ICCV 2019 Papers

Human pose and shape are two important components of 2D human body. However, how to efficiently represent both of them in images is still an open question. In this paper, we propose the Triplet Representation for Body (TRB) --- a compact 2D human body representation, with skeleton keypoints capturing human pose information and contour keypoints containing human shape information. TRB not only preserves the flexibility of skeleton keypoint representation, but also contains rich pose and human shape information. Therefore, it promises broader application areas, such as human shape editing and conditional image generation. We further introduce the challenging problem of TRB estimation, where joint learning of human pose and shape is required. We construct several large-scale TRB estimation datasets, based on the popular 2D pose datasets LSP, MPII and COCO. To effectively solve TRB estimation, we propose a two-branch network (TRB-net) with three novel techniques, namely X-structure (Xs), Directional Convolution (DC) and Pairwise mapping (PM), to enforce multi-level message passing for joint feature learning. We evaluate our proposed TRB-net and several leading approaches on our proposed TRB datasets, and demonstrate the superiority of our method through extensive evaluations.

TASED-Net is a 3D fully-convolutional network architecture for video saliency detection. It consists of two building blocks: first, the encoder network extracts low-resolution spatiotemporal features from an input clip of several consecutive frames, and then the following prediction network decodes the encoded features spatially while aggregating all the temporal information. As a result, a single prediction map is produced from an input clip of multiple frames. Frame-wise saliency maps can be predicted by applying TASED-Net in a sliding-window fashion to a video. The proposed approach assumes that the saliency map of any frame can be predicted by considering a limited number of past frames. The results of our extensive experiments on video saliency detection validate this assumption and demonstrate that our fully-convolutional model with temporal aggregation method is effective. TASED-Net significantly outperforms previous state-of-the-art approaches on all three major large-scale datasets of video saliency detection: DHF1K, Hollywood2, and UCFSports. After analyzing the results qualitatively, we observe that our model is especially better at attending to salient moving objects.

One of the most successful approaches in Multi-View Stereo estimates a depth map and a normal map for each view via PatchMatch-based optimization and fuses them into a consistent 3D points cloud. This approach relies on photo-consistency to evaluate the goodness of a depth estimate. It generally produces very accurate results; however, the reconstructed model often lacks completeness, especially in correspondence of broad untextured areas where the photo-consistency metrics are unreliable. Assuming the untextured areas piecewise planar, in this paper we generate novel PatchMatch hypotheses so to expand reliable depth estimates in neighboring untextured regions. At the same time, we modify the photo-consistency measure such to favor standard or novel PatchMatch depth hypotheses depending on the textureness of the considered area. We also propose a depth refinement step to filter wrong estimates and to fill the gaps on both the depth maps and normal maps while preserving the discontinuities. The effectiveness of our new methods has been tested against several state of the art algorithms in the publicly available ETH3D dataset containing a wide variety of high and low-resolution images.

Despite well-established baselines, learning of scene depth and ego-motion from monocular video remains an ongoing challenge, specifically when handling scaling ambiguity issues and depth inconsistencies in image sequences. Much prior work uses either a supervised mode of learning or stereo images. The former is limited by the amount of labeled data, as it requires expensive sensors, while the latter is not always readily available as monocular sequences. In this work, we demonstrate the benefit of using geometric information from synthetic images, coupled with scene depth information, to recover the scale in depth and ego-motion estimation from monocular videos. We developed our framework using synthetic image-depth pairs and unlabeled real monocular images. We had three training objectives: first, to use deep feature alignment to reduce the domain gap between synthetic and monocular images to yield more accurate depth estimation when presented with only real monocular images at test time. Second, we learn scene specific representation by exploiting self-supervision coming from multi-view synthetic images without the need for depth labels. Third, our method uses single-view depth and pose networks, which are capable of jointly training and supervising one another mutually, yielding consistent depth and ego-motion estimates. Extensive experiments demonstrate that our depth and ego-motion models surpass the state-of-the-art, unsupervised methods and compare favorably to early supervised deep models for geometric understanding. We validate the effectiveness of our training objectives against standard benchmarks thorough an ablation study.

Reading text in the wild is a very challenging task due to the diversity of text instances and the complexity of natural scenes. Recently, the community has paid increasing attention to the problem of recognizing text instances with irregular shapes. One intuitive and effective way to handle this problem is to rectify irregular text to a canonical form before recognition. However, these methods might struggle when dealing with highly curved or distorted text instances. To tackle this issue, we propose in this paper a Symmetry-constrained Rectification Network (ScRN) based on local attributes of text instances, such as center line, scale and orientation. Such constraints with an accurate description of text shape enable ScRN to generate better rectification results than existing methods and thus lead to higher recognition accuracy. Our method achieves state-of-the-art performance on text with both regular and irregular shapes. Specifically, the system outperforms existing algorithms by a large margin on datasets that contain quite a proportion of irregular text instances, e.g., ICDAR 2015, SVT-Perspective and CUTE80.

We propose a symmetric graph convolutional autoencoder which produces a low-dimensional latent representation from a graph. In contrast to the existing graph autoencoders with asymmetric decoder parts, the proposed autoencoder has a newly designed decoder which builds a completely symmetric autoencoder form. For the reconstruction of node features, the decoder is designed based on Laplacian sharpening as the counterpart of Laplacian smoothing of the encoder, which allows utilizing the graph structure in the whole processes of the proposed autoencoder architecture. In order to prevent the numerical instability of the network caused by the Laplacian sharpening introduction, we further propose a new numerically stable form of the Laplacian sharpening by incorporating the signed graphs. In addition, a new cost function which finds a latent representation and a latent affinity matrix simultaneously is devised to boost the performance of image clustering tasks. The experimental results on clustering, link prediction and visualization tasks strongly support that the proposed model is stable and outperforms various state-of-the-art algorithms.

Training accurate deep neural networks (DNNs) in the presence of noisy labels is an important and challenging task. Though a number of approaches have been proposed for learning with noisy labels, many open issues remain. In this paper, we show that DNN learning with Cross Entropy (CE) exhibits overfitting to noisy labels on some classes ("easy" classes), but more surprisingly, it also suffers from significant under learning on some other classes ("hard" classes). Intuitively, CE requires an extra term to facilitate learning of hard classes, and more importantly, this term should be noise tolerant, so as to avoid overfitting to noisy labels. Inspired by the symmetric KL-divergence, we propose the approach of Symmetric cross entropy Learning (SL), boosting CE symmetrically with a noise robust counterpart Reverse Cross Entropy (RCE). Our proposed SL approach simultaneously addresses both the under learning and overfitting problem of CE in the presence of noisy labels. We provide a theoretical analysis of SL and also empirically show, on a range of benchmark and real-world datasets, that SL outperforms state-of-the-art methods. We also show that SL can be easily incorporated into existing methods in order to further enhance their performance.

Recent advances in image-to-image translation have led to some ways to generate multiple domain images through a single network. However, there is still a limit in creating an image of a target domain without a dataset on it. We propose a method to expand the concept of `multi-domain' from data to the loss area, and to combine the characteristics of each domain to create an image. First, we introduce a sym-parameter and its learning method that can mix various losses and can synchronize them with input conditions. Then, we propose Sym-parameterized Generative Network (SGN) using it. Through experiments, we confirmed that SGN could mix the characteristics of various data and losses, and it is possible to translate images to any mixed-domain without ground truths, such as 30% Van Gogh and 20% Monet and 40% snowy.

Normalization methods are essential components in convolutional neural networks (CNNs). They either standardize or whiten data using statistics estimated in predefined sets of pixels. Unlike existing works that design normalization techniques for specific tasks, we propose Switchable Whitening (SW), which provides a general form unifying different whitening methods as well as standardization methods. SW learns to switch among these operations in an end-to-end manner. It has several advantages. First, SW adaptively selects appropriate whitening or standardization statistics for different tasks (see Fig.1), making it well suited for a wide range of tasks without manual design. Second, by integrating benefits of different normalizers, SW shows consistent improvements over its counterparts in various challenging benchmarks. Third, SW serves as a useful tool for understanding the characteristics of whitening and standardization techniques. We show that SW outperforms other alternatives on image classification (CIFAR-10/100, ImageNet), semantic segmentation (ADE20K, Cityscapes), domain adaptation (GTA5, Cityscapes), and image style transfer (COCO). For example, without bells and whistles, we achieve state-of-the-art performance with 45.33% mIoU on the ADE20K dataset.

In this paper, we introduce a novel method for reconstructing surface normals and depth of dynamic objects in water. Past shape recovery methods have leveraged various visual cues for estimating shape (e.g., depth) or surface normals. Methods that estimate both compute one from the other. We show that these two geometric surface properties can be simultaneously recovered for each pixel when the object is observed underwater. Our key idea is to leverage multi-wavelength near-infrared light absorption along different underwater light paths in conjunction with surface shading. We derive a principled theory for this surface normals and shape from water method and a practical calibration method for determining its imaging parameters values. By construction, the method can be implemented as a one-shot imaging system. We prototype both an off-line and a video-rate imaging system and demonstrate the effectiveness of the method on a number of real-world static and dynamic objects. The results show that the method can recover intricate surface features that are otherwise inaccessible.

Developing deep learning techniques for geometric data is an active and fruitful research area. This paper tackles the problem of sphere-type surface learning by developing a novel surface-to-image representation. Using this representation we are able to quickly adapt successful CNN models to the surface setting. The surface-image representation is based on a covering map from the image domain to the surface. Namely, the map wraps around the surface several times, making sure that every part of the surface is well represented in the image. Differently from previous surface-to-image representations, we provide a low distortion coverage of all surface parts in a single image. Specifically, for the use case of learning spherical signals, our representation provides a low distortion alternative to several popular spherical parameterizations used in deep learning. We have used the surface-to-image representation to apply standard CNN architectures to 3D models including spherical signals. We show that our method achieves state of the art or comparable results on the tasks of shape retrieval, shape classification and semantic shape segmentation.

Subspace clustering is the problem of partitioning data drawn from a union of multiple subspaces. The most popular subspace clustering framework in recent years is the graph clustering-based approach, which performs subspace clustering in two steps: graph construction and graph clustering. Although both steps are equally important for accurate clustering, the vast majority of work has focused on improving the graph construction step rather than the graph clustering step. In this paper, we propose a novel graph clustering framework for robust subspace clustering. By incorporating a geometry-aware term with the spectral clustering objective, we encourage our framework to be robust to noise and outliers in given affinity matrices. We also develop an efficient expectation-maximization-based algorithm for optimization. Through extensive experiments on four real-world datasets, we demonstrate that the proposed method outperforms existing methods.

Human motion prediction is a challenging and important task in many computer vision application domains. Existing work only implicitly models the spatial structure of the human skeleton. In this paper, we propose a novel approach that decomposes the prediction into individual joints by means of a structured prediction layer that explicitly models the joint dependencies. This is implemented via a hierarchy of small-sized neural networks connected analogously to the kinematic chains in the human body as well as a joint-wise decomposition in the loss function. The proposed layer is agnostic to the underlying network and can be used with existing architectures for motion modelling. Prior work typically leverages the H3.6M dataset. We show that some state-of-the-art techniques do not perform well when trained and tested on AMASS, a recently released dataset 14 times the size of H3.6M. Our experiments indicate that the proposed layer increases the performance of motion forecasting irrespective of the base network, joint-angle representation, and prediction horizon. We furthermore show that the layer also improves motion predictions qualitatively. We make code and models publicly available at https://ait.ethz.ch/projects/2019/spl.

Recent saliency models extensively explore to incorporate multi-scale contextual information from Convolutional Neural Networks (CNNs). Besides direct fusion strategies, many approaches introduce message-passing to enhance CNN features or predictions. However, the messages are mainly transmitted in two ways, by feature-to-feature passing, and by prediction-to-prediction passing. In this paper, we add message-passing between features and predictions and propose a deep unified CRF saliency model . We design a novel cascade CRFs architecture with CNN to jointly refine deep features and predictions at each scale and progressively compute a final refined saliency map. We formulate the CRF graphical model that involves message-passing of feature-feature, feature-prediction, and prediction-prediction, from the coarse scale to the finer scale, to update the features and the corresponding predictions. Also, we formulate the mean-field updates for joint end-to-end model training with CNN through back propagation. The proposed deep unified CRF saliency model is evaluated over six datasets and shows highly competitive performance among the state of the arts.

Image inpainting techniques have shown significant improvements by using deep neural networks recently. However, most of them may either fail to reconstruct reasonable structures or restore fine-grained textures. In order to solve this problem, in this paper, we propose a two-stage model which splits the inpainting task into two parts: structure reconstruction and texture generation. In the first stage, edge-preserved smooth images are employed to train a structure reconstructor which completes the missing structures of the inputs. In the second stage, based on the reconstructed structures, a texture generator using appearance flow is designed to yield image details. Experiments on multiple publicly available datasets show the superior performance of the proposed network.

The performance of multi-task learning in Convolutional Neural Networks (CNNs) hinges on the design of feature sharing between tasks within the architecture. The number of possible sharing patterns are combinatorial in the depth of the network and the number of tasks, and thus hand-crafting an architecture, purely based on the human intuitions of task relationships can be time-consuming and suboptimal. In this paper, we present a probabilistic approach to learning task-specific and shared representations in CNNs for multi-task learning. Specifically, we propose "stochastic filter groups" (SFG), a mechanism to assign convolution kernels in each layer to "specialist" and "generalist" groups, which are specific to and shared across different tasks, respectively. The SFG modules determine the connectivity between layers and the structures of task-specific and shared representations in the network. We employ variational inference to learn the posterior distribution over the possible grouping of kernels and network parameters. Experiments demonstrate the proposed method generalises across multiple tasks and shows improved performance over baseline methods.

As continuous-wave time-of-flight (C-ToF) cameras become popular in 3D imaging applications, they need to contend with the problem of multi-camera interference (MCI). In a multi-camera environment, a ToF camera may receive light from the sources of other cameras, resulting in large depth errors. In this paper, we propose stochastic exposure coding (SEC), a novel approach for mitigating. SEC involves dividing a camera's integration time into multiple slots, and switching the camera off and on stochastically during each slot. This approach has two benefits. First, by appropriately choosing the on probability for each slot, the camera can effectively filter out both the AC and DC components of interfering signals, thereby mitigating depth errors while also maintaining high signal-to-noise ratio. This enables high accuracy depth recovery with low power consumption. Second, this approach can be implemented without modifying the C-ToF camera's coding functions, and thus, can be used with a wide range of cameras with minimal changes. We demonstrate the performance benefits of SEC with theoretical analysis, simulations and real experiments, across a wide range of imaging scenarios.

This paper tackles the problem of large-scale image-based localization (IBL) where the spatial location of a query image is determined by finding out the most similar reference images in a large database. For solving this problem, a critical task is to learn discriminative image representation that captures informative information relevant for localization. We propose a novel representation learning method having higher location-discriminating power. It provides the following contributions: 1) we represent a place (location) as a set of exemplar images depicting the same landmarks and aim to maximize similarities among intra-place images while minimizing similarities among inter-place images; 2) we model a similarity measure as a probability distribution on L_2-metric distances between intra-place and inter-place image representations; 3) we propose a new Stochastic Attraction and Repulsion Embedding (SARE) loss function minimizing the KL divergence between the learned and the actual probability distributions; 4) we give theoretical comparisons between SARE, triplet ranking and contrastive losses. It provides insights into why SARE is better by analyzing gradients. Our SARE loss is easy to implement and pluggable to any CNN. Experiments show that our proposed method improves the localization performance on standard benchmarks by a large margin. Demonstrating the broad applicability of our method, we obtained the third place out of 209 teams in the 2018 Google Landmark Retrieval Challenge. Our code and model are available at https://github.com/Liumouliu/deepIBL.

We propose StartNet to address Online Detection of Action Start (ODAS) where action starts and their associated categories are detected in untrimmed, streaming videos. Previous methods aim to localize action starts by learning feature representations that can directly separate the start point from its preceding background. It is challenging due to the subtle appearance difference near the action starts and the lack of training data. Instead, StartNet decomposes ODAS into two stages: action classification (using ClsNet) and start point localization (using LocNet). ClsNet focuses on per-frame labeling and predicts action score distributions online. Based on the predicted action scores of the past and current frames, LocNet conducts class-agnostic start detection by optimizing long-term localization rewards using policy gradient methods. The proposed framework is validated on two large-scale datasets, THUMOS'14 and ActivityNet. The experimental results show that StartNet significantly outperforms the state-of-the-art by 15%-30% p-mAP under the offset tolerance of 1-10 seconds on THUMOS'14, and achieves comparable performance on ActivityNet with 10 times smaller time offset.