ICCV 2017 Papers

This paper presents a novel deep regression network to extract geometric information from Light Field (LF) data. Our network builds upon u-shaped network architectures. Those networks involve two symmetric parts, an encoding and a decoding part. In the first part the network encodes relevant information from the given input into a set of high-level feature maps. In the second part the generated feature maps are then decoded to the desired output. To predict reliable and robust depth information the proposed network examines 3D subsets of the 4D LF called Epipolar Plane Image (EPI) volumes. An important aspect of our network is the use of 3D convolutional layers, that allow to propagate information from two spatial dimensions and one directional dimension of the LF. Compared to previous work this allows for an additional spatial regularization, which reduces depth artifacts and simultaneously maintains clear depth discontinuities. Experimental results show that our approach allows to create high-quality reconstruction results, which outperform current state-of-the-art Shape from Light Field (SfLF) techniques. The main advantage of the proposed approach is the ability to provide those high-quality reconstructions at a low computation time.

In this paper, we approach the problem of segmentation-free query-by-string word spotting for handwritten documents. In other words, we use methods inspired from computer vision and machine learning to search for words in large collections of digitized manuscripts. In particular, we are interested in historical handwritten texts, which are often far more challenging than modern printed documents. This task is important, as it provides people with a way to quickly find what they are looking for in large collections that are tedious and difficult to read manually. To this end, we introduce an end-to-end trainable model based on deep neural networks that we call Ctrl-F-Net. Given a full manuscript page, the model simultaneously generates region proposals, and embeds these into a distributed word embedding space, where searches are performed. We evaluate the model on common benchmarks for handwritten word spotting, outperforming the previous state-of-the-art segmentation-free approaches by a large margin, and in some cases even segmentation-based approaches. One interesting real-life application of our approach is to help historians to find and count specific words in court records that are related to women's sustenance activities and division of labor. We provide promising preliminary experiments that validate our method on this task.

In this paper, we propose the first higher frame rate video dataset (called Need for Speed - NfS) and benchmark for visual object tracking. The dataset consists of 100 videos (380K frames) captured with now commonly available higher frame rate (240 FPS) cameras from real world scenarios. All frames are annotated with axis aligned bounding boxes and all sequences are manually labelled with nine visual attributes - such as occlusion, fast motion, background clutter, etc. Our benchmark provides an extensive evaluation of many recent and state-of-the-art trackers on higher frame rate sequences. We ranked each of these trackers according to their tracking accuracy and real-time performance. One of our surprising conclusions is that at higher frame rates, simple trackers such as correlation filters outperform complex methods based on deep networks. This suggests that for practical applications (such as in robotics or embedded vision), one needs to carefully tradeoff bandwidth constraints associated with higher frame rate acquisition, computational costs of real-time analysis, and the required application accuracy. Our dataset and benchmark allows for the first time (to our knowledge) systematic exploration of such issues, and will be made available to allow for further research in this space.

Detecting logo frequency and duration in sports videos provides sponsors an effective way to evaluate their advertising efforts. However, general-purposed object detection methods cannot address all the challenges in sports videos. In this paper, we propose a mutual-enhanced approach that can improve the detection of a logo through the information obtained from other simultaneously occurred logos. In a Fast-RCNN-based framework, we first introduce a homogeneity-enhanced re-ranking method by analyzing the characteristics of homogeneous logos in each frame, including type repetition, color consistency, and mutual exclusion. Different from conventional enhance mechanism that improves the weak proposals with the dominant proposals, our mutual method can also enhance the relatively significant proposals with weak proposals. Mutual enhancement is also included in our frame propagation mechanism that improves logo detection by utilizing the continuity of logos across frames. We use a tennis video dataset and an associated logo collection for detection evaluation. Experiments show that the proposed method outperforms existing methods with a higher accuracy.

In this work, we address multimodal learning problem with Gaussian process latent variable models (GPLVMs) and their application to cross-modal retrieval. Existing GPLVM based studies generally impose individual priors over the model parameters and ignore the intrinsic relations among these parameters. Considering the strong complementarity between modalities, we propose a novel joint prior over the parameters for multimodal GPLVMs to propagate multimodal information in both kernel hyperparameter spaces and latent space. The joint prior is formulated as a harmonization constraint on the model parameters, which enforces the agreement among the modality-specific GP kernels and the similarity in the latent space. We incorporate the harmonization mechanism into the learning process of multimodal GPLVMs. The proposed methods are evaluated on three widely used multimodal datasets for cross-modal retrieval. Experimental results show that the harmonization mechanism is beneficial to the GPLVM algorithms for learning non-linear correlation among heterogeneous modalities.

In recent years, there have been a variety of proposals for high quality 3D reconstruction by fusion of depth and normal maps that contain good low and high frequency information respectively. Typically, these methods create an initial mesh representation of the complete object or scene being scanned. Subsequently, normal estimates are assigned to each mesh vertex and a mesh-normal fusion step is carried out. In this paper, we present a complete pipeline for such depth-normal fusion. The key innovations in our pipeline are twofold. Firstly, we introduce a global multi-view non-rigid refinement step that corrects for the non-rigid misalignment present in the depth and normal maps. We demonstrate that such a correction is crucial for preserving fine-scale 3D features in the final reconstruction. Secondly, despite adequate care, the averaging of multiple normals invariably results in blurring of 3D detail. To mitigate this problem, we propose an approach that selects one out of many available normals. Our global cost for normal selection incorporates a variety of desirable properties and can be efficiently solved using graph cuts. We demonstrate the efficacy of our approach in generating high quality 3D reconstructions of both synthetic and real 3D models and compare with existing methods in the literature.

We consider 4D shape reconstructions in multi-view environments and investigate how to exploit temporal redundancy for precision refinement. In addition to being beneficial to many dynamic multi-view scenarios this also enables larger scenes where such increased precision can compensate for the reduced spatial resolution per image frame. With precision and scalability in mind, we propose a symmetric (non-causal) local time-window geometric integration scheme over temporal sequences, where shape reconstructions are refined framewise by warping local and reliable geometric regions of neighboring frames to them. This is in contrast to recent comparable approaches targeting a different context with more compact scenes and real-time applications. These usually use a single dense volumetric update space or geometric template, which they causally track and update globally frame by frame, with limitations in scalability for larger scenes and in topology and precision with a template based strategy. Our template less and local approach is a first step towards temporal shape super-resolution. We show that it improves reconstruction accuracy by considering multiple frames. To this purpose, and in addition to real data examples, we introduce a multi-camera synthetic dataset that provides ground-truth data for mid-scale dynamic scenes.

We investigate methods for combining multiple self-supervised tasks---i.e., supervised tasks where data can be collected without manual labeling---in order to train a single visual representation. First, we provide an apples-to-apples comparison of four different self-supervised tasks using the very deep ResNet-101 architecture. We then combine tasks to jointly train a network. We also explore lasso regularization to encourage the network to factorize the information in its representation, and methods for "harmonizing" network inputs in order to learn a more unified representation. We evaluate all methods on ImageNet classification, PASCAL VOC detection, and NYU depth prediction. Our results show that deeper networks work better, and that combining tasks---even via a naive multi-head architecture---always improves performance. Our best joint network nearly matches the PASCAL performance of a model pre-trained on ImageNet classification, and matches the ImageNet network on NYU depth prediction.

In the field of connectomics, neuroscientists seek to identify cortical connectivity comprehensively. Neuronal boundary detection from the Electron Microscopy (EM) images is often done to assist the automatic reconstruction of neuronal circuit. But the segmentation of EM images is a challenging problem, as it requires the detector to be able to detect both filament-like thin and blob-like thick membrane, while suppressing the ambiguous intracellular structure. In this paper, we propose multi-stage multi-recursiveinput fully convolutional networks to address this problem. The multiple recursive inputs for one stage, i.e., the multiple side outputs with different receptive field sizes learned from the lower stage, provide multi-scale contextual boundary information for the consecutive learning. This design is biologically-plausible, as it likes a human visual system to compare different possible segmentation solutions to address the ambiguous boundary issue. Our multi-stage networks are trained end-to-end. It achieves promising results on two public available EM segmentation datasets, the mouse piriform cortex dataset and the ISBI 2012 EM dataset.

Person Re-identification (re-id) aims to match people across non-overlapping camera views in a public space. It is a challenging problem because many people captured in surveillance videos wear similar clothes. Consequently, the differences in their appearance are often subtle and only detectable at the right location and scales. Existing re-id models, particularly the recently proposed deep learning based ones match people at a single scale. In contrast, in this paper, a novel multi-scale deep learning model is proposed. Our model is able to learn deep discriminative feature representations at different scales and automatically determine the most suitable scales for matching. The importance of different spatial locations for extracting discriminative features is also learned explicitly. Experiments are carried out to demonstrate that the proposed model outperforms the state-of-the art on a number of benchmarks.

Visual question answering (VQA) is challenging because it requires a simultaneous understanding of both the visual content of images and the textual content of questions. The approaches used to represent the images and questions in a fine-grained manner and questions and to fuse these multi-modal features play key roles in performance. Bilinear pooling based models have been shown to outperform traditional linear models for VQA, but their high-dimensional representations and high computational complexity may seriously limit their applicability in practice. For multi-modal feature fusion, here we develop a Multi-modal Factorized Bilinear (MFB) pooling approach to efficiently and effectively combine multi-modal features, which results in superior performance for VQA compared with other bilinear pooling approaches. For fine-grained image and question representation, we develop a co-attention mechanism using an end-to-end deep network architecture to jointly learn both the image and question attentions. Combining the proposed MFB approach with co-attention learning in a new network architecture provides a unified model for VQA. Our experimental results demonstrate that the single MFB with co-attention model achieves new state-of-the-art performance on the real-world VQA dataset. Code available at https://github.com/yuzcccc/mfb

Detecting pedestrians that are partially occluded remains a challenging problem due to variations and uncertainties of partial occlusion patterns. Following a commonly used framework of handling partial occlusions by part detection, we propose a multi-label learning approach to jointly learn part detectors to capture partial occlusion patterns. The part detectors share a set of decision trees via boosting to exploit part correlations and also reduce the computational cost of applying these part detectors. The learned decision trees capture the overall distribution of all the parts. When used as a pedestrian detector individually, our part detectors learned jointly show better performance than their counterparts learned separately in different occlusion situations. The learned part detectors can be further integrated to better detect partially occluded pedestrians. Experiments on the Caltech dataset show state-of-the-art performance of our approach for detecting heavily occluded pedestrians.

This paper proposes a novel deep architecture to address multi-label image recognition, a fundamental and practical task towards general visual understanding. Current solutions for this task usually rely on an extra step of extracting hypothesis regions (i.e., region proposals), resulting in redundant computation and sub-optimal performance. In this work, we achieve the interpretable and contextualized multi-label image classification by developing a recurrent memorized-attention module. This module consists of two alternately performed components: i) a spatial transformer layer to locate attentional regions from the convolutional feature maps in a region-proposal-free way and ii) a LSTM (Long-Short Term Memory) sub-network to sequentially predict semantic labeling scores on the located regions while capturing the global dependencies of these regions. The LSTM also output the parameters for computing the spatial transformer. On large-scale benchmarks of multi-label image classification (e.g., MS-COCO and PASCAL VOC 07), our approach demonstrates superior performances over other existing state-of-the-arts in both accuracy and efficiency.

Most of the existing denoising algorithms are developed for grayscale images. It is not trivial to extend them for color image denoising since the noise statistics in R, G, and B channels can be very different for real noisy images. In this paper, we propose a multi-channel (MC) optimization model for real color image denoising under the weighted nuclear norm minimization (WNNM) framework. We concatenate the RGB patches to make use of the channel redundancy, and introduce a weight matrix to balance the data fidelity of the three channels in consideration of their different noise statistics. The proposed MC-WNNM model does not have an analytical solution. We reformulate it into a linear equality-constrained problem and solve it via alternating direction method of multipliers. Each alternative updating step has a closed-form solution and the convergence can be guaranteed. Experiments on both synthetic and real noisy image datasets demonstrate the superiority of the proposed MC-WNNM over state-of-the-art denoising methods.

Low-rank and sparse representation based methods have attracted wide attention in background subtraction and moving object detection, where moving objects in the scene are modeled as pixel-wise sparse outliers. Since in real scenarios moving objects are also structurally sparse, recently researchers have attempted to extract moving objects using structured sparse outliers. Although existing methods with structured sparsity-inducing norms produce promising results, they are still vulnerable to various illumination changes that frequently occur in real environments, specifically for time-lapse image sequences where assumptions about sparsity between images such as group sparsity are not valid. In this paper, we first introduce a prior map obtained by illumination invariant representation of images. Next, we propose a low-rank and invariant sparse decomposition using the prior map to detect moving objects under significant illumination changes. Experiments on challenging benchmark datasets demonstrate the superior performance of our proposed method under complex illumination changes.

This paper presents an automated monocular-camera-based computer vision system for autonomous self-backing-up a vehicle towards a trailer, by continuously estimating the 3D trailer coupler position and feeding it to the vehicle control system, until the alignment of the tow hitch with the trailers coupler. This system is made possible through our proposed distance-driven Multiplexer-CNN method, which selects the most suitable CNN using the estimated coupler-to-vehicle distance. The input of the multiplexer is a group made of a CNN detector, trackers, and 3D localizer. In the CNN detector, we propose a novel algorithm to provide a presence confidence score with each detection. The score reflects the existence of the target object in a region, as well as how accurate is the 2D target detection. We demonstrate the accuracy and efficiency of the system on a large trailer database. Our system achieves an estimation error of 1.4 cm when the ball reaches the coupler, while running at 18.9 FPS on a regular PC.

Free-head 3D gaze tracking outputs both the eye location and the gaze vector in 3D space, and it has wide applications in scenarios such as driver monitoring, advertisement analysis and surveillance. A reliable and low-cost monocular solution is critical for pervasive usage in these areas. Noticing that a gaze vector is a composition of head pose and eyeball movement in a geometrically deterministic way, we propose a novel gaze transform layer to connect separate head pose and eyeball movement models. The proposed decomposition does not suffer from head-gaze correlation overfitting and makes it possible to use datasets existing for other tasks. To add stronger supervision for better network training, we propose a two-step training strategy, which first trains sub-tasks with rough labels and then jointly trains with accurate gaze labels. To enable good cross-subject performance under various conditions, we collect a large dataset which has full coverage of head poses and eyeball movements, contains 200 subjects, and has diverse illumination conditions. Our deep solution achieves state-of-the-art gaze tracking accuracy, reaching 5.6 degrees cross-subject prediction error using a small network running at 1000 fps on a s ingle CPU (excluding face alignment time) and 4.3 degrees cross-subject error with a deeper network.

This paper proposes a new approach for monocular dense 3D reconstruction of a complex dynamic scene from two perspective frames. By applying superpixel oversegmentation to the image, we model a generically dynamic (hence non-rigid) scene with a piecewise planar and rigid approximation. In this way, we reduce the dynamic reconstruction problem to a "3D jigsaw puzzle" problem which takes pieces from an unorganized "soup of superpixels". We show that our method provides an effective solution to the inherent relative scale ambiguity in structure-from-motion. Since our method does not assume a template prior, or per-object segmentation, or knowledge about the rigidity of the dynamic scene, it is applicable to a wide range of scenarios. Extensive experiments on both synthetic and real monocular sequences demonstrate the superiority of our method compared with the state-of-the-art methods.

This paper aims at estimating full-body 3D human poses from monocular images of which the biggest challenge is the inherent ambiguity introduced by lifting the 2D pose into 3D space. We propose a novel framework focusing on reducing this ambiguity by predicting the depth of human joints based on 2D human joint locations and body part images. Our approach is built on a two-level hierarchy of Long Short-Term Memory (LSTM) Networks which can be trained end-to-end. The first level consists of two components: 1) a skeleton-LSTM which learns the depth information from global human skeleton features; 2) a patch-LSTM which utilizes the local image evidence around joint locations. The both networks have tree structure defined on the kinematic relation of human skeleton, thus the information at different joints is broadcast through the whole skeleton in a top-down fashion. The two networks are first pre-trained separately on different data sources and then aggregated in the second layer for final depth prediction. The empirical evaluation on Human3.6M and HHOI dataset demonstrates the advantage of combining global 2D skeleton and local image patches for depth prediction, and our superior quantitative and qualitative performance relative to state-of-the-art methods.