ICCV 2019 论文列表

We present Extreme View Synthesis, a solution for novel view extrapolation that works even when the number of input images is small---as few as two. In this context, occlusions and depth uncertainty are two of the most pressing issues, and worsen as the degree of extrapolation increases. We follow the traditional paradigm of performing depth-based warping and refinement, with a few key improvements. First, we estimate a depth probability volume, rather than just a single depth value for each pixel of the novel view. This allows us to leverage depth uncertainty in challenging regions, such as depth discontinuities. After using it to get an initial estimate of the novel view, we explicitly combine learned image priors and the depth uncertainty to synthesize a refined image with less artifacts. Our method is the first to show visually pleasing results for baseline magnifications of up to 30x.

Driving requires reacting to a wide variety of complex environment conditions and agent behaviors. Explicitly modeling each possible scenario is unrealistic. In contrast, imitation learning can, in theory, leverage data from large fleets of human-driven cars. Behavior cloning in particular has been successfully used to learn simple visuomotor policies end-to-end, but scaling to the full spectrum of driving behaviors remains an unsolved problem. In this paper, we propose a new benchmark to experimentally investigate the scalability and limitations of behavior cloning. We show that behavior cloning leads to state-of-the-art results, executing complex lateral and longitudinal maneuvers, even in unseen environments, without being explicitly programmed to do so. However, we confirm some limitations of the behavior cloning approach: some well-known limitations (e.g., dataset bias and overfitting), new generalization issues (e.g., dynamic objects and the lack of a causal modeling), and training instabilities, all requiring further research before behavior cloning can graduate to real-world driving. The code, dataset, benchmark, and agent studied in this paper can be found at github.com/felipecode/coiltraine/blob/master/docs/exploring_limitations.md

Neural networks for image recognition have evolved through extensive manual design from simple chain-like models to structures with multiple wiring paths. The success of ResNets and DenseNets is due in large part to their innovative wiring plans. Now, neural architecture search (NAS) studies are exploring the joint optimization of wiring and operation types, however, the space of possible wirings is constrained and still driven by manual design despite being searched. In this paper, we explore a more diverse set of connectivity patterns through the lens of randomly wired neural networks. To do this, we first define the concept of a stochastic network generator that encapsulates the entire network generation process. Encapsulation provides a unified view of NAS and randomly wired networks. Then, we use three classical random graph models to generate randomly wired graphs for networks. The results are surprising: several variants of these random generators yield network instances that have competitive accuracy on the ImageNet benchmark. These results suggest that new efforts focusing on designing better network generators may lead to new breakthroughs by exploring less constrained search spaces with more room for novel design.

This paper presents a method to pursue a semantic and quantitative explanation for the knowledge encoded in a convolutional neural network (CNN). The estimation of the specific rationale of each prediction made by the CNN presents a key issue of understanding neural networks, and it is of significant values in real applications. In this study, we propose to distill knowledge from the CNN into an explainable additive model, which explains the CNN prediction quantitatively. We discuss the problem of the biased interpretation of CNN predictions. To overcome the biased interpretation, we develop prior losses to guide the learning of the explainable additive model. Experimental results have demonstrated the effectiveness of our method.

Self-attention mechanism has been widely used for various tasks. It is designed to compute the representation of each position by a weighted sum of the features at all positions. Thus, it can capture long-range relations for computer vision tasks. However, it is computationally consuming. Since the attention maps are computed w.r.t all other positions. In this paper, we formulate the attention mechanism into an expectation-maximization manner and iteratively estimate a much more compact set of bases upon which the attention maps are computed. By a weighted summation upon these bases, the resulting representation is low-rank and deprecates noisy information from the input. The proposed Expectation-Maximization Attention (EMA) module is robust to the variance of input and is also friendly in memory and computation. Moreover, we set up the bases maintenance and normalization methods to stabilize its training procedure. We conduct extensive experiments on popular semantic segmentation benchmarks including PASCAL VOC, PASCAL Context, and COCO Stuff, on which we set new records.

Several popular approaches to 3D vision tasks process multiple views of the input independently with deep neural networks pre-trained on natural images, where view permutation invariance is achieved through a single round of pooling over all views. We argue that this operation discards important information and leads to subpar global descriptors. In this paper, we propose a group convolutional approach to multiple view aggregation where convolutions are performed over a discrete subgroup of the rotation group, enabling, thus, joint reasoning over all views in an equivariant (instead of invariant) fashion, up to the very last layer. We further develop this idea to operate on smaller discrete homogeneous spaces of the rotation group, where a polar view representation is used to maintain equivariance with only a fraction of the number of input views. We set the new state of the art in several large scale 3D shape retrieval tasks, and show additional applications to panoramic scene classification.

Domain generalization (DG) is the challenging and topical problem of learning models that generalize to novel testing domains with different statistics than a set of known training domains. The simple approach of aggregating data from all source domains and training a single deep neural network end-to-end on all the data provides a surprisingly strong baseline that surpasses many prior published methods. In this paper we build on this strong baseline by designing an episodic training procedure that trains a single deep network in a way that exposes it to the domain shift that characterises a novel domain at runtime. Specifically, we decompose a deep network into feature extractor and classifier components, and then train each component by simulating it interacting with a partner who is badly tuned for the current domain. This makes both components more robust, ultimately leading to our networks producing state-of-the-art performance on three DG benchmarks. Furthermore, we consider the pervasive workflow of using an ImageNet trained CNN as a fixed feature extractor for downstream recognition tasks. Using the Visual Decathlon benchmark, we demonstrate that our episodic-DG training improves the performance of such a general purpose feature extractor by explicitly training a feature for robustness to novel problems. This shows that DG training can benefit standard practice in computer vision.

In this paper, we are interested in self-supervised learning the motion cues in videos using dynamic motion filters for a better motion representation to finally boost human action recognition in particular. Thus far, the vision community has focused on spatio-temporal approaches using standard filters, rather we here propose dynamic filters that adaptively learn the video-specific internal motion representation by predicting the short-term future frames. We name this new motion representation, as dynamic motion representation (DMR) and is embedded inside of 3D convolutional network as a new layer, which captures the visual appearance and motion dynamics throughout entire video clip via end-to-end network learning. Simultaneously, we utilize these motion representation to enrich video classification. We have designed the frame prediction task as an auxiliary task to empower the classification problem. With these overall objectives, to this end, we introduce a novel unified spatio-temporal 3D-CNN architecture (DynamoNet) that jointly optimizes the video classification and learning motion representation by predicting future frames as a multi-task learning problem. We conduct experiments on challenging human action datasets: Kinetics 400, UCF101, HMDB51. The experiments using the proposed DynamoNet show promising results on all the datasets.

We propose a method that reconstructs dynamic positron emission tomography (PET) images from given sinograms by using non-negative matrix factorization (NMF) incorporated with a deep image prior (DIP) for appropriately constraining the spatial patterns of resultant images. The proposed method can reconstruct dynamic PET images with higher signal-to-noise ratio (SNR) and blindly decompose an image matrix into pairs of spatial and temporal factors. The former represent homogeneous tissues with different kinetic parameters and the latter represent the time activity curves that are observed in the corresponding homogeneous tissues. We employ U-Nets combined in parallel for DIP and each of the U-nets is used to extract each spatial factor decomposed from the data matrix. Experimental results show that the proposed method outperforms conventional methods and can extract spatial factors that represent the homogeneous tissues.

Referring expression comprehension aims to locate the object instance described by a natural language referring expression in an image. This task is compositional and inherently requires visual reasoning on top of the relationships among the objects in the image. Meanwhile, the visual reasoning process is guided by the linguistic structure of the referring expression. However, existing approaches treat the objects in isolation or only explore the first-order relationships between objects without being aligned with the potential complexity of the expression. Thus it is hard for them to adapt to the grounding of complex referring expressions. In this paper, we explore the problem of referring expression comprehension from the perspective of language-driven visual reasoning, and propose a dynamic graph attention network to perform multi-step reasoning by modeling both the relationships among the objects in the image and the linguistic structure of the expression. In particular, we construct a graph for the image with the nodes and edges corresponding to the objects and their relationships respectively, propose a differential analyzer to predict a language-guided visual reasoning process, and perform stepwise reasoning on top of the graph to update the compound object representation at every node. Experimental results demonstrate that the proposed method can not only significantly surpass all existing state-of-the-art algorithms across three common benchmark datasets, but also generate interpretable visual evidences for stepwise locating the objects referred to in complex language descriptions.

In this paper, we investigate the audio-visual event localization problem. This task is to localize a visible and audible event in a video. Previous methods first divide a video into short segments, and then fuse visual and acoustic features at the segment level. The duration of these segments is usually short, making the visual and acoustic feature of each segment possibly not well aligned. Direct concatenation of the two features at the segment level can be vulnerable to a minor temporal misalignment of the two signals. We propose a Dual Attention Matching (DAM) module to cover a longer video duration for better high-level event information modeling, while the local temporal information is attained by the global cross-check mechanism. Our premise is that one should watch the whole video to understand the high-level event, while shorter segments should be checked in detail for localization. Specifically, the global feature of one modality queries the local feature in the other modality in a bi-directional way. With temporal co-occurrence encoded between auditory and visual signals, DAM can be readily applied in various audio-visual event localization tasks, e.g., cross-modality localization, supervised event localization. Experiments on the AVE dataset show our method outperforms the state-of-the-art by a large margin.

Single-view 3D shape reconstruction is an important but challenging problem, mainly for two reasons. First, as shape annotation is very expensive to acquire, current methods rely on synthetic data, in which ground-truth 3D annotation is easy to obtain. However, this results in domain adaptation problem when applied to natural images. The second challenge is that there are multiple shapes that can explain a given 2D image. In this paper, we propose a framework to improve over these challenges using adversarial training. On one hand, we impose domain confusion between natural and synthetic image representations to reduce the distribution gap. On the other hand, we impose the reconstruction to be `realistic' by forcing it to lie on a (learned) manifold of realistic object shapes. Our experiments show that these constraints improve performance by a large margin over baseline reconstruction models. We achieve results competitive with the state of the art with a much simpler architecture.

Predicting structured outputs such as semantic segmentation relies on expensive per-pixel annotations to learn supervised models like convolutional neural networks. However, models trained on one data domain may not generalize well to other domains without annotations for model finetuning. To avoid the labor-intensive process of annotation, we develop a domain adaptation method to adapt the source data to the unlabeled target domain. We propose to learn discriminative feature representations of patches in the source domain by discovering multiple modes of patch-wise output distribution through the construction of a clustered space. With such representations as guidance, we use an adversarial learning scheme to push the feature representations of target patches in the clustered space closer to the distributions of source patches. In addition, we show that our framework is complementary to existing domain adaptation techniques and achieves consistent improvements on semantic segmentation. Extensive ablations and results are demonstrated on numerous benchmark datasets with various settings, such as synthetic-to-real and cross-city scenarios.

Multi-exit architectures, in which a stack of processing layers is interleaved with early output layers, allow the processing of a test example to stop early and thus save computation time and/or energy. In this work, we propose a new training procedure for multi-exit architectures based on the principle of knowledge distillation. The method encourages early exits to mimic later, more accurate exits, by matching their probability outputs. Experiments on CIFAR100 and ImageNet show that distillation-based training significantly improves the accuracy of early exits while maintaining state-of-the-art accuracy for late ones. The method is particularly beneficial when training data is limited and also allows a straight-forward extension to semi-supervised learning, i.e. make use also of unlabeled data at training time. Moreover, it takes only a few lines to implement and imposes almost no computational overhead at training time, and none at all at test time.

We present a general paradigm for dynamic 3D reconstruction from multiple independent and uncontrolled image sources having arbitrary temporal sampling density and distribution. Our graph-theoretic formulation models the spatio-temporal relationships among our observations in terms of the joint estimation of their 3D geometry and its discrete Laplace operator. Towards this end, we define a tri-convex optimization framework that leverages the geometric properties and dependencies found among a Euclidean shape-space and the discrete Laplace operator describing its local and global topology. We present a reconstructability analysis, experiments on motion capture data and multi-view image datasets, as well as explore applications to geometry-based event segmentation and data association.

We present DenseRaC, a novel end-to-end framework for jointly estimating 3D human pose and body shape from a monocular RGB image. Our two-step framework takes the body pixel-to-surface correspondence map (i.e., IUV map) as proxy representation and then performs estimation of parameterized human pose and shape. Specifically, given an estimated IUV map, we develop a deep neural network optimizing 3D body reconstruction losses and further integrating a render-and-compare scheme to minimize differences between the input and the rendered output, i.e., dense body landmarks, body part masks, and adversarial priors. To boost learning, we further construct a large-scale synthetic dataset (MOCA) utilizing web-crawled Mocap sequences, 3D scans and animations. The generated data covers diversified camera views, human actions and body shapes, and is paired with full ground truth. Our model jointly learns to represent the 3D human body from hybrid datasets, mitigating the problem of unpaired training data. Our experiments show that DenseRaC obtains superior performance against state of the art on public benchmarks of various human-related tasks.

We propose DeepHuman, an image-guided volume-to-volume translation CNN for 3D human reconstruction from a single RGB image. To reduce the ambiguities associated with the reconstruction of invisible areas, our method leverages a dense semantic representation generated from SMPL model as an additional input. One key feature of our network is that it fuses different scales of image features into the 3D space through volumetric feature transformation, which helps to recover accurate surface geometry. The surface details are further refined through a normal refinement network, which can be concatenated with the volume generation network using our proposed volumetric normal projection layer. We also contribute THuman, a 3D real-world human model dataset containing approximately 7000 models. The network is trained using training data generated from the dataset. Overall, due to the specific design of our network and the diversity in our dataset, our method enables 3D human model estimation given only a single image and outperforms state-of-the-art approaches.

Convolutional Neural Networks (CNNs) achieve impressive performance in a wide variety of fields. Their success benefited from a massive boost when very deep CNN models were able to be reliably trained. Despite their merits, CNNs fail to properly address problems with non-Euclidean data. To overcome this challenge, Graph Convolutional Networks (GCNs) build graphs to represent non-Euclidean data, borrow concepts from CNNs, and apply them in training. GCNs show promising results, but they are usually limited to very shallow models due to the vanishing gradient problem. As a result, most state-of-the-art GCN models are no deeper than 3 or 4 layers. In this work, we present new ways to successfully train very deep GCNs. We do this by borrowing concepts from CNNs, specifically residual/dense connections and dilated convolutions, and adapting them to GCN architectures. Extensive experiments show the positive effect of these deep GCN frameworks. Finally, we use these new concepts to build a very deep 56-layer GCN, and show how it significantly boosts performance (+3.7% mIoU over state-of-the-art) in the task of point cloud semantic segmentation. We believe that the community can greatly benefit from this work, as it opens up many opportunities for advancing GCN-based research.

Recent modern displays are now able to render high dynamic range (HDR), high resolution (HR) videos of up to 8K UHD (Ultra High Definition). Consequently, UHD HDR broadcasting and streaming have emerged as high quality premium services. However, due to the lack of original UHD HDR video content, appropriate conversion technologies are urgently needed to transform the legacy low resolution (LR) standard dynamic range (SDR) videos into UHD HDR versions. In this paper, we propose a joint super-resolution (SR) and inverse tone-mapping (ITM) framework, called Deep SR-ITM, which learns the direct mapping from LR SDR video to their HR HDR version. Joint SR and ITM is an intricate task, where high frequency details must be restored for SR, jointly with the local contrast, for ITM. Our network is able to restore fine details by decomposing the input image and focusing on the separate base (low frequency) and detail (high frequency) layers. Moreover, the proposed modulation blocks apply location-variant operations to enhance local contrast. The Deep SR-ITM shows good subjective quality with increased contrast and details, outperforming the previous joint SR-ITM method.