ECCV 2020 论文列表

Most existing image restoration networks are designed in a disposable way and catastrophically forget previously learned distortions when trained on a new distortion removal task. To alleviate this problem, we raise the novel lifelong image restoration problem for blended distortions. We first design a base fork-join model in which multiple pre-trained expert models specializing in individual distortion removal task work cooperatively and adaptively to handle blended distortions. When the input is degraded by a new distortion, inspired by adult neurogenesis in human memory system, we develop a neural growing strategy where the previously trained model can incorporate a new expert branch and continually accumulate new knowledge without interfering with learned knowledge. Experimental results show that the proposed approach can not only achieve state-of-the-art performance on blended distortions removal tasks in both PSNR/SSIM metrics, but also maintain old expertise while learning new restoration tasks."

In this paper, we advocate the adoption of metric preservation as a powerful prior for learning latent representations of deformable 3D shapes. Key to our construction is the introduction of a geometric distortion criterion, defined directly on the decoded shapes, translating the preservation of the metric on the decoding to the formation of linear paths in the underlying latent space. Our rationale lies in the observation that training samples alone are often insufficient to endow generative models with high fidelity, motivating the need for large training datasets. In contrast, metric preservation provides a rigorous way to control the amount of geometric distortion incurring in the construction of the latent space, leading in turn to synthetic samples of higher quality. We further demonstrate, for the first time, the adoption of differentiable intrinsic distances in the backpropagation of a geodesic loss. Our geometric priors are particularly relevant in the presence of scarce training data, where learning any meaningful latent structure can be especially challenging. The effectiveness and potential of our generative model is showcased in applications of style transfer and content generation using triangle meshes and point clouds."

The ability to understand and interpret human actions is a long-standing challenge and a critical indicator of perception in artificial intelligence. However, a few imperative components of daily human activities are largely missed in prior literature, including the goal-directed actions, concurrent multi-tasks, and collaborations among multi-agents. We introduce the LEMMA dataset to provide a single home to address these missing dimensions with carefully designed settings, wherein the numbers of tasks and agents vary to highlight different learning objectives. We densely annotate the atomic-actions with human-object interactions to provide ground-truth of the compositionality, scheduling, and assignment of daily activities. We further devise challenging compositional action recognition and action/task anticipation benchmarks with baseline models to measure the capability for compositional action understanding and temporal reasoning. We hope this effort inspires the vision community to look into goal-directed human activities and further study the task scheduling and assignment in real-world scenarios."

Recent studies on facial expression editing have obtained very promising progress. On the other hand, existing methods face the constraint of requiring a large amount of expression labels which are often expensive and time-consuming to collect. This paper presents an innovative label-free expression editing via disentanglement (LEED) framework that is capable of editing the expression of both frontal and profile facial images without requiring any expression labels. The idea is to disentangle the identity and expression of a facial image in the expression manifold, where the neutral face captures the identity attribute and the displacement between the neutral image and the expressive image captures the expression attribute. Two novel losses are designed for optimal expression disentanglement and consistent synthesis, including a mutual expression information loss that aims to extract pure expression-related features and a siamese loss that aims to enhance the expression similarity between the synthesized image and the reference image. Extensive experiments over two public facial expression datasets show that LEED achieves superior facial expression editing qualitatively and quantitatively."

To understand movies, humans constantly reason over the dialogues and actions shown in specific scenes and relate them to the overall storyline already seen. Inspired by this behaviour, we design ROLL, a model for knowledge-based video story question answering that leverages three crucial aspects of movie understanding: dialog comprehension, scene reasoning, and storyline recalling. In ROLL, each of these tasks is in charge of extracting rich and diverse information by 1) processing scene dialogues, 2) generating unsupervised video scene descriptions, and 3) obtaining external knowledge in a weakly supervised fashion. To answer a given question correctly, the information generated by each inspired-cognitive task is encoded via Transformers and fused through a modality weighting mechanism, which balances the information from the different sources. Exhaustive evaluation demonstrates the effectiveness of our approach, which yields a new state-of-the-art on two challenging video question answering datasets: KnowIT VQA and TVQA+."

Knowledge Distillation (KD) based methods adopt the one-way Knowledge Transfer (KT) scheme in which training a lower-capacity student network is guided by a pre-trained high-capacity teacher network. Recently, Deep Mutual Learning (DML) presented a two-way KT strategy, showing that the student network can be also helpful to improve the teacher network. In this paper, we propose Dense Cross-layer Mutual-distillation (DCM), an improved two-way KT method in which the teacher and student networks are trained collaboratively from scratch. To augment knowledge representation learning, well-designed auxiliary classifiers are added to certain hidden layers of both teacher and student networks. To boost KT performance, we introduce dense bidirectional KD operations between the layers appended with classifiers. After training, all auxiliary classifiers are discarded, and thus there are no extra parameters introduced to final models. We test our method on a variety of KT tasks, showing its superiorities over related methods. Code is available at https://github.com/sundw2014/DCM"

Knowledge distillation, which involves extracting the ""dark knowledge"" from a teacher network to guide the learning of a student network, has emerged as an important technique for model compression and transfer learning. Unlike previous works that exploit architecture-specific cues such as activation and attention for distillation, here we wish to explore a more general and model-agnostic approach for extracting ""richer dark knowledge"" from the pre-trained teacher model. We show that the seemingly different self-supervision task can serve as a simple yet powerful solution. For example, when performing contrastive learning between transformed entities, the noisy predictions of the teacher network reflect its intrinsic composition of semantic and pose information. By exploiting the similarity between those self-supervision signals as an auxiliary task, one can effectively transfer the hidden information from the teacher to the student. In this paper, we discuss practical ways to exploit those noisy self-supervision signals with selective transfer for distillation. We further show that self-supervision signals improve conventional distillation with substantial gains under few-shot and noisy-label scenarios. Given the richer knowledge mined from self-supervision, our knowledge distillation approach achieves state-of-the-art performance on standard benchmarks, i.e., CIFAR100 and ImageNet, under both similar-architecture and cross-architecture settings. The advantage is even more pronounced under the cross-architecture setting, where our method outperforms the state of the art CRD by an average of 2.3% in accuracy rate on CIFAR100 across six different teacher-student pairs."

Current state-of-the-art trackers rely only on a target appearance model in order to localize the object in each frame. Such approaches are however prone to fail in case of e.g. fast appearance changes or presence of distractor objects, where a target appearance model alone is insufficient for robust tracking. Having the knowledge about the presence and locations of other objects in the surrounding scene can be highly beneficial in such cases. This scene information can be propagated through the sequence and used to, for instance, explicitly avoid distractor objects and eliminate target candidate regions.

Kinship verification is a well-explored task: identifying whether or not two persons are kin. In contrast, kinship identification has been largely ignored so far. Kinship identification aims to further identify the particular type of kinship. An extension to kinship verification run short to properly obtain identification, because existing verification networks are individually trained on specific kinships and do not consider the context between different kinship types. Also, existing kinship verification datasets have biased positive-negative distributions which are different than real-world distributions.

Perceiving the physical world in 3D is fundamental for self-driving applications. Although temporal motion is an invaluable resource to human vision for detection, tracking, and depth perception, such features have not been thoroughly utilized in modern 3D object detectors. In this work, we propose a novel method for monocular video-based 3D object detection which carefully leverages kinematic motion to improve precision of 3D localization. Specifically, we first propose a novel decomposition of object orientation as well as a self-balancing 3D confidence. We show that both components are critical to enable our kinematic model to work effectively. Collectively, using only a single model, we efficiently leverage 3D kinematics from monocular videos to improve the overall localization precision in 3D object detection while also producing useful by-products of scene dynamics (ego-motion and per-object velocity). We achieve state-of-the-art performance on monocular 3D object detection and the Bird's Eye View tasks within the KITTI self-driving dataset. "

Human pose estimation in video relies on local information by either estimating each frame independently or tracking poses across frames. In this paper, we propose a novel method combining local approaches with global context. We introduce a light weighted, unsupervised, key-frame proposal network (K-FPN) to select informative frames and a learned dictionary to recover the entire pose sequence from these frames. The K-FPN speeds up the pose estimation and provides robustness to bad frames with occlusion, motion blur, and illumination changes, while the learned dictionary provides global dynamic context. Experiments on Penn Action and sub-JHMDB datasets show that the proposed method achieves state-of-the-art accuracy, with substantial speed-up."

Semi-supervised video object segmentation (VOS) is a task that involves predicting a target object in a video when the ground truth segmentation mask of the target object is given in the first frame. Recently, space-time memory networks (STM) have received significant attention as a promising solution for semi-supervised VOS. However, an important point is overlooked when applying STM to VOS. The solution (STM) is non-local, but the problem (VOS) is predominantly local. To solve the mismatch between STM and VOS, we propose a kernelized memory network (KMN). Before being trained on real videos, our KMN is pre-trained on static images, as in previous works. Unlike in previous works, we use the Hide-and-Seek strategy in pre-training to obtain the best possible results in handling occlusions and segment boundary extraction. The proposed KMN surpasses the state-of-the-art on standard benchmarks by a significant margin (+5% on DAVIS 2017 test-dev set). In addition, the runtime of KMN is 0.12 seconds per frame on the DAVIS 2016 validation set, and the KMN rarely requires extra computation, when compared with STM."

Recent works in single-image perceptual super-resolution (SR) have demonstrated unprecedented performance in generating realistic textures by means of deep convolutional networks. However, these convolutional models are large and expensive, preventing them from being deployed to devices that require them. In this work, we propose a neural architecture search (NAS) approach that integrates NAS and generative adversarial networks (GANs) with recent advances in perceptual SR and pushes the efficiency of small perceptual SR models to facilitate on-device execution. Specifically, we search over the architectures of both the generator and the discriminator sequentially, highlighting the unique challenges and key observations of searching for an SR-optimised discriminator and comparing them with existing discriminator architectures in the literature. Our tiny perceptual SR (TPSR) models outperform SRGAN and EnhanceNet on both full-reference perceptual metric (LPIPS) and distortion metric (PSNR) while being up to 26.4$ imes$ more memory efficient and 33.6$ imes$ more compute efficient respectively."

We propose the first network to jointly learn visual motion and confidence from events in spatially local patches. Event-based sensors deliver high temporal resolution motion information in a sparse, non-redundant format. This creates the potential for low computation, low latency motion recognition. Neural networks which extract global motion information, however, are generally computationally expensive. Here, we introduce a novel shallow and compact neural architecture and learning approach to capture reliable visual motion information along with the corresponding confidence of inference. Our network makes a prediction of the visual motion at each spatial location using only local events. Our confidence network then identifies which of these predictions will be accurate. In the task of recovering pan-tilt ego velocities from events, we show that each individual confident local prediction of our network can be expected to be as accurate as state of the art optimization approaches which utilize the full image. Furthermore, on a publicly available dataset, we find our local predictions generalize to scenes with camera motions and the presence of independently moving objects. This makes the output of our network well suited for motion based tasks, such as the segmentation of independently moving objects. We demonstrate on a publicly available motion segmentation dataset that restricting predictions to confident regions is sufficient to achieve results that exceed state of the art methods. "

We address the problem of bias in automated face recognition and demographic attribute estimation algorithms, where errors are lower on certain cohorts belonging to specific demographic groups. We present a novel de-biasing adversarial network (DebFace) that learns to extract disentangled feature representations for both unbiased face recognition and demographics estimation. The proposed network consists of one identity classifier and three demographic classifiers (for gender, age, and race) that are trained to distinguish identity and demographic attributes, respectively. Adversarial learning is adopted to minimize correlation among feature factors so as to abate bias influence from other factors. We also design a new scheme to combine demographics with identity features to strengthen robustness of face representation in different demographic groups. The experimental results show that our approach is able to reduce bias in face recognition as well as demographics estimation while achieving state-of-the-art performance."

Unsupervised domain adaptive person Re-IDentification (ReID) is challenging because of the large domain gap between source and target domains, as well as the lackage of labeled data on the target domain. This paper tackles this challenge through jointly enforcing visual and temporal consistency in the combination of a local one-hot classification and a global multi-class classification. The local one-hot classification assigns images in a training batch with different person IDs, then adopts a Self-Adaptive Classification (SAC) model to classify them. The global multi-class classification is achieved by predicting labels on the entire unlabeled training set with the Memory-based Temporal-guided Cluster (MTC). MTC predicts multi-class labels by considering both visual similarity and temporal consistency to ensure the quality of label prediction. The two classification models are combined in a unified framework, which effectively leverages the unlabeled data for discriminative feature learning. Experimental results on three large-scale ReID datasets demonstrate the superiority of proposed method in both unsupervised and unsupervised domain adaptive ReID tasks. For example, under unsupervised setting, our method outperforms recent unsupervised domain adaptive methods, which leverage more labels for training."

Semantic instance segmentation is the task of simultaneously partitioning an image into distinct segments while associating each pixel with a class label. In commonly used pipelines, segmentation and label assignment are solved separately since joint optimization is computationally expensive. We propose a greedy algorithm for joint graph partitioning and labeling derived from the efficient Mutex Watershed partitioning algorithm. It optimizes an objective function closely related to the Symmetric Multiway Cut objective and empirically shows efficient scaling behavior. Due to the algorithm's efficiency, it can operate directly on pixels without prior over-segmentation of the image into superpixels. We evaluate the performance on the Cityscapes dataset (2D urban scenes) and on a 3D microscopy volume. In urban scenes, the proposed algorithm combined with current deep neural networks outperforms the strong baseline of `Panoptic Feature Pyramid Networks' by Kirillov et al. (2019). In the 3D electron microscopy images, we show explicitly that our joint formulation outperforms a separate optimization of the partitioning and labeling problems. "

We propose a markerless end-to-end training framework for parametric 3D human shape models. The training of statistical 3D human shape models with minimal supervision is an important problem in computer vision. Contrary to prior work, the whole training process (i) uses a differentiable shape model surface and (ii) is trained end-to-end by jointly optimizing all parameters of a single, self-contained objective that can be solved with slightly modified off-the-shelf non-linear least squares solvers. The training process only requires a compact model definition and an off-the-shelf 2D RGB pose estimator. No pre-trained shape models are required. For training (iii) a medium-sized dataset of approximately 1000 low-resolution human body scans is sufficient to achieve competitive performance on the challenging FAUST surface correspondence benchmark. The training and evaluation code will be made available for research purposes to facilitate end-to-end shape model training on novel datasets with minimal setup cost."

Individuals Action and Sub-group Activities in Videos","The state-of-the art solutions for human activity understanding from a video stream formulate the task as a spatio-temporal problem which requires joint localization of all individuals in the scene and classification of their actions or group activity over time. Who is interacting with whom, e.g. not everyone in a queue is interacting with each other, is often not predicted. There are scenarios where people are best to be split into sub-groups, which we call social groups, and each social group may be engaged in a different social activity. In this paper, we solve the problem of simultaneously grouping people by their social interactions, predicting their individual actions and the social activity of each social group, which we call the social task. Our main contributions are: i) we propose an end-to-end trainable framework for the social task; ii) our proposed method also sets the state-of-the-art results on two widely adopted benchmarks for the traditional group activity recognition task (assuming individuals of the scene form a single group and predicting a single group activity label for the scene); iii) we introduce new annotations on an existing group activity dataset, re-purposing it for the social task."