ECCV 2018 论文列表

This paper studies the object transfiguration problem in wild images. The generative network in classical GANs for object transfiguration often undertakes a dual responsibility: to detect the objects of interests and to convert the object from source domain to another domain. In contrast, we decompose the generative network into two separated networks, each of which is only dedicated to one particular sub-task. The attention network predicts spatial attention maps of images, and the transformation network focuses on translating objects. Attention maps produced by attention network are encouraged to be sparse, so that major attention can be paid on objects of interests. No matter before or after object transfiguration, attention maps should remain constant. In addition, learning attention network can receive more instruction, given the available segmentation annotations of images. Experimental results demonstrate the necessity of investigating attention in object transfiguration, and that the proposed algorithm can learn accurate attention to improve quality of generated images.

We propose a novel attention mechanism to enhance Convolutional Neural Networks for fine-grained recognition. It learns to attend to lower-level feature activations without requiring part annotations and uses these activations to update and rectify the output likelihood distribution. In contrast to other approaches, the proposed mechanism is modular, architecture-independent and efficient both in terms of parameters and computation required. Experiments show that networks augmented with our approach systematically improve their classification accuracy and become more robust to clutter. As a result, Wide Residual Networks augmented with our proposal surpasses the state of the art classification accuracies in CIFAR-10, the Adience gender recognition task, Stanford dogs, and UEC Food-100.

We present a method that leverages the complementarity of event cameras and standard cameras to track visual features with low-latency. Event cameras are novel sensors that output pixel-level brightness changes, called "events". They offer significant advantages over standard cameras, namely a very high dynamic range, no motion blur, and a latency in the order of microseconds. However, because the same scene pattern can produce different events depending on the motion direction, establishing event correspondences across time is challenging. By contrast, standard cameras provide intensity measurements (frames) that do not depend on motion direction. Our method extracts features on frames and subsequently tracks them asynchronously using events, thereby exploiting the best of both types of data: the frames provide a photometric representation that does not depend on motion direction and the events provide low-latency updates. In contrast to previous works, which are based on heuristics, this is the first principled method that uses raw intensity measurements directly, based on a generative event model within a maximum-likelihood framework. As a result, our method produces feature tracks that are both more accurate (subpixel accuracy) and longer than the state of the art, across a wide variety of scenes.

Effectively bridging between image level keyword annotations and corresponding image pixels is one of the main challenges in weakly supervised semantic segmentation. In this paper, we use an instance-level salient object detector to automatically generate salient instances (candidate objects) for training images. Using similarity features extracted from each salient instance in the whole training set, we build a similarity graph, then use a graph partitioning algorithm to separate it into multiple subgraphs, each of which is associated with a single keyword (tag). Our graph-partitioning-based clustering algorithm allows us to consider the relationships between all salient instances in the training set as well as the information within them. We further show that with the help of attention information, our clustering algorithm is able to correct certain wrong assignments, leading to more accurate results. The proposed framework is general, and any state-of-the-art fully-supervised network structure can be incorporated to learn the segmentation network. When working with DeepLab for semantic segmentation, our method outperforms state-of-the-art weakly supervised alternatives by a large margin, achieving 65.6% mIoU on the PASCAL VOC 2012 dataset. We also combine our method with Mask R-CNN for instance segmentation, and demonstrated for the first time the ability of weakly supervised instance segmentation using only keyword annotations.

Deep learning models are susceptible to input specific noise, called adversarial perturbations. Moreover, there exist input-agnostic noise, called Universal Adversarial Perturbations (UAP) that can affect inference of the models over most input samples. Given a model, there exist broadly two approaches to craft UAPs: (i) data-driven: that require data, and (ii) data-free: that do not require data samples. Data-driven approaches require actual samples from the underlying data distribution and craft UAPs with high success (fooling) rate. However, data-free approaches craft UAPs without utilizing any data samples and therefore result in lesser success rates. In this paper, for data-free scenarios, we propose a novel approach that emulates the effect of data samples with class impressions in order to craft UAPs using data-driven objectives. Class impression for a given pair of category and model is a generic representation (in the input space) of the samples belonging to that category. Further, we present a neural network based generative model that utilizes the acquired class impressions to learn crafting UAPs. Experimental evaluation demonstrates that the learned generative model, (i) readily crafts UAPs via simple feed-forwarding through neural network layers, and (ii) achieves state-of-the-art success rates for data-free scenario and closer to that for data-driven setting without actually utilizing any data samples.

This paper proposes a real-time dynamic scene reconstruction method capable of reproducing the motion, geometry, and segmentation simultaneously given live depth stream from a single RGB-D camera. Our approach fuses geometry frame by frame and uses a segmentation-enhanced node graph structure to drive the deformation of geometry in registration step. A two-level node motion optimization is proposed. The optimization space of node motions and the range of physically-plausible deformations are largely reduced by taking advantage of the articulated motion prior, which is solved by an efficient node graph segmentation method. Compared to previous fusion-based dynamic scene reconstruction methods, our experiments show robust and improved reconstruction results for tangential and occluded motions.

Eye gaze estimation has been increasingly demanded by recent intelligent systems to accomplish a range of interaction-related tasks, by using simple eye images as input. However, learning the highly complex regression between eye images and gaze directions is nontrivial, and thus the problem is yet to be solved efficiently. In this paper, we propose the Asymmetric Regression-Evaluation Network (ARE-Net), and try to improve the gaze estimation performance to its full extent. At the core of our method is the notion of ``two eye asymmetry'' observed during gaze estimation for the left and right eyes. Inspired by this, we design the multi-stream ARE-Net; one asymmetric regression network (AR-Net) predicts 3D gaze directions for both eyes with a novel asymmetric strategy, and the evaluation network (E-Net) adaptively adjusts the strategy by evaluating the two eyes in terms of their performance during optimization. By training the whole network, our method achieves promising results and surpasses the state-of-the-art methods on multiple public datasets.

Recent capture technologies and methods allow not only to retrieve 3D model sequence of moving people in clothing, but also to separate and extract the underlying body geometry, motion component and the clothing as a geometric layer. So far this clothing layer has only been used as raw offsets for individual applications such as retargeting a different body capture sequence with the clothing layer of another sequence, with limited scope, e.g. using identical or similar motions. The structured, semantics and motion-correlated nature of the information contained in this layer has yet to be fully understood and exploited. To this purpose we propose a comprehensive analysis of the statistics of this layer with a simple two-component model, based on PCA subspace reduction of the layer information on one hand, and a generic parameter regression model using neural networks on the other hand, designed to regress from any semantic parameter whose variation is observed in a training set, to the layer parameterization space. We show that this model not only allows to reproduce previous retargeting works, but generalizes the data generation capabilities to other semantic parameters such as clothing variation and size, or physical material parameters with synthetically generated training sequence, paving the way for many kinds of capture data-driven creation and augmentation applications.

While deep neural networks have demonstrated competitive results for many visual recognition and image retrieval tasks, the major challenge lies in distinguishing similar images from different categories (i.e., hard negative examples) while clustering images with large variations from the same category (i.e., hard positive examples). The current state-of-the-art is to mine the most hard triplet examples from the mini-batch to train the network. However, mining-based methods tend to look into these triplets that are hard in terms of the current estimated network, rather than deliberately generating those hard triplets that really matter in globally optimizing the network. For this purpose, we propose an adversarial network for Hard Triplet Generation (HTG) to optimize the network ability in distinguishing similar examples of different categories as well as grouping varied examples of the same categories. We evaluate our method on the real-world challenging datasets, such as CUB200-2011, CARS196, DeepFashion and VehicleID datasets, and show that our method outperforms the state-of-the-art methods significantly.

We present an instance segmentation scheme based on pixel affinity information, which is the relationship of two pixels belonging to a same instance. In our scheme, we use two neural networks with similar structure. One is to predict pixel level semantic score and the other is designed to derive pixel affinities. Regarding pixels as the vertexes and affinities as edges, we then propose a simple yet effective graph merge algorithm to cluster pixels into instances. Experimental results show that our scheme can generate fine grained instance mask. With Cityscapes training data, the proposed scheme achieves 27.3 AP on test set.

This paper presents a novel algorithm to estimate the relative pose, i.e. the 3D rotation and translation of two cameras, from two affine correspondences (ACs) considering any central camera model. The solver is built on new epipolar constraints describing the relationship of an AC and any central views. We also show that the pinhole case is a specialization of the proposed approach. Benefiting from the low number of required correspondences, robust estimators like LO-RANSAC need fewer samples, and thus terminate earlier than using the five-point method. Tests on publicly available datasets containing pinhole, fisheye and catadioptric camera images confirmed that the method often leads to results superior to the state-of-the-art in terms of geometric accuracy.

In a typical real-world application of re-id, a watch-list (gallery set) of a handful of target people (e.g. suspects) to track around a large volume of non-target people are demanded across camera views, and this is called the open-world person re-id. Different from conventional (closed-world) person re-id, a large portion of probe samples are not from target people in the open-world setting. And, it always happens that a non-target person would look similar to a target one and therefore would seriously challenge a re-id system. In this work, we introduce a deep open-world group-based person re-id model based on adversarial learning to alleviate the attack problem caused by similar non-target people. The main idea is learning to attack feature extractor on the target people by using GAN to generate very target-like images (imposters), and in the meantime the model will make the feature extractor learn to tolerate the attack by discriminative learning so as to realize group-based verification. The framework we proposed is called the adversarial open-world person re-identification, and this is realized by our Adversarial PersonNet (APN) that jointly learns a generator, a person discriminator, a target discriminator and a feature extractor, where the feature extractor and target discriminator share the same weights so as to makes the feature extractor learn to tolerate the attack by imposters for better group-based verification. While open-world person re-id is challenging, we show for the first time that the adversarial-based approach helps stabilize person re-id system under imposter attack more effectively.

We explore an approach to forecasting human motion in a few milliseconds given an input 3D skeleton sequence based on a recurrent encoder-decoder framework. Current approaches suffer from the problem of prediction discontinuities and may fail to predict human-like motion in longer time horizons due to error accumulation. We address these critical issues by incorporating local geometric structure constraints and regularizing predictions with plausible temporal smoothness and continuity from a global perspective. Specifically, rather than using the conventional Euclidean loss, we propose a novel frame-wise geodesic loss as a geometrically meaningful, more precise distance measurement. Moreover, inspired by the adversarial training mechanism, we present a new learning procedure to simultaneously validate the sequence-level plausibility of the prediction and its coherence with the input sequence by introducing two global recurrent discriminators. An unconditional, fidelity discriminator and a conditional, continuity discriminator are jointly trained along with the predictor in an adversarial manner. Our resulting adversarial geometry-aware encoder-decoder (AGED) model significantly outperforms state-of-the-art deep learning based approaches on the heavily benchmarked H3.6M dataset in both short-term and long-term predictions.

Recurrent neural networks (RNNs) are capable of modeling the temporal dynamics of complex sequential information. However, the structures of existing RNN neurons mainly focus on controlling the contributions of current and historical information but do not explore the different importance levels of different elements in an input vector of a time slot. We propose adding a simple yet effective Element-wise-Attention Gate (EleAttG) to an RNN block (e.g., all RNN neurons in a network layer) that empowers the RNN neurons to have the attentiveness capability. For an RNN block, an EleAttG is added to adaptively modulate the input by assigning different levels of importance, i.e., attention, to each element/dimension of the input. We refer to an RNN block equipped with an EleAttG as an EleAtt-RNN block. Specically, the modulation of the input is content adaptive and is performed at ne granularity, being element-wise rather than input-wise. The proposed EleAttG, as an additional fundamental unit, is general and can be applied to any RNN structures, e.g., standard RNN, Long Short-Term Memory (LSTM), or Gated Recurrent Unit (GRU). We demonstrate the effectiveness of the proposed EleAtt-RNN by applying it to the action recognition tasks on both 3D human skeleton data and RGB videos. Experiments show that adding attentiveness through EleAttGs to RNN blocks signicantly boosts the power of RNNs.

Recently, many graph matching methods that incorporate pairwise constraints and that can be formulated as a quadratic assignment problem (QAP) have been proposed. Although these methods demonstrate promising results for the graph matching problem, they have high complexity in space or time. In this paper, we introduce an adaptively transforming graph matching (ATGM) method from the perspective of functional representation. More precisely, under a transformation formulation, we aim to match two graphs by minimizing the discrepancy between the original graph and the transformed graph. With a linear representation map of the transformation, the pairwise edge attributes of graphs are explicitly represented by unary node attributes, which enables us to reduce the space and time complexity significantly. Due to an efficient Frank-Wolfe method-based optimization strategy, we can handle graphs with hundreds and thousands of nodes within an acceptable amount of time. Meanwhile, because transformation map can preserve graph structures, a domain adaptation-based strategy is proposed to remove the outliers. The experimental results demonstrate that our proposed method outperforms the state-of-the-art graph matching algorithms.

Existing semantic segmentation methods mostly rely on per-pixel supervision, unable to capture structural regularity present in natural images. Instead of learning to enforce semantic labels on individual pixels, we propose to enforce affinity field patterns in individual pixel neighbourhoods, i.e., the semantic label patterns of whether neighbour pixels are in the same segment should match between the prediction and the ground-truth. The affinity fields thus characterize the intrinsic geometric relationships inside a given scene, such as ``motorcycles have round wheels''. We further develop a novel method for learning the optimal neighbourhood size for each semantic category, with an adversarial loss that optimizes over worst-case scenarios. Unlike the popular Conditional Random Field approaches, our adaptive affinity field method has no extra parameters during inference, and is also less sensitive to input appearance changes. Extensive evaluations on Cityscapes, PASCAL VOC 2012 and GTA5 datasets demonstrate AAF provides an effective, efficient, and robust solution for semantic segmentation.

Current state-of-the-art approaches for spatio-temporal action localization rely on detections at the frame level and model temporal context with 3D ConvNets. Here, we go one step further and model spatio-temporal relations to capture the interactions between human actors, relevant objects and scene elements essential to differentiate similar human actions. Our approach is weakly supervised and mines the relevant elements automatically with an actor-centric relational network (ACRN). ACRN computes and accumulates pair-wise relation information from actor and global scene features, and generates relation features for action classification. It is implemented as neural networks and can be trained jointly with an existing action detection system. We show that ACRN outperforms alternative approaches which capture relation information, and that the proposed framework improves upon the state-of-the-art performance on JHMDB and AVA. A visualization of the learned relation features confirms that our approach is able to attend to the relevant relations for each action.

In this paper we present ActiveStereoNet, the first deep learning solution for active stereo systems. Due to the lack of ground truth, our method is fully self-supervised, yet it produces precise depth with a subpixel precision of 1/30th of a pixel; it does not suffer from the common over-smoothing issues of previous approaches; it preserves the edges; and it explicitly handles occlusions. We introduce a novel reconstruction loss that is more robust to noise and texture-less patches, and is invariant to illumination changes. The proposed loss is optimized using a window-based cost aggregation with an adaptive support weight scheme. This cost aggregation is edge-preserving and smooths the loss function, which is key to allow the network to reach compelling results. Finally we show how the task of predicting invalid regions, such as occlusions, can be trained end-to-end without ground-truth. This component is crucial to reduce blur and particularly improves predictions along depth discontinuities. Extensive quantitatively and qualitatively evaluations on real and synthetic data demonstrate state of the art results in many challenging scenes.

State-of-the-art temporal action detectors inefficiently search the entire video for specific actions. Despite the encouraging progress these methods achieve, it is crucial to design automated approaches that only explore parts of the video which are the most relevant to the actions being searched for. To address this need, we propose the new problem of action spotting in video, which we define as finding a specific action in a video while observing a small portion of that video. Inspired by the observation that humans are extremely efficient and accurate in spotting and finding action instances in video, we propose Action Search, a novel Recurrent Neural Network approach that mimics the way humans spot actions. Moreover, to address the absence of data recording the behavior of human annotators, we put forward the Human Searches dataset, which compiles the search sequences employed by human annotators spotting actions in the AVA and THUMOS14 datasets. We consider temporal action localization as an application of the action spotting problem. Experiments on the THUMOS14 dataset reveal that our model is not only able to explore the video efficiently (observing on average 17.3% of the video) but it also accurately finds human activities with 30.8% mAP.