CVPR 2020 Papers

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).

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.

Face recognition systems are usually faced with unseen domains in real-world applications and show unsatisfactory performance due to their poor generalization. For example, a well-trained model on webface data cannot deal with the ID vs. Spot task in surveillance scenario. In this paper, we aim to learn a generalized model that can directly handle new unseen domains without any model updating. To this end, we propose a novel face recognition method via meta-learning named Meta Face Recognition (MFR). MFR synthesizes the source/target domain shift with a meta-optimization objective, which requires the model to learn effective representations not only on synthesized source domains but also on synthesized target domains. Specifically, we build domain-shift batches through a domain-level sampling strategy and get back-propagated gradients/meta-gradients on synthesized source/target domains by optimizing multi-domain distributions. The gradients and meta-gradients are further combined to update the model to improve generalization. Besides, we propose two benchmarks for generalized face recognition evaluation. Experiments on our benchmarks validate the generalization of our method compared to several baselines and other state-of-the-arts. The proposed benchmarks and code will be available at https://github.com/cleardusk/MFR.

Interactions between people are often governed by their relationships. On the flip side, social relationships are built upon several interactions. Two strangers are more likely to greet and introduce themselves while becoming friends over time. We are fascinated by this interplay between interactions and relationships, and believe that it is an important aspect of understanding social situations. In this work, we propose neural models to learn and jointly predict interactions, relationships, and the pair of characters that are involved. We note that interactions are informed by a mixture of visual and dialog cues, and present a multimodal architecture to extract meaningful information from them. Localizing the pair of interacting characters in video is a time-consuming process, instead, we train our model to learn from clip-level weak labels. We evaluate our models on the MovieGraphs dataset and show the impact of modalities, use of longer temporal context for predicting relationships, and achieve encouraging performance using weak labels as compared with ground-truth labels. Code is online.

In this paper, we present a learning-based framework for light field view synthesis from a subset of input views. Building upon a light-weight optical flow estimation network to obtain depth maps, our method employs two reconstruction modules in pixel and feature domains respectively. For the pixel-wise reconstruction, occlusions are explicitly handled by a disparity-dependent interpolation filter, whereas inpainting on disoccluded areas is learned by convolutional layers. Due to disparity inconsistencies, the pixel-based reconstruction may lead to blurriness in highly textured areas as well as on object contours. On the contrary, the feature-based reconstruction well performs on high frequencies, making the reconstruction in the two domains complementary. End-to-end learning is finally performed including a fusion module merging pixel and feature-based reconstructions. Experimental results show that our method achieves state-of-the-art performance on both synthetic and real-world datasets, moreover, it is even able to extend light fields' baseline by extrapolating high quality views without additional training.

Despite great success in human parsing, progress for parsing other deformable articulated objects, like animals, is still limited by the lack of labeled data. In this paper, we use synthetic images and ground truth generated from CAD animal models to address this challenge. To bridge the domain gap between real and synthetic images, we propose a novel consistency-constrained semi-supervised learning method (CC-SSL). Our method leverages both spatial and temporal consistencies, to bootstrap weak models trained on synthetic data with unlabeled real images. We demonstrate the effectiveness of our method on highly deformable animals, such as horses and tigers. Without using any real image label, our method allows for accurate keypoint prediction on real images. Moreover, we quantitatively show that models using synthetic data achieve better generalization performance than models trained on real images across different domains in the Visual Domain Adaptation Challenge dataset. Our synthetic dataset contains 10+ animals with diverse poses and rich ground truth, which enables us to use the multi-task learning strategy to further boost models' performance.

Video object segmentation (VOS) is a highly challenging problem since the initial mask, defining the target object, is only given at test-time. The main difficulty is to effectively handle appearance changes and similar background objects, while maintaining accurate segmentation. Most previous approaches fine-tune segmentation networks on the first frame, resulting in impractical frame-rates and risk of overfitting. More recent methods integrate generative target appearance models, but either achieve limited robustness or require large amounts of training data. We propose a novel VOS architecture consisting of two network components. The target appearance model consists of a light-weight module, which is learned during the inference stage using fast optimization techniques to predict a coarse but robust target segmentation. The segmentation model is exclusively trained offline, designed to process the coarse scores into high quality segmentation masks. Our method is fast, easily trainable and remains highly effective in cases of limited training data. We perform extensive experiments on the challenging YouTube-VOS and DAVIS datasets. Our network achieves favorable performance, while operating at higher frame-rates compared to state-of-the-art. Code and trained models are available at https://github.com/andr345/frtm-vos.

Recently, numerous handcrafted and searched networks have been applied for semantic segmentation. However, previous works intend to handle inputs with various scales in pre-defined static architectures, such as FCN, U-Net, and DeepLab series. This paper studies a conceptually new method to alleviate the scale variance in semantic representation, named dynamic routing. The proposed framework generates data-dependent routes, adapting to the scale distribution of each image. To this end, a differentiable gating function, called soft conditional gate, is proposed to select scale transform paths on the fly. In addition, the computational cost can be further reduced in an end-to-end manner by giving budget constraints to the gating function. We further relax the network level routing space to support multi-path propagations and skip-connections in each forward, bringing substantial network capacity. To demonstrate the superiority of the dynamic property, we compare with several static architectures, which can be modeled as special cases in the routing space. Extensive experiments are conducted on Cityscapes and PASCAL VOC 2012 to illustrate the effectiveness of the dynamic framework. Code is available at https://github.com/yanwei-li/DynamicRouting.

Learning-based approaches to graph matching have been developed and explored for more than a decade, have grown rapidly in scope and popularity in recent years. However, previous learning-based algorithms, with or without deep learning strategy, mainly focus on the learning of node and/or edge affinities generation, and pay less attention on the learning of the combinatorial solver. In this paper we propose a fully trainable framework for graph matching, in which learning of affinities and solving for combinatorial optimization are not explicitly separated as in many previous arts. We firstly convert the problem of building node correspondences between two input graphs to the problem of selecting reliable nodes from a constructed assignment graph. Subsequently, the graph network block module is adopted to perform computation on the graph to form structured representations for each node. It finally predicts a label for each node that is used for node classification, and the training is performed under the supervision of both permutation differences and the one-to-one matching constraints. The proposed method is evaluated on four public benchmarks in comparison with several state-of-the-art algorithms, and the experimental results illustrate its excellent performance.

Training with more data has always been the most stable and effective way of improving performance in deep learn-ing era. As the largest object detection dataset so far, OpenImages brings great opportunities and challenges for object detection in general and sophisticated scenarios. However, owing to its semi-automatic collecting and labeling pipeline to deal with the huge data scale, Open Images dataset suffers from label-related problems that objects may explicitly or implicitly have multiple labels and the label distribution is extremely imbalanced. In this work, we quantitatively analyze these label problems and provide a simple but effective solution. We design a concurrent softmax to handle the multi-label problems in object detection and propose a soft-sampling methods with hybrid training scheduler to deal with the label imbalance. Overall, our method yields a dramatic improvement of 3.34 points, leading to the best single model with 60.90 mAP on the public object detection test set of Open Images. And our ensembling result achieves 67.17mAP, which is 4.29 points higher than the first place method last year.

We study the energy minimization problem in low-level vision tasks from a novel perspective. We replace the heuristic regularization term with a data-driven learnable subspace constraint, and preserve the data term to exploit domain knowledge derived from the first principles of a task. This learning subspace minimization (LSM) framework unifies the network structures and the parameters for many different low-level vision tasks, which allows us to train a single network for multiple tasks simultaneously with shared parameters, and even generalizes the trained network to an unseen task as long as the data term can be formulated. We validate our LSM frame on four low-level tasks including edge detection, interactive segmentation, stereo matching, and optical flow, and validate the network on various datasets. The experiments demonstrate that the proposed LSM generates state-of-the-art results with smaller model size, faster training convergence, and real-time inference.

Temporal camera relocalization estimates the pose with respect to each video frame in sequence, as opposed to one-shot relocalization which focuses on a still image. Even though the time dependency has been taken into account, current temporal relocalization methods still generally underperform the state-of-the-art one-shot approaches in terms of accuracy. In this work, we improve the temporal relocalization method by using a network architecture that incorporates Kalman filtering (KFNet) for online camera relocalization. In particular, KFNet extends the scene coordinate regression problem to the time domain in order to recursively establish 2D and 3D correspondences for the pose determination. The network architecture design and the loss formulation are based on Kalman filtering in the context of Bayesian learning. Extensive experiments on multiple relocalization benchmarks demonstrate the high accuracy of KFNet at the top of both one-shot and temporal relocalization approaches.

When interacting with highly dynamic environments, scene flow allows autonomous systems to reason about the non-rigid motion of multiple independent objects. This is of particular interest in the field of autonomous driving, in which many cars, people, bicycles, and other objects need to be accurately tracked. Current state-of-the-art methods require annotated scene flow data from autonomous driving scenes to train scene flow networks with supervised learning. As an alternative, we present a method of training scene flow that uses two self-supervised losses, based on nearest neighbors and cycle consistency. These self-supervised losses allow us to train our method on large unlabeled autonomous driving datasets; the resulting method matches current state-of-the-art supervised performance using no real world annotations and exceeds state-of-the-art performance when combining our self-supervised approach with supervised learning on a smaller labeled dataset.

Visual dialog is a challenging task that requires the comprehension of the semantic dependencies among implicit visual and textual contexts. This task can refer to the relation inference in a graphical model with sparse contexts and unknown graph structure (relation descriptor), and how to model the underlying context-aware relation inference is critical. To this end, we propose a novel Context-Aware Graph (CAG) neural network. Each node in the graph corresponds to a joint semantic feature, including both object-based (visual) and history-related (textual) context representations. The graph structure (relations in dialog) is iteratively updated using an adaptive top-K message passing mechanism. Specifically, in every message passing step, each node selects the most K relevant nodes, and only receives messages from them. Then, after the update, we impose graph attention on all the nodes to get the final graph embedding and infer the answer. In CAG, each node has dynamic relations in the graph (different related K neighbor nodes), and only the most relevant nodes are attributive to the context-aware relational graph inference. Experimental results on VisDial v0.9 and v1.0 datasets show that CAG outperforms comparative methods. Visualization results further validate the interpretability of our method.

Many visual scenes contain text that carries crucial information, and it is thus essential to understand text in images for downstream reasoning tasks. For example, a deep water label on a warning sign warns people about the danger in the scene. Recent work has explored the TextVQA task that requires reading and understanding text in images to answer a question. However, existing approaches for TextVQA are mostly based on custom pairwise fusion mechanisms between a pair of two modalities and are restricted to a single prediction step by casting TextVQA as a classification task. In this work, we propose a novel model for the TextVQA task based on a multimodal transformer architecture accompanied by a rich representation for text in images. Our model naturally fuses different modalities homogeneously by embedding them into a common semantic space where self-attention is applied to model inter- and intra- modality context. Furthermore, it enables iterative answer decoding with a dynamic pointer network, allowing the model to form an answer through multi-step prediction instead of one-step classification. Our model outperforms existing approaches on three benchmark datasets for the TextVQA task by a large margin.

We propose a deep inverse rendering framework for indoor scenes. From a single RGB image of an arbitrary indoor scene, we obtain a complete scene reconstruction, estimating shape, spatially-varying lighting, and spatially-varying, non-Lambertian surface reflectance. Our novel inverse rendering network incorporates physical insights -- including a spatially-varying spherical Gaussian lighting representation, a differentiable rendering layer to model scene appearance, a cascade structure to iteratively refine the predictions and a bilateral solver for refinement -- allowing us to jointly reason about shape, lighting, and reflectance. Since no existing dataset provides ground truth high quality spatially-varying material and spatially-varying lighting, we propose novel methods to map complex materials to existing indoor scene datasets and a new physically-based GPU renderer to create a large-scale, photorealistic indoor dataset. Experiments show that our framework outperforms previous methods and enables various novel applications like photorealistic object insertion and material editing.

We present an interactive approach to synthesizing realistic variations in facial hair in images, ranging from subtle edits to existing hair to the addition of complex and challenging hair in images of clean-shaven subjects. To circumvent the tedious and computationally expensive tasks of modeling, rendering and compositing the 3D geometry of the target hairstyle using the traditional graphics pipeline, we employ a neural network pipeline that synthesizes realistic and detailed images of facial hair directly in the target image in under one second. The synthesis is controlled by simple and sparse guide strokes from the user defining the general structural and color properties of the target hairstyle. We qualitatively and quantitatively evaluate our chosen method compared to several alternative approaches. We show compelling interactive editing results with a prototype user interface that allows novice users to progressively refine the generated image to match their desired hairstyle, and demonstrate that our approach also allows for flexible and high-fidelity scalp hair synthesis.

Medicine is an important application area for deep learning models. Research in this field is a combination of medical expertise and data science knowledge. In this paper, instead of 2D medical images, we introduce an open-access 3D intracranial aneurysm dataset, IntrA, that makes the application of points-based and mesh-based classification and segmentation models available. Our dataset can be used to diagnose intracranial aneurysms and to extract the neck for a clipping operation in medicine and other areas of deep learning, such as normal estimation and surface reconstruction. We provide a large-scale benchmark of classification and part segmentation by testing state-of-the-art networks. We also discuss the performance of each method and demonstrate the challenges of our dataset. The published dataset can be accessed here: https://github.com/intra2d2019/IntrA.