ECCV 2018 Papers

Critical to the registration of point clouds is the establishment of a set of accurate correspondences between points in 3D space. The correspondence problem is generally addressed by the design of discriminative 3D local descriptors on the one hand, and the development of robust matching strategies on the other hand. In this work, we first propose a multi-view local descriptor, which is learned from the images of multiple views, for the description of 3D keypoints. Then, we develop a robust matching approach, aiming at rejecting outlier matches based on the efficient inference via belief propagation on the defined graphical model. We have demonstrated the boost of our approaches to registration on the public scanning and multi-view stereo datasets. The superior performance has been verified by the intensive comparisons against a variety of descriptors and matching methods.

This paper studies the problem of blind face restoration from an unconstrained blurry, noisy, low-resolution, or compressed image (i.e., degraded observation). For better recovery of fine facial details, we modify the problem setting by taking both the degraded observation and a high-quality guided image of the same identity as input to our guided face restoration network (GFRNet). However, the degraded observation and guided image generally are different in pose, illumination and expression, thereby making plain CNNs (e.g., U-Net) fail to recover fine and identity-aware facial details. To tackle this issue, our GFRNet model includes both a warping subnetwork (WarpNet) and a reconstruction subnetwork (RecNet). The WarpNet is introduced to predict flow field for warping the guided image to correct pose and expression (i.e., warped guidance), while the RecNet takes the degraded observation and warped guidance as input to produce the restoration result. Due to that the ground-truth flow field is unavailable, landmark loss together with total variation regularization are incorporated to guide the learning of WarpNet. Furthermore, to make the model applicable to blind restoration, our GFRNet is trained on the synthetic data with versatile settings on blur kernel, noise level, downsampling scale factor, and JPEG quality factor. Experiments show that our GFRNet not only performs favorably against the state-of-the-art image and face restoration methods, but also generates visually photo-realistic results on real degraded facial images.

Attention mechanisms in biological perception are thought to select subsets of perceptual information for more sophisticated processing which would be prohibitive to perform on all sensory inputs. In computer vision, however, there has been relatively little exploration of hard attention, where some information is selectively ignored, in spite of the success of soft attention, where information is re-weighted and aggregated, but never filtered out. Here, we introduce a new approach for hard attention and find it achieves very competitive performance on a recently-released visual question answering dataset, equalling and in some cases surpassing similar soft attention architectures while entirely ignoring some features. Even though the hard attention mechanism is not differentiable, we found that the feature magnitudes correlate with semantic relevance, and provided a useful signal for our mechanism's attentional selection criterion. Because hard attention selects important features of the input information, it can also be more efficient than analogous soft attention mechanisms. This is especially important for recent approaches that use non-local pairwise operations, whereby computational and memory costs are quadratic in the size of the set of features.

Outfits in online fashion data are composed of items of many different types (e.g. top, bottom, shoes) that share some stylistic relationship with one another. A representation for building outfits requires a method that can learn both notions of similarity (for example, when two tops are interchangeable) and compatibility (items of possibly different type that can go together in an outfit). This paper presents an approach to learning an image embedding that respects item type, and jointly learns notions of item similarity and compatibility in an end-to-end model. To evaluate the learned representation, we crawled 68,306 outfits created by users on the Polyvore website. Our approach obtains 3-5% improvement over the state-of-the-art on outfit compatibility prediction and fill-in-the-blank tasks using our dataset, as well as an established smaller dataset, while supporting a variety of useful queries.

It is expensive to label images with 3D structure or precise camera pose. Yet, this is precisely the kind of annotation required to train single-view 3D reconstruction models. In contrast, unlabeled images or images with just category labels are easy to acquire, but few current models can use this weak supervision. We present a unified framework that can combine both types of supervision: a small amount of camera pose annotations are used to enforce pose-invariance and view-point consistency, and unlabeled images combined with an adversarial loss are used to enforce the realism of rendered, generated models. We use this unified framework to measure the impact of each form of supervision in three paradigms: semi-supervised, multi-task, and transfer learning. We show that with a combination of these ideas, we can train single-view reconstruction models that improve up to 7 points in performance (AP) when using only 1% pose annotated training data.

The problem of single-view 3D shape completion or reconstruction is challenging, because among the many possible shapes that explain an observation, most are implausible and do not correspond to natural objects. Recent research in the field has tackled this problem by exploiting the expressiveness of deep convolutional networks. In fact, there is another level of ambiguity that is often overlooked: there are usually multiple plausible shapes that fit the 2D image equally well; i.e., the ground truth shape is non-deterministic given the input. Existing fully supervised approaches fail to address this issue, and often produce blurry mean shapes with smooth surfaces but no fine details. In this paper, we propose ShapeHD, pushing the limit of single-view shape completion and reconstruction by integrating deep generative models with adversarially learned shape priors. The learned priors serve as a regularizer, penalizing the model only if its output is unrealistic, not if it deviates from the ground truth. Our design thus overcomes both types of ambiguities aforementioned. Experiments demonstrate that ShapeHD outperforms state-of-the-arts by a large margin on both shape completion and shape reconstruction.

We address the problem of 3D rotation equivariance in convolutional neural networks. 3D rotations have been a challenging nuisance in 3D classification tasks requiring higher capacity and extended data augmentation in order to tackle it. We model 3D data with multi-valued spherical functions and we propose a novel spherical convolutional network that implements exact convolutions on the sphere by realizing them in the spherical harmonic domain. Resulting filters have local symmetry and are localized by enforcing smooth spectra. We apply a novel pooling on the spectral domain and our operations are independent of the underlying spherical resolution throughout the network. We show that networks with much lower capacity and without requiring data augmentation can exhibit performance comparable to the state of the art in standard retrieval and classification benchmarks.

Estimation of 3D motion in a dynamic scene from a temporal pair of images is a core task in many scene understanding problems. In real world applications, a dynamic scene is commonly captured by a moving camera (i.e., panning, tilting or hand-held), increasing the task complexity because the scene is observed from different view points. The main challenge is the disambiguation of the camera motion from scene motion, which becomes more difficult as the amount of rigidity observed decreases, even with successful estimation of 2D image correspondences. Compared to other state-of-the-art 3D scene flow estimation methods, in this paper we propose to emph{learn} the rigidity of a scene in a supervised manner from a large collection of dynamic scene data, and directly infer a rigidity mask from two sequential images with depths. With the learned network, we show how we can effectively estimate camera motion and projected scene flow using computed 2D optical flow and the inferred rigidity mask. For training and testing the rigidity network, we also provide a new semi-synthetic dynamic scene dataset (synthetic foreground objects with a real background) and an evaluation split that accounts for the percentage of observed non-rigid pixels. Through our evaluation, we show the proposed framework outperforms current state-of-the-art scene flow estimation methods in challenging dynamic scenes.

While most steps in the modern object detection methods are learnable, the region feature extraction step remains largely hand-crafted, featured by RoI pooling methods. This work proposes a general viewpoint that unifies existing region feature extraction methods and a novel method that is end-to-end learnable. The proposed method removes most heuristic choices and outperforms its RoI pooling counterparts. It moves further towards emph{fully learnable object detection}.

We present a novel semantic 3D reconstruction framework which embeds variational regularization into a neural network. Our network performs a fixed number of unrolled multi-scale optimization iterations with shared interaction weights. In contrast to existing variational methods for semantic 3D reconstruction, our model is end-to-end trainable and captures more complex dependencies between the semantic labels and the 3D geometry. Compared to previous learning-based approaches to 3D reconstruction, we integrate powerful long-range dependencies using variational coarse-to-fine optimization. As a result, our network architecture requires only a moderate number of parameters while keeping a high level of expressiveness which enables learning from very little data. Experiments on real and synthetic datasets demonstrate that our network achieves higher accuracy compared to a purely variational approach while at the same time requiring two orders of magnitude less iterations to converge. Moreover, our approach handles ten times more semantic class labels using the same computational resources.

Monocular depth estimation aims at estimating a pixelwise depth map for a single image, which has wide applications in scene understanding and autonomous driving. Existing supervised and unsupervised methods face great challenges. Supervised methods require large amounts of depth measurement data, which are generally difficult to obtain, while unsupervised methods are usually limited in estimation accuracy. Synthetic data generated by graphics engines provide a possible solution for collecting large amounts of depth data. However, the large domain gaps between synthetic and realistic data make directly training with them challenging. In this paper, we propose to use the stereo matching network as a proxy to learn depth from synthetic data and use predicted stereo disparity maps for supervising the monocular depth estimation network. Cross-domain synthetic data could be fully utilized in this novel framework. Different strategies are proposed to ensure learned depth perception capability well transferred across different domains. Our extensive experiments show state-of-the-art results of monocular depth estimation on KITTI dataset.

This paper addresses the task of detecting and recognizing human-object interactions (HOI) in images and videos. We introduce the Graph Parsing Neural Network (GPNN), a framework that incorporates structural knowledge while being differentiable end-to-end. For a given scene, GPNN infers a parse graph that includes i) the HOI graph structure represented by an adjacency matrix, and ii) the node labels. Within a message passing inference framework, GPNN iteratively computes the adjacency matrices and node labels. We extensively evaluate our model on three HOI detection benchmarks on images and videos: HICO-DET, V-COCO, and CAD-120 datasets. Our approach significantly outperforms state-of-art methods, verifying that GPNN is scalable to large datasets and applies to spatial-temporal settings.

Though Faster R-CNN based two-stage detectors have witnessed significant boost in pedestrian detection accuracy, it is still slow for practical applications. One solution is to simplify this working flow as a single-stage detector. However, current single-stage detectors (e.g. SSD) have not presented competitive accuracy on common pedestrian detection benchmarks. This paper is towards a successful pedestrian detector enjoying the speed of SSD while maintaining the accuracy of Faster R-CNN. Specifically, a structurally simple but effective module called emph{Asymptotic Localization Fitting} (ALF) is proposed, which stacks a series of predictors to directly evolve the default anchor boxes of SSD step by step into improving detection results. As a result, during training the latter predictors enjoy more and better-quality positive samples, meanwhile harder negatives could be mined with increasing IoU thresholds. On top of this, an efficient single-stage pedestrian detection architecture (denoted as ALFNet) is designed, achieving state-of-the-art performance on CityPersons and Caltech, two of the largest pedestrian detection benchmarks, and hence resulting in an attractive pedestrian detector in both accuracy and speed. Code is available at href{https://github.com/VideoObjectSearch/ALFNet}{https://github.com/VideoObjectSearch/ALFNet}.

Template-matching methods for visual tracking have gained popularity recently due to their comparable performance and fast speed. However, they lack effective ways to adapt to changes in the target object’s appearance, making their tracking accuracy still far from state-of-the-art. In this paper, we propose a dynamic memory network to adapt the template to the target’s appearance variations during tracking. An LSTM is used as a memory controller, where the input is the search feature map and the outputs are the control signals for the reading and writing process of the memory block. As the location of the target is at first unknown in the search feature map, an attention mechanism is applied to concentrate the LSTM input on the potential target. To prevent aggressive model adaptivity, we apply gated residual template learning to control the amount of retrieved memory that is used to combine with the initial template. Unlike tracking-by-detection methods where the object’s information is maintained by the weight parameters of neural networks, which requires expensive online fine-tuning to be adaptable, our tracker runs completely feed-forward and adapts to the target’s appearance changes by updating the external memory. Moreover, unlike other tracking methods where the model capacity is fixed after offline training – the capacity of our tracker can be easily enlarged as the memory require- ments of a task increase, which is favorable for memorizing long-term ob- ject information. Extensive experiments on OTB and VOT demonstrates that our tracker MemTrack performs favorably against state-of-the-art tracking methods while retaining real-time speed of 50 fps.

Adversarial perturbations are noise-like patterns that can subtly change the data, while failing an otherwise accurate classifier. In this paper, we propose to use such perturbations for improving the robustness of video representations. To this end, given a well-trained deep model for per-frame video recognition, we first generate adversarial noise adapted to this model. Using the original data features from the full video sequence and their perturbed counterparts, as two separate bags, we develop a binary classification problem that learns a set of discriminative hyperplanes -- as a subspace -- that will separate the two bags from each other. This subspace is then used as a descriptor for the video, dubbed discriminative subspace pooling. As the perturbed features belong to data classes that are likely to be confused with the original features, the discriminative subspace will characterize parts of the feature space that are more representative of the original data, and thus may provide robust video representations. To learn such descriptors, we formulate a subspace learning objective on the Stiefel manifold and resort to Riemannian optimization methods for solving it efficiently. We provide experiments on several video datasets and demonstrate state-of-the-art results.

Knowledge Transfer (KT) techniques tackle the problem of transferring the knowledge from a large and complex neural network into a smaller and faster one. However, existing KT methods are tailored towards classification tasks and they cannot be used efficiently for other representation learning tasks. In this paper we propose a novel probabilistic knowledge transfer method that works by matching the probability distribution of the data in the feature space instead of their actual representation. Apart from outperforming existing KT techniques, the proposed method allows for overcoming several of their limitations providing new insight into KT as well as novel KT applications, ranging from KT from handcrafted feature extractors to cross-modal KT from the textual modality into the representation extracted from the visual modality of the data.

Existing deblurring methods mainly focus on developing effective image priors and assume that blurred images contain insignificant amounts of noise. However, state-of-the-art deblurring methods do not perform well on real-world images degraded with significant noise or outliers. To address these issues, we show that it is critical to learn data fitting terms beyond the commonly used L1 or L2 norm. We propose a simple and effective discriminative framework to learn data terms that can adaptively handle blurred images in the presence of severe noise and outliers. Instead of learning the distribution of the data fitting errors, we directly learn the associated shrinkage function for the data term using a cascaded architecture, which is more flexible and efficient. Our analysis shows that the shrinkage functions learned at the intermediate stages can effectively suppress noise and preserve image structures. Extensive experimental results show that the proposed algorithm performs favorably against state-of-the-art methods.

Convolutional neural networks (CNNs) have dramatically advanced the state-of-art in a number of domains. However, most models are both computation and memory intensive, which arouse the interest of network compression. While existing compression methods achieve good performance, they suffer from three limitations: 1) the inevitable retraining with enormous labeled data; 2) the massive GPU hours for retraining; 3) the training tricks for model compression. Especially the requirement of retraining on original datasets makes it difficult to apply in many real-world scenarios, where training data is not publicly available. In this paper, we reveal that re-normalization is the practical and effective way to alleviate the above limitations. Through quantization or pruning, most methods may compress a large number of parameters but ignore the core role in performance degradation, which is the Gaussian conjugate prior induced by batch normalization. By employing the re-estimated statistics in batch normalization, we significantly improve the accuracy of compressed CNNs. Extensive experiments on ImageNet show it outperforms baselines by a large margin and is comparable to label-based methods. Besides, the fine-tuning process takes less than 5 minutes on CPU, using 1000 unlabeled images.

Zero-shot learning (ZSL) aims to recognize objects of novel classes without any training samples of specific classes, which is achieved by exploiting the semantic information and auxiliary datasets. Recently most ZSL approaches focus on learning visual-semantic embeddings to transfer knowledge from the auxiliary datasets to the novel classes. However, few works study whether the semantic information is discriminative or not for the recognition task. To tackle such problem, we propose a coupled dictionary learning approach to align the visual-semantic structures using the class prototypes, where the discriminative information lying in the visual space is utilized to improve the less discriminative semantic space. Then, zero-shot recognition can be performed in different spaces by the simple nearest neighbor approach using the learned class prototypes. Extensive experiments on four benchmark datasets show the effectiveness of the proposed approach.