ICCV 2017 Papers

Imagery texts are usually organized as a hierarchy of several visual elements, i.e. characters, words, text lines and text blocks. Among these elements, character is the most basic one for various languages such as Western, Chinese, Japanese, mathematical expression and etc. It is natural and convenient to construct a common text detection engine based on character detectors. However, training character detectors requires a vast of location annotated characters, which are expensive to obtain. Actually, the existing real text datasets are mostly annotated in word or line level. To remedy this dilemma, we propose a weakly supervised framework that can utilize word annotations, either in tight quadrangles or the more loose bounding boxes, for character detector training. When applied in scene text detection, we are thus able to train a robust character detector by exploiting word annotations in the rich large-scale real scene text datasets, e.g. ICDAR15 [??] and COCO-text [??]. The character detector acts as a key role in the pipeline of our text detection engine. It achieves the state-of-the-art performance on several challenging scene text detection benchmarks. We also demonstrate the flexibility of our pipeline by various scenarios, including deformed text detection and math expression recognition.

Domain adaption (DA) allows machine learning methods trained on data sampled from one distribution to be applied to data sampled from another. It is thus of great practical importance to the application of such methods. Despite the fact that tensor representations are widely used in Computer Vision to capture multi-linear relationships that affect the data, most existing DA methods are applicable to vectors only. This renders them incapable of reflecting and preserving important structure in many problems. We thus propose here a learning-based method to adapt the source and target tensor representations directly, without vectorization. In particular, a set of alignment matrices is introduced to align the tensor representations from both domains into the invariant tensor subspace. These alignment matrices and the tensor subspace are modeled as a joint optimization problem and can be learned adaptively from the data using the proposed alternative minimization scheme. Extensive experiments show that our approach is capable of preserving the discriminative power of the source domain, of resisting the effects of label noise, and works effectively for small sample sizes, and even one-shot DA. We show that our method outperforms the state-of-the-art on the task of cross-domain visual recognition in both efficacy and efficiency, and particularly that it outperforms all comparators when applied to DA of the convolutional activations of deep convolutional networks.

A large number of very popular team sports involve the act of one team trying to score a goal against the other. During this game play, defending players constantly try to predict the next move of the attackers to prevent them from scoring, whereas attackers constantly try to predict the next move of the defenders in order to defy them and score. Such behavior is a prime example of the general human faculty to make predictions about the future and is an important facet of human intelligence. An algorithmic solution to learning a model of the external world from sensory inputs in order to make forecasts is an important unsolved problem. In this work we develop a generic framework for forecasting future events in team sports videos directly from visual inputs. We introduce water polo and basketball datasets towards this end and compare the predictions of the proposed methods against expert and non-expert humans.

How much does a single image reveal about the environment it was taken in? In this paper, we investigate how much of that information can be retrieved from a foreground object, combined with the background (i.e. the visible part of the environment). Assuming it is not perfectly diffuse, the foreground object acts as a complexly shaped and far-from-perfect mirror. An additional challenge is that its appearance confounds the light coming from the environment with the unknown materials it is made of. We propose a learning-based approach to predict the environment from multiple reflectance maps that are computed from approximate surface normals. The proposed method allows us to jointly model the statistics of environments and material properties. We train our system from synthesized training data, but demonstrate its applicability to real-world data. Interestingly, our analysis shows that the information obtained from objects made out of multiple materials often is complementary and leads to better performance.

What is the right way to reason about human activities? What directions forward are most promising? In this work, we analyze the current state of human activity understanding in videos. The goal of this paper is to examine datasets, evaluation metrics, algorithms, and potential future directions. We look at the qualitative attributes that define activities such as pose variability, brevity, and density. The experiments consider multiple state-of-the-art algorithms and multiple datasets. The results demonstrate that while there is inherent ambiguity in the temporal extent of activities, current datasets still permit effective benchmarking. We discover that fine-grained understanding of objects and pose when combined with temporal reasoning is likely to yield substantial improvements in algorithmic accuracy. We present the many kinds of information that will be needed to achieve substantial gains in activity understanding: objects, verbs, intent, and sequential reasoning. The software and additional information will be made available to provide other researchers detailed diagnostics to understand their own algorithms.

Most of the prior works summarize videos by either exploring different heuristically designed criteria in an unsupervised way or developing fully supervised algorithms by leveraging human-crafted training data in form of video-summary pairs or importance annotations. However, unsupervised methods are blind to the video category and often fail to produce semantically meaningful video summaries. On the other hand, acquisition of large amount of training data in supervised approaches is non-trivial and may lead to a biased model. Different from existing methods, we introduce a weakly supervised approach that requires only video-level annotation for summarizing web videos. Casting the problem as a weakly supervised learning problem, we propose a flexible deep 3D CNN architecture to learn the notion of importance using only video-level annotation, and without any human-crafted training data. Specifically, our main idea is to leverage multiple videos of a category to automatically learn a parametric model for categorizing videos and then adopt the model to find important segments from a given video as the ones which have maximum influence to the model output. Furthermore, to unleash the full potential of our 3D CNN architecture, we also explored a series of good practices to reduce the influence of limited training data while summarizing videos. Experiments on two challenging and diverse datasets well demonstrate that our approach produces superior quality video summaries compared to several recently proposed approaches.

We propose to help weakly supervised object localization for classes where location annotations are not available, by transferring things and stuff knowledge from a source set with available annotations. The source and target classes might share similar appearance (e.g. bear fur is similar to cat fur) or appear against similar background (e.g. horse and sheep appear against grass). To exploit this, we acquire three types of knowledge from the source set: a segmentation model trained on both thing and stuff classes; similarity relations between target and source classes; and co-occurrence relations between thing and stuff classes in the source. The segmentation model is used to generate thing and stuff segmentation maps on a target image, while the class similarity and co-occurrence knowledge help refining them. We then incorporate these maps as new cues into a multiple instance learning framework (MIL), propagating the transferred knowledge from the pixel level to the object proposal level. In extensive experiments, we conduct our transfer from the PASCAL Context dataset (source) to the ILSVRC, COCO and PASCAL VOC 2007 datasets (targets). We evaluate our transfer across widely different thing classes, including some that are not similar in appearance, but appear against similar background. The results demonstrate significant improvement over standard MIL, and we outperform the state-of-the-art in the transfer setting.

The goal of the semantic object correspondence problem is to compute dense association maps for a pair of images such that the same object parts get matched for very different appearing object instances. Our method builds on the recent findings that deep convolutional neural networks (DCNNs) implicitly learn a latent model of object parts even when trained for classification. We also leverage a key correspondence problem insight that the geometric structure between object parts is consistent across multiple object instances. These two concepts are then combined in the form of a novel optimization scheme. This optimization learns a feature embedding by rewarding for projecting features closer on the manifold if they have low feature-space distance. Simultaneously, the optimization penalizes feature clusters whose geometric structure is inconsistent with the observed geometric structure of object parts. In this manner, by accounting for feature space similarities and feature neighborhood context together, a manifold is learned where features belonging to semantically similar object parts cluster together. We also describe transferring these embedded features to the sister tasks of semantic keypoint classification and localization task via a Siamese DCNN. We provide qualitative results on the Pascal VOC 2012 images and quantitative results on the Pascal Berkeley dataset where we improve on the state of the art by over 5% on classification and over 9% on localization tasks.

Deep-learning metrics have recently demonstrated extremely good performance to match image patches for stereo reconstruction. However, training such metrics requires large amount of labeled stereo images, which can be difficult or costly to collect for certain applications (consider for example satellite stereo imaging). Moreover, labels from the depth sensors are often noisy. The main contribution of our work is a new weakly-supervised method for learning deep metrics from unlabeled stereo images, given coarse information about the scenes and the optical system. Our method alternatively optimizes the metric with a standard stochastic gradient descent, and applies stereo constraints to regularize its prediction. Experiments on reference data-sets show that, for a given network architecture, training with this new method without ground-truth produces a metric with performance as good as state-of-the-art baselines trained with the said ground-truth. This work has three practical implications. Firstly, it helps to overcome limitations of training sets, in particular noisy ground truth. Secondly it allows to use much more training data during learning. Thirdly, it allows to tune deep metric for a particular stereo system, even if ground truth is not available.

The requiring of large amounts of annotated training data has become a common constraint on various deep learning systems. In this paper, we propose a weakly supervised scene text detection method (WeText) that trains robust and accurate scene text detection models by learning from unannotated or weakly annotated data. With a "light" supervised model trained on a small fully annotated dataset, we explore semi-supervised and weakly supervised learning on a large unannotated dataset and a large weakly annotated dataset, respectively. For the unsupervised learning, the light supervised model is applied to the unannotated dataset to search for more character training samples, which are further combined with the small annotated dataset to retrain a superior character detection model. For the weakly supervised learning, the character searching is guided by high-level annotations of words/text lines that are widely available and also much easier to prepare. In addition, we design an unified scene character detector by adapting regression based deep networks, which greatly relieves the error accumulation issue that widely exists in most traditional approaches. Extensive experiments across different unannotated and weakly annotated datasets show that the scene text detection performance can be clearly boosted under both scenarios, where the weakly supervised learning can achieve the state-of-the-art performance by using only 229 fully annotated scene text images.

Most modern face super-resolution methods resort to convolutional neural networks (CNN) to infer high-resolution (HR) face images. When dealing with very low resolution (LR) images, the performance of these CNN based methods greatly degrades. Meanwhile, these methods tend to produce over-smoothed outputs and miss some textural details. To address these challenges, this paper presents a wavelet-based CNN approach that can ultra-resolve a very low resolution face image of 16x16 or smaller pixel-size to its larger version of multiple scaling factors (2x, 4x, 8x and even 16x) in a unified framework. Different from conventional CNN methods directly inferring HR images, our approach firstly learns to predict the LR's corresponding series of HR's wavelet coefficients before reconstructing HR images from them. To capture both global topology information and local texture details of human faces, we present a flexible and extensible convolutional neural network with three types of loss: wavelet prediction loss, texture loss and full-image loss. Extensive experiments demonstrate that the proposed approach achieves more appealing results both quantitatively and qualitatively than state-of-the-art super-resolution methods.

In experimental fluid dynamics, the flow in a volume of fluid is observed by injecting high-contrast tracer particles and tracking them in multi-view video. Fluid dynamics researchers have developed variants of space-carving to reconstruct the 3D particle distribution at a given time-step, and then use relatively simple local matching to recover the motion over time. On the contrary, estimating the optical flow between two consecutive images is a long-standing standard problem in computer vision, but only little work exists about volumetric 3D flow. Here, we propose a variational method for 3D fluid flow estimation from multi-view data. We start from a 3D version of the standard variational flow model, and investigate different regularization schemes that ensure divergence-free flow fields, to account for the physics of incompressible fluids. Moreover, we propose a semi-dense formulation, to cope with the computational demands of large volumetric datasets. Flow is estimated and regularized at a lower spatial resolution, while the data term is evaluated at full resolution to preserve the discriminative power and geometric precision of the local particle distribution. Extensive experiments reveal that a simple sum of squared differences (SSD) is the most suitable data term for our application. For regularization, an energy whose Euler-Lagrange equations correspond to the stationary Stokes equations leads to the best results. This strictly enforces a divergence-free flow and additionally penalizes the squared gradient of the flow.

Automatic kinship verification from facial information is a relatively new and open research problem in computer vision. This paper explores the possibility of learning an efficient facial representation for video-based kinship verification by exploiting the visual transformation between facial appearance of kin pairs. To this end, a Siamese-like coupled convolutional encoder-decoder network is proposed. To reveal resemblance patterns of kinship while discarding the similarity patterns that can also be observed between people who do not have a kin relationship, a novel contrastive loss function is defined in the visual appearance space. For further optimization, the learned representation is fine-tuned using a feature-based contrastive loss. An expression matching procedure is employed in the model to minimize the negative influence of expression differences between kin pairs. Each kin video is analyzed by a sliding temporal window to leverage short-term facial dynamics. The effectiveness of the proposed method is assessed on seven different kin relationships using smile videos of kin pairs. On the average, 93.65% verification accuracy is achieved, improving the state of the art.

A crucial capability of real-world intelligent agents is their ability to plan a sequence of actions to achieve their goals in the visual world. In this work, we address the problem of visual semantic planning: the task of predicting a sequence of actions from visual observations that transform a dynamic environment from an initial state to a goal state. Doing so entails knowledge about objects and their affordances, as well as actions and their preconditions and effects. We propose learning these through interacting with a visual and dynamic environment. Our proposed solution involves bootstrapping reinforcement learning with imitation learning. To ensure cross task generalization, we develop a deep predictive model based on successor representations. Our experimental results show near optimal results across a wide range of tasks in the challenging THOR environment. The supplementary video can be accessed at the following link: https://goo.gl/vXsbQP.

Understanding the visual relationship between two objects involves identifying the subject, the object, and a predicate relating them.We leverage the strong correlations between the predicate and the (subj,obj) pair (both semantically and spatially) to predict predicates conditioned on the subjects and the objects. Modeling the three entities jointly more accurately reflects their relationships compared to modeling them independently, but it complicates learning since the semantic space of visual relationships is huge and training data is limited, especially for long-tail relationships that have few instances. To overcome this, we use knowledge of linguistic statistics to regularize visual model learning. We obtain linguistic knowledge by mining from both training annotations (internal knowledge) and publicly available text, e.g., Wikipedia (external knowledge), computing the conditional probability distribution of a predicate given a (subj,obj) pair. As we train the visual model, we distill this knowledge into the deep model to achieve better generalization. Our experimental results on the Visual Relationship Detection (VRD) and Visual Genome datasets suggest that with this linguistic knowledge distillation, our model outperforms the state-of-the-art methods significantly, especially when predicting unseen relationships (e.g., recall improved from 8.45% to 19.17% on VRD zero-shot testing set).

Skeleton-based human action recognition has recently attracted increasing attention due to the popularity of 3D skeleton data. One main challenge lies in the large view variations in captured human actions. We propose a novel view adaptation scheme to automatically regulate observation viewpoints during the occurrence of an action. Rather than re-positioning the skeletons based on a human defined prior criterion, we design a view adaptive recurrent neural network (RNN) with LSTM architecture, which enables the network itself to adapt to the most suitable observation viewpoints from end to end. Extensive experiment analyses show that the proposed view adaptive RNN model strives to (1) transform the skeletons of various views to much more consistent viewpoints and (2) maintain the continuity of the action rather than transforming every frame to the same position with the same body orientation. Our model achieves significant improvement over the state-of-the-art approaches on three benchmark datasets.

Video scene parsing is challenging due to the following two reasons: firstly, it is non-trivial to learn meaningful video representations for producing the temporally consistent labeling map; secondly, such a learning process becomes more difficult with insufficient labeled video training data. In this work, we propose a unified framework to address the above two problems, which is to our knowledge the first model to employ predictive feature learning in the video scene parsing. The predictive feature learning is carried out in two predictive tasks: frame prediction and predictive parsing. It is experimentally proved that the learned predictive features in our model are able to significantly enhance the video parsing performance by combining with the standard image parsing network. Interestingly, the performance gain brought by the predictive learning is almost costless as the features are learned from a large amount of unlabeled video data in an unsupervised way. Extensive experiments over two challenging datasets, Cityscapes and Camvid, have demonstrated the effectiveness of our model by showing remarkable improvement over well-established baselines.

Reflections can obstruct content during video capture and hence their removal is desirable. Current removal techniques are designed for still images, extracting only one reflection (foreground) and one background layer from the input. When extended to videos, unpleasant artifacts such as temporal flickering and incomplete separation are generated. We present a technique for video reflection removal by jointly solving for motion and separation. The novelty of our work is in our optimization formulation as well as the motion initialization strategy. We present a novel spatio-temporal optimization that takes n frames as input and directly estimates 2n frames as output, n for each layer. We aim to fully utilize spatio-temporal information in our objective terms. Our motion initialization is based on iterative frame-to-frame alignment instead of the direct alignment used by current approaches. We compare against advanced video extensions of the state of the art, and we significantly reduce temporal flickering and improve separation. In addition, we reduce image blur and recover moving objects more accurately. We validate our approach through subjective and objective evaluations on real and controlled data.

Standard video frame interpolation methods first estimate optical flow between input frames and then synthesize an intermediate frame guided by motion. Recent approaches merge these two steps into a single convolution process by convolving input frames with spatially adaptive kernels that account for motion and re-sampling simultaneously. These methods require large kernels to handle large motion, which limits the number of pixels whose kernels can be estimated at once due to the large memory demand. To address this problem, this paper formulates frame interpolation as local separable convolution over input frames using pairs of 1D kernels. Compared to regular 2D kernels, the 1D kernels require significantly fewer parameters to be estimated. Our method develops a deep fully convolutional neural network that takes two input frames and estimates pairs of 1D kernels for all pixels simultaneously. Since our method is able to estimate kernels and synthesizes the whole video frame at once, it allows for the incorporation of perceptual loss to train the neural network to produce visually pleasing frames. This deep neural network is trained end-to-end using widely available video data without any human annotation. Both qualitative and quantitative experiments show that our method provides a practical solution to high-quality video frame interpolation.