ICCV 2017 Papers

Learning-based hashing methods are widely used for nearest neighbor retrieval, and recently, online hashing methods have demonstrated good performance-complexity trade-offs by learning hash functions from streaming data. In this paper, we first address a key challenge for online hashing: the binary codes for indexed data must be recomputed to keep pace with updates to the hash functions. We propose an efficient quality measure for hash functions, based on an information-theoretic quantity, mutual information, and use it successfully as a criterion to eliminate unnecessary hash table updates. Next, we also show how to optimize the mutual information objective using stochastic gradient descent. We thus develop a novel hashing method, MIHash, that can be used in both online and batch settings. Experiments on image retrieval benchmarks (including a 2.5M image dataset) confirm the effectiveness of our formulation, both in reducing hash table recomputations and in learning high-quality hash functions.

Low rank tensor completion, which solves a linear inverse problem with the principle of parsimony, is a powerful technique used in many application domains in computer vision and pattern recognition. As a surrogate function of the matrix rank that is non-convex and discontinuous, the nuclear norm is often used instead to derive efficient algorithms for recovering missing information in matrices and higher order tensors. However, the nuclear norm is a loose approximation of the matrix rank, and what is more, the tensor nuclear norm is not guaranteed to be the tightest convex envelope of a multilinear rank. Alternative algorithms either require specifying/tuning several parameters (e.g., the tensor rank), and/or have a performance far from reaching the theoretical limit where the number of observed elements equals the degree of freedom in the unknown low-rank tensor. In this paper, we propose a pseudo-Bayesian approach, where a Bayesian-inspired cost function is adjusted using appropriate approximations that lead to desirable attributes including concavity and symmetry. Although deviating from the original Bayesian model, the resulting non-convex cost function is proved to have the ability to recover the true tensor with a low multilinear rank. A computational efficient algorithm is derived to solve the resulting non-convex optimization problem. We demonstrate the superior performance of the proposed algorithm in comparison with state-of-the-art alternatives by conducting extensive experiments on both synthetic data and several visual data recovery tasks.

We present a robust algorithm for personalizing a sphere-mesh tracking model to a user from a collection of depth measurements. Our core contribution is to demonstrate how simple geometric reasoning can be exploited to build a shape-space, and how its performance is comparable to shape-spaces constructed from datasets of carefully calibrated models. We achieve this goal by first re-parameterizing the geometry of the tracking template, and introducing a multi-stage calibration optimization. Our novel parameterization decouples the degrees of freedom for pose and shape, resulting in improved convergence properties. Our analytically differentiable multi-stage calibration pipeline optimizes for the model in the natural low-dimensional space of local anisotropic scalings, leading to an effective solution that can be easily embedded in other tracking/calibration algorithms. Compared to existing sphere-mesh calibration algorithms, quantitative experiments assess our algorithm possesses a larger convergence basin, and our personalized models allows to perform motion tracking with superior accuracy. Code and data are available at http://github.com/edoRemelli/hadjust

Numerous computer vision problems such as stereo depth estimation, object-class segmentation and foreground/background segmentation can be formulated as per-pixel image labeling tasks. Given one or many images as input, the desired output of these methods is usually a spatially smooth assignment of labels. The large amount of such computer vision problems has lead to significant research efforts, with the state of art moving from CRF-based approaches to deep CNNs and more recently, hybrids of the two. Although these approaches have significantly advanced the state of the art, the vast majority has solely focused on improving quantitative results and are not designed for low-compute scenarios. In this paper, we present a new general framework for a variety of computer vision labeling tasks, called HashMatch. Our approach is designed to be both fully parallel, i.e. each pixel is independently processed, and low-compute, with a model complexity an order of magnitude less than existing CNN and CRF-based approaches. We evaluate HashMatch extensively on several problems such as disparity estimation, image retrieval, feature approximation and background subtraction, for which HashMatch achieves high computational efficiency while producing high quality results.

Recently, CNN-based models have achieved remarkable success in image-based salient object detection (SOD). In these models, a key issue is to find a proper network architecture that best fits for the task of SOD. Toward this end, this paper proposes two-stream fixation-semantic CNNs, whose architecture is inspired by the fact that salient objects in complex images can be unambiguously annotated by selecting the pre-segmented semantic objects that receive the highest fixation density in eye-tracking experiments. In the two-stream CNNs, a fixation stream is pre-trained on eye-tracking data whose architecture well fits for the task of fixation prediction, and a semantic stream is pre-trained on images with semantic tags that has a proper architecture for semantic perception. By fusing these two streams into an inception-segmentation module and jointly fine-tuning them on images with manually annotated salient objects, the proposed networks show impressive performance in segmenting salient objects. Experimental results show that our approach outperforms 10 state-of-the-art models (5 deep, 5 non-deep) on 4 datasets.

We consider the question: what can be learnt by looking at and listening to a large number of unlabelled videos? There is a valuable, but so far untapped, source of information contained in the video itself -- the correspondence between the visual and the audio streams, and we introduce a novel "Audio-Visual Correspondence" learning task that makes use of this. Training visual and audio networks from scratch, without any additional supervision other than the raw unconstrained videos themselves, is shown to successfully solve this task, and, more interestingly, result in good visual and audio representations. These features set the new state-of-the-art on two sound classification benchmarks, and perform on par with the state-of-the-art self-supervised approaches on ImageNet classification. We also demonstrate that the network is able to localize objects in both modalities, as well as perform fine-grained recognition tasks.

One-shot pose estimation for tasks such as body joint localization, camera pose estimation, and object tracking are generally noisy, and temporal filters have been extensively used for regularization. One of the most widely-used methods is the Kalman filter, which is both extremely simple and general. However, Kalman filters require a motion model and measurement model to be specified a priori, which burdens the modeler and simultaneously demands that we use explicit models that are often only crude approximations of reality. For example, in the pose-estimation tasks mentioned above, it is common to use motion models that assume constant velocity or constant acceleration, and we believe that these simplified representations are severely inhibitive. In this work, we propose to instead learn rich, dynamic representations of the motion and noise models. In particular, we propose learning these models from data using long short-term memory, which allows representations that depend on all previous observations and all previous states. We evaluate our method using three of the most popular pose estimation tasks in computer vision, and in all cases we obtain state-of-the-art performance.

Texture classification has been extensively studied in computer vision. Recent research shows that the combination of Fisher vector (FV) encoding and convolutional neural network (CNN) provides significant improvement in texture classification over the previous feature representation methods. However, by truncating the CNN model at the last convolutional layer, the CNN-based FV descriptors would not incorporate the full capability of neural networks in feature learning. In this study, we propose that we can further transform the CNN-based FV descriptors in a neural network model to obtain more discriminative feature representations. In particular, we design a locally-transferred Fisher vector (LFV) method, which involves a multi-layer neural network model containing locally connected layers to transform the input FV descriptors with filters of locally shared weights. The network is optimized based on the hinge loss of classification, and transferred FV descriptors are then used for image classification. Our results on three challenging texture image datasets show improved performance over the state of the art.

We consider retrieving a specific temporal segment, or moment, from a video given a natural language text description. Methods designed to retrieve whole video clips with natural language determine what occurs in a video but not when. To address this issue, we propose the Moment Context Network (MCN) which effectively localizes natural language queries in videos by integrating local and global video features over time. A key obstacle to training our MCN model is that current video datasets do not include pairs of localized video segments and referring expressions, or text descriptions which uniquely identify a corresponding moment. Therefore, we collect the Distinct Describable Moments (DiDeMo) dataset which consists of over 10,000 unedited, personal videos in diverse visual settings with pairs of localized video segments and referring expressions. We demonstrate that MCN outperforms several baseline methods and believe that our initial results together with release of DiDeMo will inspire further research on localizing video moments with natural language.

Local-to global point cloud registration is a challenging task due to the substantial differences between these two types of data, and the different techniques used to acquire them. Global clouds cover large-scale environments and are usually acquired aerially, e.g., 3D modeling of a city using Airborne Laser Scanning (ALS). In contrast, local clouds are often acquired from ground level and at a much smaller range, for example, using Terrestrial Laser Scanning (TLS). The differences are often manifested in point density distribution, occlusions nature, and measurement noise. As a result of these differences, existing point cloud registration approaches, such as keypoint-based registration, tend to fail. We improve upon a different approach, recently proposed, based on converting the global cloud into a viewpoint-based cloud dictionary. We propose a local-to-global registration method where we replace the dictionary clouds with viewpoint descriptors, consisting of panoramic range-images. We then use an efficient dictionary search in the Discrete Fourier Transform (DFT) domain, using phase correlation, to rapidly find plausible transformations from the local to the global reference frame. We demonstrate our method's significant advantages over the previous cloud dictionary approach, in terms of computational efficiency and memory requirements. In addition, We show its superior registration performance in comparison to a state-of-the-art, keypoint-based method (FPFH). For the evaluation, we use a challenging dataset of TLS local clouds and an ALS large-scale global cloud, in an urban environment.

Existing video event classification approaches suffer from limited human-labeled semantic annotations. Weak semantic annotations can be harvested from Web-knowledge without involving any human interaction. However such weak annotations are noisy, thus can not be effectively utilized without distinguishing its reliability. In this paper, we propose a novel approach to automatically maximize the utility of weak semantic annotations (formalized as the semantic relevance of video shots to the target event) to facilitate video event classification. A novel attention model is designed to determine the attention scores of video shots, where the weak semantic relevance is considered as attentional guidance. Specifically, our model jointly optimizes two objectives at different levels. The first one is the classification loss corresponding to video-level groundtruth labels, and the second is the shot-level relevance loss corresponding to weak semantic relevance. We use a long short-term memory (LSTM) layer to capture the temporal information carried by the shots of a video. In each timestep, the LSTM employs the attention model to weight the current shot under the guidance of its weak semantic relevance to the event of interest. Thus, we can automatically exploit weak semantic relevance to assist video event classification. Extensive experiments have been conducted on three complex large-scale video event datasets i.e., MEDTest14, ActivityNet and FCVID. Our approach achieves the state-of-the-art classification performance on all three datasets. The significant performance improvement upon the conventional attention model also demonstrates the effectiveness of our model.

Unsupervised learning with generative adversarial networks (GANs) has proven hugely successful. Regular GANs hypothesize the discriminator as a classifier with the sigmoid cross entropy loss function. However, we found that this loss function may lead to the vanishing gradients problem during the learning process. To overcome such a problem, we propose in this paper the Least Squares Generative Adversarial Networks (LSGANs) which adopt the least squares loss function for the discriminator. We show that minimizing the objective function of LSGAN yields minimizing the Pearson Chi^2 divergence. There are two benefits of LSGANs over regular GANs. First, LSGANs are able to generate higher quality images than regular GANs. Second, LSGANs perform more stable during the learning process. We evaluate LSGANs on LSUN and CIFAR-10 datasets and the experimental results show that the images generated by LSGANs are of better quality than the ones generated by regular GANs. We also conduct two comparison experiments between LSGANs and regular GANs to illustrate the stability of LSGANs.

Image understanding enables better reconstruction of the physical world from images and videos. Existing methods focus largely on geometry and visual appearance of the reconstructed scene. In this paper, we extend the frontier in image understanding and present a new technique to recover the material properties of cloth from a video.Previous cloth material recovery methods often require markers or complex experimental set-up to acquire physical properties, or are limited to certain types of images/videos. Our approach takes advantages of the appearance changes of the moving cloth to infer its physical properties. To extract information about the cloth, our method characterizes both the motion space and the visual appearance of the cloth geometry. We apply the Convolutional Neural Network (CNN) and the Long Short Term Memory (LSTM) neural network to material recovery of cloth properties from videos. We also exploit simulated data to help statistical learning of mapping between the visual appearance and motion dynamics of the cloth. The effectiveness of our method is demonstrated via validation using simulated datasets and real-life recorded videos.

We present an algorithm to directly restore a clear high-resolution image from a blurry low-resolution input. This problem is highly ill-posed and the basic assumptions for existing super-resolution methods (requiring clear input) and deblurring methods (requiring high-resolution input) no longer hold. We focus on face and text images and adopt a generative adversarial network (GAN) to learn a category-specific prior to solve this problem. However, the basic GAN formulation does not generate realistic high-resolution images. In this work, we introduce novel training losses that help recover fine details. We also present a multi-class GAN that can process multi-class image restoration tasks, i.e., face and text images, using a single generator network. Extensive experiments demonstrate that our method performs favorably against the state-of-the-art methods on both synthetic and real-world images at a lower computational cost.

Most recent approaches to monocular 3D human pose estimation rely on Deep Learning. They typically involve regressing from an image to either 3D joint coordinates directly or 2D joint locations from which 3D coordinates are inferred. Both approaches have their strengths and weaknesses and we therefore propose a novel architecture designed to deliver the best of both worlds by performing both simultaneously and fusing the information along the way. At the heart of our framework is a trainable fusion scheme that learns how to fuse the information optimally instead of being hand-designed. This yields significant improvements upon the state-of-the-art on standard 3D human pose estimation benchmarks.

Low-cost consumer depth cameras and deep learning have enabled reasonable 3D hand pose estimation from single depth images. In this paper, we present an approach that estimates 3D hand pose from regular RGB images. This task has far more ambiguities due to the missing depth information. To this end, we propose a deep network that learns a network-implicit 3D articulation prior. Together with detected keypoints in the images, this network yields good estimates of the 3D pose. We introduce a large scale 3D hand pose dataset based on synthetic hand models for training the involved networks. Experiments on a variety of test sets, including one on sign language recognition, demonstrate the feasibility of 3D hand pose estimation on single color images.

The ability to ask questions is a powerful tool to gather information in order to learn about the world and resolve ambiguities. In this paper, we explore a novel problem of generating discriminative questions to help disambiguate visual instances. Our work can be seen as a complement and new extension to the rich research studies on image captioning and question answering. We introduce the first large-scale dataset with over 10,000 carefully annotated images-question tuples to facilitate benchmarking. In particular, each tuple consists of a pair of images and 4.6 discriminative questions (as positive samples) and 5.9 non-discriminative questions (as negative samples) on average. In addition, we present an effective method for visual discriminative question generation. The method can be trained in a weakly supervised manner without discriminative images-question tuples but just existing visual question answering datasets. Promising results are shown against representative baselines through quantitative evaluations and user studies.

What defines a visual style? Fashion styles emerge organically from how people assemble outfits of clothing, making them difficult to pin down with a computational model. Low-level visual similarity can be too specific to detect stylistically similar images, while manually crafted style categories can be too abstract to capture subtle style differences. We propose an unsupervised approach to learn a style-coherent representation. Our method leverages probabilistic polylingual topic models based on visual attributes to discover a set of latent style factors. Given a collection of unlabeled fashion images, our approach mines for the latent styles, then summarizes outfits by how they mix those styles. Our approach can organize galleries of outfits by style without requiring any style labels. Experiments on over 100K images demonstrate its promise for retrieving, mixing, and summarizing fashion images by their style.

Many prediction tasks contain uncertainty. In some cases, uncertainty is inherent in the task itself. In future prediction, for example, many distinct outcomes are equally valid. In other cases, uncertainty arises from the way data is labeled. For example, in object detection, many objects of interest often go unlabeled, and in human pose estimation, occluded joints are often labeled with ambiguous values. In this work we focus on a principled approach for handling such scenarios. In particular, we propose a framework for reformulating existing single-prediction models as multiple hypothesis prediction (MHP) models and an associated meta loss and optimization procedure to train them. To demonstrate our approach, we consider four diverse applications: human pose estimation, future prediction, image classification and segmentation. We find that MHP models outperform their single-hypothesis counterparts in all cases, and that MHP models simultaneously expose valuable insights into the variability of predictions.