ECCV 2020 论文列表

Recent works in deep-learning have shown that second-order information is beneficial in many computer-vision tasks. Second-order information can be enforced both in the spatial context and the abstract feature dimensions. In this work, we explore two second-order components. One is focused on second-order spatial information to increase the performance of image descriptors, both local and global. It is used to re-weight feature maps, and thus emphasise salient image locations that are subsequently used for description. The second component is concerned with a second-order similarity (SOS) loss, that we extend to global descriptors for image retrieval, and is used to enhance the triplet loss with hard-negative mining. We validate our approach on two different tasks and datasets for image retrieval and image matching. The results show that our two second-order components complement each other, bringing significant performance improvements in both tasks and lead to state-of-the-art results across the public benchmarks. Code available at: http://github.com/tonyngjichun/SOLAR"

Dense Video Captioning (DVC) is a challenging task that localizes all events in a short video and describes them with natural language sentences. The main goal of DVC is video story description, that is, to generate a concise video story that supports human video comprehension without watching it. In recent years, DVC has attracted increasing attention in the research community of Vision and Language, and has been employed as a task of the workshop, ActivityNet Challenge. In the current research community, the official scorer provided by AcivityNet Challenge is the de-facto standard evaluation framework for DVC systems. It computes averaged METEOR scores for matched pairs between generated and reference captions whose Interval of Union (IoU) exceeds a specific threshold value. However, the current framework does not take into account the story of the video, the ordering of captions. It also tends to give high scores to systems that generate redundant, several hundred captions, that humans cannot read. This paper proposes a new evaluation framework, Story Oriented Dense video cAptioning evaluation framework (SODA), for measuring the performance of video story description systems. SODA first tries to find temporally optimal matching between generated and reference captions to capture a story for a video. Then, it computes METEOR scores for the matching and derives F-measure scores from the METEOR scores to penalize redundant captions. To demonstrate that SODA gives low scores for inadequate captions in terms of video story description, we evaluate two state-of-the-art systems with it, varying the number of captions. The results show that SODA can give low scores against too many or too few captions and high scores against captions whose number equals to that of a reference, while the current framework gives good scores for all the cases. Furthermore, we show that SODA tends to give lower scores than the current evaluation framework in evaluating captions with incorrect order."

Freespace detection is an essential component of visual perception for self-driving cars. The recent efforts made in data-fusion convolutional neural networks (CNNs) have significantly improved semantic driving scene segmentation. Freespace can be hypothesized as a ground plane, on which the points have similar surface normals. Hence, in this paper, we first introduce a novel module, named surface normal estimator (SNE), which can infer surface normal information from dense depth/disparity images with high accuracy and efficiency. Furthermore, we propose a data-fusion CNN architecture, referred to as RoadSeg, which can extract and fuse features from both RGB images and the inferred surface normal information for accurate freespace detection. For research purposes, we publish a large-scale synthetic freespace detection dataset, named Ready-to-Drive (R2D) road dataset, collected under different illumination and weather conditions. The experimental results demonstrate that our proposed SNE module can benefit all the state-of-the-art CNNs for freespace detection, and our SNE-RoadSeg achieves the best overall performance among different datasets."

We propose advances that address two key challenges in future trajectory prediction: (i) multimodality in both training data and predictions and (ii) constant time inference regardless of number of agents. Existing trajectory predictions are fundamentally limited by lack of diversity in training data, which is difficult to acquire with sufficient coverage of possible modes. Our first contribution is an automatic method to simulate diverse trajectories in the top-view. It uses pre-existing datasets and maps as initialization, mines existing trajectories to represent realistic driving behaviors and uses a multi-agent vehicle dynamics simulator to generate diverse new trajectories that cover various modes and are consistent with scene layout constraints. Our second contribution is a novel method that generates diverse predictions while accounting for scene semantics and multi-agent interactions, with constant-time inference independent of the number of agents. We propose a convLSTM with novel state pooling operations and losses to predict scene-consistent states of multiple agents in a single forward pass, along with a CVAE for diversity. We validate our proposed multi-agent trajectory prediction approach by training and testing on the proposed simulated dataset and existing real datasets of traffic scenes. In both cases, our approach outperforms SOTA methods by a large margin, highlighting the benefits of both our diverse dataset simulation and constant-time diverse trajectory prediction methods."

Recovering multi-person 3D poses with absolute scales from a single RGB image is a challenging problem due to the inherent depth and scale ambiguity from a single view. Addressing this ambiguity requires to aggregate various cues over the entire image, such as body sizes, scene layouts, and inter-person relationships. However, most previous methods adopt a top-down scheme that first performs 2D pose detection and then regresses the 3D pose and scale for each detected person individually, ignoring global contextual cues. In this paper, we propose a novel system that first regresses a set of 2.5D representations of body parts and then reconstructs the 3D absolute poses based on these 2.5D representations with a depth-aware part association algorithm. Such a single-shot bottom-up scheme allows the system to better learn and reason about the inter-person depth relationship, improving both 3D and 2D pose estimation. The experiments demonstrate that the proposed approach achieves the state-of-the-art performance on the CMU Panoptic and MuPoTS-3D datasets and is applicable to in-the-wild videos."

While models of 3D clothing learned from real data exist, no method can predict clothing deformation as a function of garment size. In this paper, we introduce SizerNet to predict 3D clothing conditioned on human body shape and garment size parameters, and ParserNet to infer garment meshes and shape under clothing with personal details in a single pass from an input mesh. SizerNet allows to estimate and visualize the dressing effect of a garment in various sizes, and ParserNet allows to edit clothing of an input mesh directly, removing the need for scan segmentation, which is a challenging problem in itself. To learn these models, we introduce the SIZER dataset of clothing size variation which includes 100 different subjects wearing casual clothing items in various sizes, totaling to approximately 2000 scans. This dataset includes the scans, registrations to the SMPL model, scans segmented in clothing parts, garment category and size labels. Our experiments show better parsing accuracy and size prediction than baseline methods trained on SIZER. The code, model and dataset will be released for research purposes at: https://virtualhumans.mpi-inf.mpg.de/sizer/."

Deep generative models have been used in recent years to learn coherent latent representations in order to synthesize high-quality images. In this work, we propose a neural network to learn a generative model for sampling consistent indoor scene layouts. Our method learns the co-occurrences, and appearance parameters such as shape and pose, for different objects categories through a grammar-based auto-encoder, resulting in a compact and accurate representation for scene layouts. In contrast to existing grammar-based methods with a user-specified grammar, we construct the grammar automatically by extracting a set of production rules on reasoning about object co-occurrences in training data. The extracted grammar is able to represent a scene by an augmented parse tree. The proposed auto-encoder encodes these parse trees to a latent code, and decodes the latent code to a parse tree, thereby ensuring the generated scene is always valid. We experimentally demonstrate that the proposed auto-encoder learns not only to generate valid scenes (i.e. the arrangements and appearances of objects), but it also learns coherent latent representations where nearby latent samples decode to similar scene outputs. The obtained generative model is applicable to several computer vision tasks such as 3D pose and layout estimation from RGB-D data. "

In this paper, we study an intermediate form of supervision, i.e., single-frame supervision, for temporal action localization (TAL). To obtain the single-frame supervision, the annotators are asked to identify only a single frame within the temporal window of an action. This can significantly reduce the labor cost of obtaining full supervision which requires annotating the action boundary. Compared to the weak supervision that only annotates the video-level label, the single-frame supervision introduces extra temporal action signals while maintaining low annotation overhead. To make full use of such single-frame supervision, we propose a unified system called SF-Net. First, we propose to predict an actionness score for each video frame. Along with a typical category score, the actionness score can provide comprehensive information about the occurrence of a potential action and aid the temporal boundary refinement during inference. Second, we mine pseudo action and background frames based on the single-frame annotations. We identify pseudo action frames by adaptively expanding each annotated single frame to its nearby, contextual frames and we mine pseudo background frames from all the unannotated frames across multiple videos. Together with the ground-truth labeled frames, these pseudo-labeled frames are further used for training the classifier."

Manipulating or Erasing Objects","Recent advances in image generation gave rise to powerful tools for semantic image editing. However, existing approaches can either operate on a single image or require an abundance of additional information. They are not capable of handling the complete set of editing operations, that is addition, manipulation or removal of semantic concepts. To address these limitations, we propose SESAME, a novel generator-discriminator pair for Semantic Editing of Scenes by Adding, Manipulating or Erasing objects. In our setup, the user provides the semantic labels of the areas to be edited and the generator synthesizes the corresponding pixels. In contrast to previous methods that employ a discriminator that concatenates semantics and image as an input, the SESAME discriminator is composed of two input streams that independently process the image and its semantics, using the latter to manipulate the results of the former. We evaluate our model on a diverse set of datasets and report state-of-the-art performance on two tasks: (a) image manipulation and (b) image generation conditioned on semantic labels."

In this paper, we equip Prototypical Networks (PNs) with a novel dissimilarity measure to enable discriminative feature normalization for few-shot learning. The embedding onto the hypersphere requires no direct normalization and is easy to optimize. Our theoretical analysis shows that the proposed dissimilarity measure, denoted the Squared root of the Euclidean distance and the Norm distance (SEN), forces embedding points to be attracted to its correct prototype, while being repelled from all other prototypes, keeping the norm of all points the same. The resulting SEN PN outperforms the regular PN with a considerable margin, with no additional parameters as well as with negligible computational overhead."

Can we automatically group images into semantically meaningful clusters when ground-truth annotations are absent? The task of unsupervised image classification remains an important, and open challenge in computer vision. Several recent approaches have tried to tackle this problem in an end-to-end fashion. In this paper, we deviate from recent works, and advocate a two-step approach where feature learning and clustering are decoupled. First, a self-supervised task from representation learning is employed to obtain semantically meaningful features. Second, we use the obtained features as a prior in a learnable clustering approach. In doing so, we remove the ability for cluster learning to depend on low-level features, which is present in current end-to-end learning approaches. Experimental evaluation shows that we outperform state-of-the-art methods by large margins, in particular +26.6% on CIFAR10, +25.0% on CIFAR100-20 and +21.3% on STL10 in terms of classification accuracy. Furthermore, our method is the first to perform well on a large-scale dataset for image classification. In particular, we obtain promising results on ImageNet, and outperform several semi-supervised learning methods in the low-data regime without the use of any ground-truth annotations."

Recently, cybersickness assessment for VR content is required to deal with viewing safety issues. Assessing physical symptoms of individual viewers is challenging but important to provide detailed and personalized guides for viewing safety. In this paper, we propose a novel symptom-aware cybersickness assessment network (SACA Net) that quantifies physical symptom levels for assessing cybersickness of individual viewers. The SACA Net is designed to utilize the relational characteristics of symptoms for complementary effects among relevant symptoms. The proposed network consists of three main parts: a stimulus symptom context guider, a physiological symptom guider, and a symptom relation embedder. The stimulus symptom context guider and the physiological symptom guider extract symptom features from VR content and human physiology, respectively. The symptom relation embedder refines the stimulus-response symptom features to effectively predict cybersickness by embedding relational characteristics with graph formulation. For validation, we utilize two public 360-degree video datasets that contain cybersickness scores and physiological signals. Experimental results show that the proposed method is effective in predicting human cybersickness with physical symptoms. Further, latent relations among symptoms are interpretable by analyzing relational weights in the proposed network."

Establishing robust and accurate correspondences is a fundamental backbone to many computer vision algorithms. While recent learning-based feature matching methods have shown promising results in providing robust correspondences under challenging conditions, they are often limited in terms of precision. In this paper, we introduce S2DNet, a novel feature matching pipeline, designed and trained to efficiently establish both robust and accurate correspondences. By leveraging a sparse-to-dense matching paradigm, we cast the correspondence learning problem as a supervised classification task to learn to output highly peaked correspondence maps. We show that S2DNet achieves state-of-the-art results on the HPatches benchmark, as well as on several long-term visual localization datasets. "

Recently, dynamic inference has emerged as a promising way to reduce the computational cost of deep convolutional neural networks (CNNs). In contrast to static methods (e.g., weight pruning), dynamic inference adaptively adjusts the inference process according to each input sample, which can considerably reduce the computational cost on ""easy"" samples while maintaining the overall model performance. In this paper, we introduce a general framework, S2DNAS, which can transform various static CNN models to support dynamic inference via neural architecture search. To this end, based on a given CNN model, we first generate a CNN architecture space in which each architecture is a multi-stage CNN generated from the given model using some predefined transformations. Then, we propose a reinforcement learning based approach to automatically search for the optimal CNN architecture in the generated space. At last, with the searched multi-stage network, we can perform dynamic inference by adaptively choosing a stage to evaluate for each sample. Unlike previous works that introduce irregular computations or complex controllers in the inference or re-design a CNN model from scratch, our method can generalize to most of the popular CNN architectures and the searched dynamic network can be directly deployed using existing deep learning frameworks in various hardware devices."

Video action recognition is a complex task dependent on modeling spatial and temporal context. Standard approaches rely on 2D or 3D convolutions to process such context, resulting in expensive operations with millions of parameters. Recent efficient architectures leverage a channel-wise shift-based primitive as a replacement for temporal convolutions, but remain bottlenecked by spatial convolution operations to maintain strong accuracy and a fixed-shift scheme. Naively extending such developments to a 3D setting is a difficult, intractable goal. To this end, we introduce RubiksNet, a new efficient architecture for video action recognition which is based on a proposed learnable 3D spatiotemporal shift operation instead. We analyze the suitability of our new primitive for video action recognition and explore several novel variations of our approach to enable stronger representational flexibility while maintaining an efficient design. We benchmark our approach on several standard video recognition datasets, and observe that our method achieves comparable or better accuracy than prior work on efficient video action recognition at a fraction of the performance cost, with 2.9-5.9x fewer parameters and 2.1-3.7x fewer FLOPs. We also perform a series of controlled ablation studies to verify our significant boost in the efficiency-accuracy tradeoff curve is rooted in the core contributions of our RubiksNet architecture. "

Novel view video synthesis aims to synthesize novel viewpoints videos given input captures of a human performance taken from multiple reference viewpoints and over consecutive time steps. Despite great advances in model-free novel view synthesis, existing methods present three limitations when applied to complex and time-varying human performance. First, these methods (and related datasets) mainly consider simple and symmetric objects. Second, they do not enforce explicit consistency across generated views. Third, they focus on static and non-moving objects. The fine-grained details of a human subject can therefore suffer from inconsistencies when synthesized across different viewpoints or time steps. To tackle these challenges, we introduce a human-specific framework that employs a learned 3D-aware representation. Specifically, we first introduce a novel siamese network that employs a gating layer for better reconstruction of the latent volumetric representation and, consequently, final visual results. Moreover, features from consecutive time steps are shared inside the network to improve temporal consistency. Second, we introduce a novel loss to explicitly enforce consistency across generated views both in space and in time. Third, we present the Multi-View Human Action (MVHA) dataset, consisting of near 1200 synthetic human performance captured from 54 viewpoints. Experiments on the MVHA, Pose-Varying Human Model and ShapeNet datasets show that our method outperforms the state-of-the-art baselines both in view generation quality and spatio-temporal consistency."

In the last few years, there has been an increasing trend to consider Structure from Motion (SfM, in computer vision) and Simultaneous Localization and Mapping (SLAM, in robotics) problems from the point of view of pose averaging (also known as global SfM, in computer vision) or Pose Graph Optimization (PGO, in robotics), where the motion of the camera is reconstructed by considering only relative rigid body transformations instead of including also 3-D points (as done in a full Bundle Adjustment). At a high level, the advantage of this approach is that modern solvers can effectively avoid most of the problems of local minima, and that it is easier to reason about outlier poses (caused by feature mismatches and repetitive structures in the images). In this paper, we contribute to the state of the art of the latter, by proposing a method to detect incorrect orientation measurements prior to pose graph optimization by checking the geometric consistency of rotation measurements. The novel aspects of our method are the use of Expectation-Maximization to fine-tune the covariance of the noise in inlier measurements, and a new approximate graph inference procedure, of independent interest, that is specifically designed to take advantage of evidence on cycles with better performance than standard approaches (Belief Propagation). The paper includes simulation and experimental results that evaluate the performance of our outlier detection and cycle-based inference algorithms on synthetic and real-world data."

In this paper, we propose a Rotation-robust Intersection over Union ($ extit{RIoU}$) for 3D object detection, which aims to jointly learn the overlap of rotated bounding boxes. In most existing 3D object detection methods, the norm-based loss is adopted to individually regress the parameters of bounding boxes, which may suffer from the loss-metric mismatch due to the scaling problem. Motivated by the IoU loss in the axis-aligned 2D object detection which is invariant to the scale, our method jointly optimizes the parameters via the $ extit{RIoU}$ loss. To tackle the uncertainty of convex caused by rotation, a projection operation is defined to estimate the intersection area. The calculation process of $ extit{RIoU}$ and its loss function is robust to the rotation condition and feasible for back-propagation, which only comprises basic numerical operations. By incorporating the $ extit{RIoU}$ loss with the conventional norm-based loss function, we enforce the network to directly optimize the $ extit{RIoU}$ . Experimental results on the KITTI and nuScenes datasets validate the effectiveness of our proposed method. Moreover, we show that our method is suitable for the detection task of 2D rotated objects, such as text boxes and cluttered targets in the aerial images. "

The attention-based encoder-decoder framework has recently achieved impressive results for scene text recognition, and many variants have emerged with improvements in recognition quality. However, it performs poorly on contextless texts (e.g., random character sequences) which is unacceptable in most of real application scenarios. In this paper, we first deeply investigate the decoding process of the decoder. We empirically find that a representative character-level sequence decoder utilizes not only context information but also positional information. Contextual information, which the existing approaches heavily rely on, causes the problem of attention drift. To suppress such side-effect, we propose a novel position enhancement branch, and dynamically fuse its outputs with those of the decoder attention module for scene text recognition. Specifically, it contains a position aware module to enable the encoder to output feature vectors encoding their own spatial positions, and an attention module to estimate glimpses using the positional clue (i.e., the current decoding time step) only. The dynamic fusion is conducted for more robust feature via an element-wise gate mechanism. Theoretically, our proposed method, dubbed RobustScanner, decodes individual characters with dynamic ratio between context and positional clues, and utilizes more positional ones when the decoding sequences with scarce context, and thus is robust and practical. Empirically, it has achieved new state-of-the-art results on popular regular and irregular text recognition benchmarks while without much performance drop on contextless benchmarks, validating its robustness in both contextual and contextless application scenarios."