CVPR 2017 Papers

In recent years, Discriminative Correlation Filter (DCF) based methods have significantly advanced the state-of-the-art in tracking. However, in the pursuit of ever increasing tracking performance, their characteristic speed and real-time capability have gradually faded. Further, the increasingly complex models, with massive number of trainable parameters, have introduced the risk of severe over-fitting. In this work, we tackle the key causes behind the problems of computational complexity and over-fitting, with the aim of simultaneously improving both speed and performance. We revisit the core DCF formulation and introduce: (i) a factorized convolution operator, which drastically reduces the number of parameters in the model; (ii) a compact generative model of the training sample distribution, that significantly reduces memory and time complexity, while providing better diversity of samples; (iii) a conservative model update strategy with improved robustness and reduced complexity. We perform comprehensive experiments on four benchmarks: VOT2016, UAV123, OTB-2015, and TempleColor. When using expensive deep features, our tracker provides a 20-fold speedup and achieves a 13.0% relative gain in Expected Average Overlap compared to the top ranked method in the VOT2016 challenge. Moreover, our fast variant, using hand-crafted features, operates at 60 Hz on a single CPU, while obtaining 65.0% AUC on OTB-2015.

Previous approaches for scene text detection have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with challenging scenarios, even when equipped with deep neural network models, because the overall performance is determined by the interplay of multiple stages and components in the pipelines. In this work, we propose a simple yet powerful pipeline that yields fast and accurate text detection in natural scenes. The pipeline directly predicts words or text lines of arbitrary orientations and quadrilateral shapes in full images, eliminating unnecessary intermediate steps (e.g., candidate aggregation and word partitioning), with a single neural network. The simplicity of our pipeline allows concentrating efforts on designing loss functions and neural network architecture. Experiments on standard datasets including ICDAR 2015, COCO-Text and MSRA-TD500 demonstrate that the proposed algorithm significantly outperforms state-of-the-art methods in terms of both accuracy and efficiency. On the ICDAR 2015 dataset, the proposed algorithm achieves an F-score of 0.7820 at 13.2fps at 720p resolution.

Time-of-flight (TOF) depth cameras provide robust depth inference at low power requirements in a wide variety of consumer and industrial applications. These cameras reconstruct a single depth frame from a given set of infrared (IR) frames captured over a very short exposure period. Operating in this mode the camera essentially forgets all information previously captured - and performs depth inference from scratch for every frame. We challenge this practice and propose using previously captured information when inferring depth. An inherent problem we have to address is camera motion over this longer period of collecting observations. We derive a probabilistic framework combining a simple but robust model of camera and object motion, together with an observation model. This combination allows us to integrate information over multiple frames while remaining robust to rapid changes. Operating the camera in this manner has implications in terms of both computational efficiency and how information should be captured. We address these two issues and demonstrate a realtime TOF system with robust temporal integration that improves depth accuracy over strong baseline methods including adaptive spatio-temporal filters.

Facial analysis in videos, including head pose estimation and facial landmark localization, is key for many applications such as facial animation capture, human activity recognition, and human-computer interaction. In this paper, we propose to use a recurrent neural network (RNN) for joint estimation and tracking of facial features in videos. We are inspired by the fact that the computation performed in an RNN bears resemblance to Bayesian filters, which have been used for tracking in many previous methods for facial analysis from videos. Bayesian filters used in these methods, however, require complicated, problem-specific design and tuning. In contrast, our proposed RNN-based method avoids such tracker-engineering by learning from training data, similar to how a convolutional neural network (CNN) avoids feature-engineering for image classification. As an end-to-end network, the proposed RNN-based method provides a generic and holistic solution for joint estimation and tracking of various types of facial features from consecutive video frames. Extensive experimental results on head pose estimation and facial landmark localization from videos demonstrate that the proposed RNN-based method outperforms frame-wise models and Bayesian filtering. In addition, we create a large-scale synthetic dataset for head pose estimation, with which we achieve state-of-the-art performance on a benchmark dataset.

While the ready availability of 3D scan data has influenced research throughout computer vision, less attention has focused on 4D data; that is 3D scans of moving non-rigid objects, captured over time. To be useful for vision research, such 4D scans need to be registered, or aligned, to a common topology. Consequently, extending mesh registration methods to 4D is important. Unfortunately, no ground-truth datasets are available for quantitative evaluation and comparison of 4D registration methods. To address this we create a novel dataset of high-resolution 4D scans of human subjects in motion, captured at 60 fps. We propose a new mesh registration method that uses both 3D geometry and texture information to register all scans in a sequence to a common reference topology. The approach exploits consistency in texture over both short and long time intervals and deals with temporal offsets between shape and texture capture. We show how using geometry alone results in significant errors in alignment when the motions are fast and non-rigid. We evaluate the accuracy of our registration and provide a dataset of 40,000 raw and aligned meshes. Dynamic FAUST extends the popular FAUST dataset to dynamic 4D data, and is available for research purposes at http://dfaust.is.tue.mpg.de.

A number of problems can be formulated as prediction on graph-structured data. In this work, we generalize the convolution operator from regular grids to arbitrary graphs while avoiding the spectral domain, which allows us to handle graphs of varying size and connectivity. To move beyond a simple diffusion, filter weights are conditioned on the specific edge labels in the neighborhood of a vertex. Together with the proper choice of graph coarsening, we explore constructing deep neural networks for graph classification. In particular, we demonstrate the generality of our formulation in point cloud classification, where we set the new state of the art, and on a graph classification dataset, where we outperform other deep learning approaches.

We present a new Cascaded Shape Regression (CSR) architecture, namely Dynamic Attention-Controlled CSR (DAC-CSR), for robust facial landmark detection on unconstrained faces. Our DAC-CSR divides facial landmark detection into three cascaded sub-tasks: face bounding box refinement, general CSR and attention-controlled CSR. The first two stages refine initial face bounding boxes and output intermediate facial landmarks. Then, an online dynamic model selection method is used to choose appropriate domain-specific CSRs for further landmark refinement. The key innovation of our DAC-CSR is the fault-tolerant mechanism, using fuzzy set sample weighting, for attention-controlled domain-specific model training. Moreover, we advocate data augmentation with a simple but effective 2D profile face generator, and context-aware feature extraction for better facial feature representation. Experimental results obtained on challenging datasets demonstrate the merits of our DAC-CSR over the state-of-the-art methods.

We propose Dual Attention Networks (DANs) which jointly leverage visual and textual attention mechanisms to capture fine-grained interplay between vision and language. DANs attend to specific regions in images and words in text through multiple steps and gather essential information from both modalities. Based on this framework, we introduce two types of DANs for multimodal reasoning and matching, respectively. The reasoning model allows visual and textual attentions to steer each other during collaborative inference, which is useful for tasks such as Visual Question Answering (VQA). In addition, the matching model exploits the two attention mechanisms to estimate the similarity between images and sentences by focusing on their shared semantics. Our extensive experiments validate the effectiveness of DANs in combining vision and language, achieving the state-of-the-art performance on public benchmarks for VQA and image-text matching.

In this paper, we propose an approach to the domain adaptation, dubbed Second- or Higher-order Transfer of Knowledge (So-HoT), based on the mixture of alignments of second- or higher-order scatter statistics between the source and target domains. The human ability to learn from few labeled samples is a recurring motivation in the literature for domain adaptation. Towards this end, we investigate the supervised target scenario for which few labeled target training samples per category exist. Specifically, we utilize two CNN streams: the source and target networks fused at the classifier level. Features from the fully connected layers fc7 of each network are used to compute second- or even higher-order scatter tensors; one per network stream per class. As the source and target distributions are somewhat different despite being related, we align the scatters of the two network streams of the same class (within-class scatters) to a desired degree with our bespoke loss while maintaining good separation of the between-class scatters. We train the entire network in end-to-end fashion. We provide evaluations on the standard Office benchmark (visual domains) and RGB-D combined with Caltech256 (depth-to-rgb transfer). We attain state-of-the-art results.

Recent progresses on deep discriminative and generative modeling have shown promising results on texture synthesis. However, existing feed-forward based methods trade off generality for efficiency, which suffer from many issues, such as shortage of generality (i.e., build one network per texture), lack of diversity (i.e., always produce visually identical output) and suboptimality (i.e., generate less satisfying visual effects). In this work, we focus on solving these issues for improved texture synthesis. We propose a deep generative feed-forward network which enables efficient synthesis of multiple textures within one single network and meaningful interpolation between them. Meanwhile, a suite of important techniques are introduced to achieve better convergence and diversity. With extensive experiments, we demonstrate the effectiveness of the proposed model and techniques for synthesizing a large number of textures and show its applications with the stylization.

In this work we study the task of image annotation, of which the goal is to describe an image using a few tags. Instead of predicting the full list of tags, here we target for providing a short list of tags under a limited number (e.g., 3), to cover as much information as possible of the image. The tags in such a short list should be representative and diverse. It means they are required to be not only corresponding to the contents of the image, but also be different to each other. To this end, we treat the image annotation as a subset selection problem based on the conditional determinantal point process (DPP) model, which formulates the representation and diversity jointly. We further explore the semantic hierarchy and synonyms among the candidate tags, and require that two tags in a semantic hierarchy or in a pair of synonyms should not be selected simultaneously. This requirement is then embedded into the sampling algorithm according to the learned conditional DPP model. Besides, we find that traditional metrics for image annotation (e.g., precision, recall and F1 score) only consider the representation, but ignore the diversity. Thus we propose new metrics to evaluate the quality of the selected subset (i.e., the tag list), based on the semantic hierarchy and synonyms. Human study through Amazon Mechanical Turk verifies that the proposed metrics are more close to the human's judgment than traditional metrics. Experiments on two benchmark datasets show that the proposed method can produce more representative and diverse tags, compared with existing image annotation methods.

A perennial problem in structure from motion (SfM) is visual ambiguity posed by repetitive structures. Recent disambiguating algorithms infer ambiguities mainly via explicit background context, thus face limitations in highly ambiguous scenes which are visually indistinguishable. Instead of analyzing local visual information, we propose a novel algorithm for SfM disambiguation that explores the global topology as encoded in photo collections. An important adaptation of this work is to approximate the available imagery using a manifold of viewpoints. We note that, while ambiguous images appear deceptively similar in appearance, they are actually located far apart on geodesics. We establish the manifold by adaptively identifying cameras with adjacent viewpoint, and detect ambiguities via a new measure, geodesic consistency. We demonstrate the accuracy and efficiency of the proposed approach on a range of complex ambiguity datasets, even including the challenging scenes without background conflicts.

The large pose discrepancy between two face images is one of the key challenges in face recognition. Conventional approaches for pose-invariant face recognition either perform face frontalization on, or learn a pose-invariant representation from, a non-frontal face image. We argue that it is more desirable to perform both tasks jointly to allow them to leverage each other. To this end, this paper proposes Disentangled Representation learning-Generative Adversarial Network (DR-GAN) with three distinct novelties. First, the encoder-decoder structure of the generator allows DR-GAN to learn a generative and discriminative representation, in addition to image synthesis. Second, this representation is explicitly disentangled from other face variations such as pose, through the pose code provided to the decoder and pose estimation in the discriminator. Third, DR-GAN can take one or multiple images as the input, and generate one unified representation along with an arbitrary number of synthetic images. Quantitative and qualitative evaluation on both controlled and in-the-wild databases demonstrate the superiority of DR-GAN over the state of the art.

Many computer vision problems are formulated as the optimization of a cost function. This approach faces two main challenges: (1) designing a cost function with a local optimum at an acceptable solution, and (2) developing an efficient numerical method to search for one (or multiple) of these local optima. While designing such functions is feasible in the noiseless case, the stability and location of local optima are mostly unknown under noise, occlusion, or missing data. In practice, this can result in undesirable local optima or not having a local optimum in the expected place. On the other hand, numerical optimization algorithms in high-dimensional spaces are typically local and often rely on expensive first or second order information to guide the search. To overcome these limitations, this paper proposes Discriminative Optimization (DO), a method that learns search directions from data without the need of a cost function. Specifically, DO explicitly learns a sequence of updates in the search space that leads to stationary points that correspond to desired solutions. We provide a formal analysis of DO and illustrate its benefits in the problem of 2D and 3D point cloud registration both in synthetic and range-scan data. We show that DO outperforms state-of-the-art algorithms by a large margin in terms of accuracy, robustness to perturbations, and computational efficiency.

For face recognition with image sets, while most existing works mainly focus on building robust set models with hand-crafted feature, it remains a research gap to learn better image representations which can closely match the subsequent image set modeling and classification. Taking sample covariance matrix as set model in the light of its recent promising success, we present a Discriminative Covariance oriented Representation Learning (DCRL) framework to bridge the above gap. The framework constructs a feature learning network (e.g. a CNN) to project the face images into a target representation space, and the network is trained towards the goal that the set covariance matrix calculated in the target space has maximum discriminative ability. To encode the discriminative ability of set covariance matrices, we elaborately design two different loss functions, which respectively lead to two different representation learning schemes, i.e., the Graph Embedding scheme and the Softmax Regression scheme. Both schemes optimize the whole network containing both image representation mapping and set model classification in a joint learning manner. The proposed method is extensively validated on three challenging and large scale databases for the task of face recognition with image sets, i.e., YouTube Celebrities, YouTube Face DB and Point-and-Shoot Challenge.

Short-term tracking is an open and challenging problem for which discriminative correlation filters (DCF) have shown excellent performance. We introduce the channel and spatial reliability concepts to DCF tracking and provide a novel learning algorithm for its efficient and seamless integration in the filter update and the tracking process. The spatial reliability map adjusts the filter support to the part of the object suitable for tracking. This allows tracking of non-rectangular objects as well as extending the search region. Channel reliability reflects the quality of the learned filter and it is used as a feature weighting coefficient in localization. Experimentally, with only two simple standard features, HOGs and Colornames, the novel CSR-DCF method -- DCF with Channel and Spatial Reliability -- achieves state-of-the-art results on VOT 2016, VOT 2015 and OTB. The CSR-DCF runs in real-time on a CPU.

Associating image regions with text queries has been recently explored as a new way to bridge visual and linguistic representations. A few pioneering approaches have been proposed based on recurrent neural language models trained generatively (e.g., generating captions), but achieving somewhat limited localization accuracy. To better address natural-language-based visual entity localization, we propose a discriminative approach. We formulate a discriminative bimodal neural network (DBNet), which can be trained by a classifier with extensive use of negative samples. Our training objective encourages better localization on single images, incorporates text phrases in a broad range, and properly pairs image regions with text phrases into positive and negative examples. Experiments on the Visual Genome dataset demonstrate the proposed DBNet significantly outperforms previous state-of-the-art methods both for localization on single images and for detection on multiple images. We we also establish an evaluation protocol for natural-language visual detection.

Learning to hash has been recognized to accomplish highly efficient storage and retrieval for large-scale visual data. Particularly, ranking-based hashing techniques have recently attracted broad research attention because ranking accuracy among the retrieved data is well explored and their objective is more applicable to realistic search tasks. However, directly optimizing discrete hash codes without continuous-relaxations on a nonlinear ranking objective is infeasible by either traditional optimization methods or even recent discrete hashing algorithms. To address this challenging issue, in this paper, we introduce a novel supervised hashing method, dubbed Discrete Semantic Ranking Hashing (DSeRH), which aims to directly embed semantic rank orders into binary codes. In DSeRH, a generalized Adaptive Discrete Minimization (ADM) approach is proposed to discretely optimize binary codes with the quadratic nonlinear ranking objective in an iterative manner and is guaranteed to converge quickly. Additionally, instead of using 0/1 independent labels to form rank orders as in previous works, we generate the listwise rank orders from the high-level semantic word embeddings which can quantitatively capture the intrinsic correlation between different categories. We evaluate our DSeRH, coupled with both linear and deep convolutional neural network (CNN) hash functions, on three image datasets, i.e., CIFAR-10, SUN397 and ImageNet100, and the results manifest that DSeRH can outperform the state-of-the-art ranking-based hashing methods.

This paper establishes the existence of observable footprints that reveal the "causal dispositions" of the object categories appearing in collections of images. We achieve this goal in two steps. First, we take a learning approach to observational causal discovery, and build a classifier that achieves state-of-the-art performance on finding the causal direction between pairs of random variables, given samples from their joint distribution. Second, we use our causal direction classifier to effectively distinguish between features of objects and features of their contexts in collections of static images. Our experiments demonstrate the existence of a relation between the direction of causality and the difference between objects and their contexts, and by the same token, the existence of observable signals that reveal the causal dispositions of objects.