levels. Besides, a general graph-based loss function beyond constraints between consecutive views is employed for training the network in an end-to-end fashion. Extensive experiments conducted on both indoor and outdoor datasets demonstrate that our method outperforms previous approaches especially in large-scale and challenging scenarios.

In this work we contribute a novel pipeline to automatically generate training data, and to improve over state-of-the-art multi-object tracking and segmentation (MOTS) methods. Our proposed track mining algorithm turns raw street-level videos into high-fidelity MOTS training data, is scalable and overcomes the need of expensive and time-consuming manual annotation approaches. We leverage state-of-the-art instance segmentation results in combination with optical flow predictions, also trained on automatically harvested training data. Our second major contribution is MOTSNet - a deep learning, tracking-by-detection architecture for MOTS - deploying a novel mask-pooling layer for improved object association over time. Training MOTSNet with our automatically extracted data leads to significantly improved sMOTSA scores on the novel KITTI MOTS dataset (+1.9%/+7.5% on cars/pedestrians), and MOTSNet improves by +4.1% over previously best methods on the MOTSChallenge dataset. Our most impressive finding is that we can improve over previous best-performing works, even in complete absence of manually annotated MOTS training data.

Video-based person re-identification (re-ID) is an important research topic in computer vision. The key to tackling the challenging task is to exploit both spatial and temporal clues in video sequences. In this work, we propose a novel graph-based framework, namely Multi-Granular Hypergraph (MGH), to pursue better representational capabilities by modeling spatiotemporal dependencies in terms of multiple granularities. Specifically, hypergraphs with different spatial granularities are constructed using various levels of part-based features across the video sequence. In each hypergraph, different temporal granularities are captured by hyperedges that connect a set of graph nodes (i.e., part-based features) across different temporal ranges. Two critical issues (misalignment and occlusion) are explicitly addressed by the proposed hypergraph propagation and feature aggregation schemes. Finally, we further enhance the overall video representation by learning more diversified graph-level representations of multiple granularities based on mutual information minimization. Extensive experiments on three widely-adopted benchmarks clearly demonstrate the effectiveness of the proposed framework. Notably, 90.0% top-1 accuracy on MARS is achieved using MGH, outperforming the state-of-the-arts.