ECCV 2018 论文列表

Designing discriminative and invariant features is the key to visual recognition. Recently, the bilinear pooled feature matrix of Convolutional Neural Network (CNN) has shown to achieve state-of-the-art performance on a range of fine-grained visual recognition tasks. The bilinear feature matrix collects second-order statistics and is closely related to the covariance matrix descriptor. However, the bilinear feature could suffer from the visual burstiness phenomenon similar to other visual representations such as VLAD and Fisher Vector. The reason is that the bilinear feature matrix is sensitive to the magnitudes and correlations of local CNN feature elements which can be measured by its singular values. On the other hand, the singular vectors are more invariant and reasonable to be adopted as the feature representation. Motivated by this point, we advocate an alternative pooling method which transforms the CNN feature matrix to an orthonormal matrix consists of its principal singular vectors. Geometrically, such orthonormal matrix lies on the Grassmann manifold, a Riemannian manifold whose points represent subspaces of the Euclidean space. Similarity measurement of images reduces to comparing the principal angles between these ``homogeneous" subspaces and thus is independent of the magnitudes and correlations of local CNN activations. In particular, we demonstrate that the projection distance on the Grassmann manifold deduces a bilinear feature mapping without explicitly computing the bilinear feature matrix, which enables a very compact feature and classifier representation. Experimental results show that our method achieves an excellent balance of model complexity and accuracy on a variety of fine-grained image classification datasets.

We propose a novel scene graph generation model called Graph R-CNN, that is both effective and efficient at detecting objects and their relations in images. Our model contains a Relation Proposal Network (RePN) that efficiently deals with the quadratic number of potential relations between objects in an image. We also propose an attentional Graph Convolutional Network (aGCN) that effectively captures contextual information between objects and relations. Finally, we introduce a novel scene graph evaluation metric that is more holistic and realistic than existing metrics. We report state-of-the-art performance on scene graph generation as evaluated using both existing and our proposed metrics.

We propose a technique that tackles action detection in multimodal videos under a realistic and challenging condition in which only limited training data and partially observed modalities are available. Common methods in transfer learning do not take advantage of the extra modalities potentially available in the source domain. On the other hand, previous work on multimodal learning only focuses on a single domain or task and does not handle the modality discrepancy between training and testing. In this work, we propose a method termed graph distillation that incorporates rich privileged information from a large-scale multimodal dataset in the source domain, and improves the learning in the target domain where training data and modalities are scarce. We evaluate our approach on action classification and detection tasks in multimodal videos, and show that our model outperforms the state-of-the-art by a large margin on the NTU RGB+D and PKU-MMD benchmarks. The code is released at http://alan.vision/eccv18_graph/.

Unsupervised domain adaptation has caught appealing attentions as it facilitates the unlabeled target learning by borrowing existing well-established source domain knowledge. Recent practice on domain adaptation manages to extract effective features by incorporating the pseudo labels for the target domain to better solve cross-domain distribution divergences. However, existing approaches separate target label optimization and domain-invariant feature learning as different steps. To address that issue, we develop a novel Graph Adaptive Knowledge Transfer (GAKT) model to jointly optimize target labels and domain-free features in a unified framework. Specifically, semi-supervised knowledge adaptation and label propagation on target data are coupled to benefit each other, and hence the marginal and conditional disparities across different domains will be better alleviated. Experimental evaluation on two cross-domain visual datasets demonstrates the effectiveness of our designed approach on facilitating the unlabeled target task learning, compared to the state-of-the-art domain adaptation approaches.

In this paper, we propose a graininess-aware deep feature learning method for pedestrian detection. Unlike most existing pedestrian detection methods which only consider low resolution feature maps, we incorporate fine-grained information into convolutional features to make them more discriminative for human body parts. Specifically, we propose a pedestrian attention mechanism which efficiently identifies pedestrian regions. Our method encodes fine-grained attention masks into convolutional feature maps, which significantly suppresses background interference and highlights pedestrians. Hence, our graininess-aware features become more focused on pedestrians, in particular those of small size and with occlusion. We further introduce a zoom-in-zoom-out module, which enhances the features by incorporating local details and context information. We integrate these two modules into a deep neural network, forming an end-to-end trainable pedestrian detector. Comprehensive experimental results on four challenging pedestrian benchmarks demonstrate the effectiveness of the proposed approach.

This paper tackles a problem in line-assisted VO/VSLAM: accurately solving the least squares pose optimization with unreliable 3D line input. The solution we present is good line cutting, which extracts the most-informative sub-segment from each 3D line for use within the pose optimization formulation. By studying the impact of line cutting towards the information gain of pose estimation in line-based least squares problem, we demonstrate the applicability of improving pose estimation accuracy with good line cutting. To that end, we describe an efficient algorithm that approximately approaches the joint optimization problem of good line cutting. The proposed algorithm is integrated into a state-of-the-art line-assisted VSLAM system. When evaluated in two target scenarios of line-assisted VO/VSLAM, low-texture and motion blur, the accuracy of pose tracking is improved, while the robustness is preserved.

Despite significant progress in a variety of vision-and-language problems, developing a method capable of asking intelligent, goal-oriented questions about images is proven to be an inscrutable challenge. Towards this end, we propose a Deep Reinforcement Learning framework based on three new intermediate rewards, namely goal-achieved, progressive and informativeness that encourage the generation of succinct questions, which in turn uncover valuable information towards the overall goal. By directly optimizing for questions that work quickly towards fulfilling the overall goal, we avoid the tendency of existing methods to generate long series of inane queries that add little value. We evaluate our model on the GuessWhat?! dataset and show that the resulting questions can help a standard `Guesser' identify a specific object in an image at a much higher success rate.

Lane detection is playing an indispensable role in advanced driver assistance systems. The existing approaches for lane detection can be categorized as lane area segmentation and lane boundary detection. Most of these methods abandon a great quantity of complementary information, such as geometric priors, when exploiting the lane area and the lane boundaries alternatively. In this paper, we establish a multiple-task learning framework to segment lane areas and detect lane boundaries simultaneously. The main contributions of the proposed frame- work are highlighted in two facets: (1) We put forward a multiple-task learning framework with mutually interlinked sub-structures between lane segmentation and lane boundary detection to improve overall performance. (2) A novel loss function is proposed with two geometric constraints considered, as assumed that the lane boundary is predicted as the outer contour of the lane area while the lane area is predicted as the area integration result within the lane boundary lines. With an end-to-end training process, these improvements extremely enhance the robustness and accuracy of our approach on several metrics. The proposed framework is evaluated on KITTI dataset, CULane dataset and RVD dataset. Compared with the state of the arts, our approach achieves the best performance on the metrics and a more robust detection in varied traffic scenes.

While the successful estimation of a photo's geolocation enables a number of interesting applications, it is also a very challenging task. Due to the complexity of the problem, most existing approaches are restricted to specific areas, imagery, or worldwide landmarks. Only a few proposals predict GPS coordinates without any limitations. In this paper, we introduce several deep learning methods, which pursue the latter approach and treat geolocalization as a classification problem where the earth is subdivided into geographical cells. We propose to exploit hierarchical knowledge of multiple partitionings and additionally extract and take the photo's scene content into account, i.e., indoor, natural, or urban setting etc. As a result, contextual information at different spatial resolutions as well as more specific features for various environmental settings are incorporated in the learning process of the convolutional neural network. Experimental results on two benchmarks demonstrate the effectiveness of our approach outperforming the state of the art while using a significant lower number of training images and without relying on retrieval methods that require an appropriate reference dataset.

Learned local descriptors based on Convolutional Neural Networks (CNNs) have achieved significant improvements on patch-based benchmarks, whereas not having demonstrated strong generalization ability on recent benchmarks of image-based 3D reconstruction. In this paper, we mitigate this limitation by proposing a novel local descriptor learning approach that integrates geometry constraints from multi-view reconstructions, which benefit the learning process in data generation, data sampling and loss computation. We refer to the proposed descriptor as GeoDesc, and demonstrate its superior performance on various large-scale benchmarks, and in particular show its great success on challenging reconstruction cases. Moreover, we provide guidelines towards practical integration of learned descriptors in Structure-from-Motion (SfM) pipelines, showing the good trade-off that GeoDesc delivers to 3D reconstruction tasks between accuracy and efficiency.

Despite the promising results on paired/unpaired image-to-image translation achieved by Generative Adversarial Networks (GANs), prior works often only transfer the low-level information (e.g. color or texture changes), but fail to manipulate high-level semantic meanings (e.g., geometric structure or content) of different object regions. On the other hand, while some researches can synthesize compelling real-world images given a class label or caption, they cannot condition on arbitrary shapes or structures, which largely limits their application scenarios and interpretive capability of model results. In this work, we focus on a more challenging semantic manipulation task, aiming at modifying the semantic meaning of an object while preserving its own characteristics (e.g. viewpoints and shapes), such as cow$ ightarrow$sheep, motor$ ightarrow$ bicycle, cat$ ightarrow$dog. To tackle such large semantic changes, we introduce a contrasting GAN (contrast-GAN) with a novel adversarial contrasting objective which is able to perform all types of semantic translations with one category-conditional generator. Instead of directly making the synthesized samples close to target data as previous GANs did, our adversarial contrasting objective optimizes over the distance comparisons between samples, that is, enforcing the manipulated data be semantically closer to the real data with target category than the input data. Equipped with the new contrasting objective, a novel mask-conditional contrast-GAN architecture is proposed to enable disentangle image background with object semantic changes. Extensive qualitative and quantitative experiments on several semantic manipulation tasks on ImageNet and MSCOCO dataset show considerable performance gain by our contrast-GAN over other conditional GANs.

Due to the succinct nature of free-hand sketch drawings, sketch-based image retrieval (SBIR) has abundant practical use cases in consumer electronics. However, SBIR remains a long-standing unsolved problem mainly due to the significant discrepancy between the sketch domain and the image domain. In this work, we propose a Generative Domain-migration Hashing (GDH) approach, which for the first time generates hashing codes from synthetic natural images that are migrated from sketches. The generative model learns a mapping that the distributions of sketches can be indistinguishable from the distribution of natural images using an adversarial loss, and simultaneously learns an inverse mapping based on the cycle consistency loss in order to enhance the indistinguishability. With the robust mapping learned from the generative model, GDH can migrate sketches to their indistinguishable image counterparts while preserving the domain-invariant information of sketches. With an end-to-end multi-task learning framework, the generative model and binarized hashing codes can be jointly optimized. Comprehensive experiments of both category-level and fine-grained SBIR on multiple large-scale datasets demonstrate the consistently balanced superiority of GDH in terms of efficiency, memory costs and effectiveness.

Face attribute editing aims at editing the face image with the given attribute. Most existing works employ Generative Adversarial Network (GAN) to operate face attribute editing. However, these methods inevitably change the attribute-irrelevant regions, as shown in Fig.~ ef{fig1}. Therefore, we introduce the spatial attention mechanism into GAN framework (referred to as SaGAN), to only alter the attribute-specific region and keep the rest unchanged. Our approach SaGAN consists of a generator and a discriminator. The generator contains an attribute manipulation network (AMN) to edit the face image, and a spatial attention network (SAN) to localize the attribute-specific region which restricts the alternation of AMN within this region. The discriminator endeavors to distinguish the generated images from the real ones, and classify the face attribute. Experiments demonstrate that our approach can achieve promising visual results, and keep those attribute-irrelevant regions unchanged. Besides, our approach can benefit the face recognition by data augmentation.

Learned 3D representations of human faces are useful for computer vision problems such as 3D face tracking and reconstruction from images, as well as graphics applications such as character generation and animation. Traditional models learn a latent representation of a face using linear subspaces or higher-order tensor generalizations. Due to this linearity, they can not capture extreme deformations and non-linear expressions. To address this, we introduce a versatile model that learns a non-linear representation of a face using spectral convolutions on a mesh surface. We introduce mesh sampling operations that enable a hierarchical mesh representation that captures non-linear variations in shape and expression at multiple scales within the model. In a variational setting, our model samples diverse realistic 3D faces from a multivariate Gaussian distribution. Our training data consists of 20,466 meshes of extreme expressions captured over 12 different subjects. Despite limited training data, our trained model outperforms state-of-the-art face models with 50% lower reconstruction error, while using 75% fewer parameters. We also show that, replacing the expression space of an existing state-of-the-art face model with our autoencoder, achieves a lower reconstruction error. Our data, model and code are available at http://coma.is.tue.mpg.de/.

Person re-identification (re-ID) poses unique challenges for unsupervised domain adaptation (UDA) in that classes in the source and target sets (domains) are entirely different and that image variations are largely caused by cameras. Given a labeled source training set and an unlabeled target training set, we aim to improve the generalization ability of re-ID models on the target testing set. To this end, we introduce a Hetero-Homogeneous Learning (HHL) method. Our method enforces two properties simultaneously: 1) camera invariance, learned via positive pairs formed by unlabeled target images and their camera style transferred counterparts; 2) domain connectedness, by regarding source / target images as negative matching pairs to the target / source images. The first property is implemented by homogeneous learning because training pairs are collected from the same domain. The second property is achieved by heterogeneous learning because we sample training pairs from both the source and target domains. On Market-1501, DukeMTMC-reID and CUHK03, we show that the two properties contribute indispensably and that very competitive re-ID UDA accuracy is achieved. Code is available at: https://github.com/zhunzhong07/HHL

The classic Generative Adversarial Net and its variants can be roughly categorized into two large families: the unregularized ver- sus regularized GANs. By relaxing the non-parametric assumption on the discriminator in the classic GAN, the regularized GANs have better generalization ability to produce new samples drawn from the real dis- tribution. It is well known that the real data like natural images are not uniformly distributed over the whole data space. Instead, they are often restricted to a low-dimensional manifold of the ambient space. Such a manifold assumption suggests the distance over the manifold should be a better measure to characterize the distinct between real and fake sam- ples. Thus, we define a pullback operator to map samples back to their data manifold, and a manifold margin is defined as the distance between the pullback representations to distinguish between real and fake sam- ples and learn the optimal generators. We justify the effectiveness of the proposed model both theoretically and empirically.

Recent advances in Generative Adversarial Networks (GANs) have shown impressive results for task of facial expression synthesis. The most successful architecture is StarGAN, that conditions GANs' generation process with images of a specific domain, namely a set of images of persons sharing the same expression. While effective, this approach can only generate a discrete number of expressions, determined by the content of the dataset. To address this limitation, in this paper, we introduce a novel GAN conditioning scheme based on Action Units (AU) annotations, which describes in a continuous manifold the anatomical facial movements defining a human expression. Our approach allows controlling the magnitude of activation of each AU and combine several of them. Additionally, we propose a fully unsupervised strategy to train the model, that only requires images annotated with their activated AUs, and exploit attention mechanisms that make our network robust to changing backgrounds and lighting conditions. Extensive evaluation show that our approach goes beyond competing conditional generators both in the capability to synthesize a much wider range of expressions ruled by anatomically feasible muscle movements, as in the capacity of dealing with images in the wild.

In this paper, we present a framework for reconstructing a point-based 3D model of an object from a single view image. Distance metrics, like Chamfer distance, were used in previous work to measure the difference of two point sets and serve as the loss function in point-based reconstruction. However, such point-point loss does not constrain the 3D model from a global perspective. We propose to add geometric adversarial loss (GAL). It is composed of two terms where the geometric loss ensures consistent shape of reconstructed 3D models close to ground-truth from different viewpoints, and the conditional adversarial loss generates a semantically-meaningful point cloud. GAL benefits predicting the obscured part of objects and maintaining geometric structure of the predicted 3D model. Both the qualitative results and quantitative analysis manifest the generality and suitability of our method.

This work proposes a general-purpose, fully-convolutional network architecture for efficiently processing large-scale 3D data. One striking characteristic of our approach is its ability to process unorganized 3D representations such as point clouds as input, then transforming them internally to ordered structures to be processed via 3D convolutions. In contrast to conventional approaches that maintain either unorganized or organized representations, from input to output, our approach has the advantage of operating on memory efficient input data representations while at the same time exploiting the natural structure of convolutional operations to avoid the redundant computing and storing of spatial information in the network. The network eliminates the need to pre- or post process the raw sensor data. This, together with the fully-convolutional nature of the network, makes it an end-to-end method able to process point clouds of huge spaces or even entire rooms with up to 200k points at once. Another advantage is that our network can produce either an ordered output or map predictions directly onto the input cloud, thus making it suitable as a general-purpose point cloud descriptor applicable to many 3D tasks. We demonstrate our network’s ability to effectively learn both low-level features as well as complex compositional relationships by evaluating it on benchmark datasets for semantic voxel segmentation, semantic part segmentation and 3D scene captioning.