CVPR 2017 论文列表

Deep learning methods achieve great success recently on many computer vision problems. In spite of these practical successes, optimization of deep networks remains an active topic in deep learning research. In this work, we focus on investigation of the network solution properties that can potentially lead to good performance. Our research is inspired by theoretical and empirical results that use orthogonal matrices to initialize networks, but we are interested in investigating how orthogonal weight matrices perform when network training converges. To this end, we propose to constrain the solutions of weight matrices in the orthogonal feasible set during the whole process of network training, and achieve this by a simple yet effective method called Singular Value Bounding (SVB). In SVB, all singular values of each weight matrix are simply bounded in a narrow band around the value of 1. Based on the same motivation, we also propose Bounded Batch Normalization (BBN), which improves Batch Normalization by removing its potential risk of ill-conditioned layer transform. We present both theoretical and empirical results to justify our proposed methods. Experiments on benchmark image classification datasets show the efficacy of our proposed SVB and BBN. In particular, we achieve the state-of-the-art results of 3.06% error rate on CIFAR10 and 16.90% on CIFAR100, using off-the-shelf network architectures (Wide ResNets). Our preliminary results on ImageNet also show the promise in large-scale learning. We release the implementation code of our methods at www.aperture-lab.net/research/svb.

In this work, we present a method for improving a random sample consensus (RANSAC) based image segmentation algorithm by encapsulating it within a convolutional neural network (CNN). The improvements are gained by gradient descent training on the set of pre-RANSAC filtering and thresholding operations using a novel RANSAC-based loss function, which is geared toward optimizing the strength of the correct model relative to the most convincing false model. Thus, it can be said that our loss function trains the network on metrics that directly dictate the success or failure of the final segmentation rather than metrics that are merely correlated to the success or failure. We demonstrate successful application of this method to a RANSAC method for identifying the pupil boundary in images from the CASIA-IrisV3 iris recognition data set, and we expect that this method could be successfully applied to any RANSAC-based segmentation algorithm.

Learning to rank has recently emerged as an attractive technique to train deep convolutional neural networks for various computer vision tasks. Pairwise ranking, in particular, has been successful in multi-label image classification, achieving state-of-the-art results on various benchmarks. However, most existing approaches use the hinge loss to train their models, which is non-smooth and thus is difficult to optimize especially with deep networks. Furthermore, they employ simple heuristics, such as top-k or thresholding, to determine which labels to include in the output from a ranked list of labels, which limits their use in the real-world setting. In this work, we propose two techniques to improve pairwise ranking based multi-label image classification by solving the aforementioned problems: (1) we propose a novel loss function for pairwise ranking, which is smooth everywhere; and (2) we incorporate a label decision module into the model, estimating the optimal confidence thresholds for each visual concept. We provide theoretical analyses of our loss function from the point of view of the Bayes consistency and risk minimization, and show its benefit over existing pairwise ranking formulations. We also demonstrate the effectiveness of our approach on two large-scale datasets, NUS-WIDE and MS-COCO, achieving the best reported result in the literature.

Interpretability of deep neural networks (DNNs) is essential since it enables users to understand the overall strengths and weaknesses of the models, conveys an understanding of how the models will behave in the future, and how to diagnose and correct potential problems. However, it is challenging to reason about what a DNN actually does due to its opaque or black-box nature. To address this issue, we propose a novel technique to improve the interpretability of DNNs by leveraging the rich semantic information embedded in human descriptions. By concentrating on the video captioning task, we first extract a set of semantically meaningful topics from the human descriptions that cover a wide range of visual concepts, and integrate them into the model with an interpretive loss. We then propose a prediction difference maximization algorithm to interpret the learned features of each neuron. Experimental results demonstrate its effectiveness in video captioning using the interpretable features, which can also be transferred to video action recognition. By clearly understanding the learned features, users can easily revise false predictions via a human-in-the-loop procedure.

Attributes are semantically meaningful characteristics whose applicability widely crosses category boundaries. They are particularly important in describing and recognizing concepts where no explicit training example is given, e.g., zero-shot learning. Additionally, since attributes are human describable, they can be used for efficient human-computer interaction. In this paper, we propose to employ semantic segmentation to improve facial attribute prediction. The core idea lies in the fact that many facial attributes describe local properties. In other words, the probability of an attribute to appear in a face image is far from being uniform in the spatial domain. We build our facial attribute prediction model jointly with a deep semantic segmentation network. This harnesses the localization cues learned by the semantic segmentation to guide the attention of the attribute prediction to the regions where different attributes naturally show up. As a result of this approach, in addition to recognition, we are able to localize the attributes, despite merely having access to image level labels (weak supervision) during training. We evaluate our proposed method on CelebA and LFWA datasets and achieve superior results to the prior arts. Furthermore, we show that in the reverse problem, semantic face parsing improves when facial attributes are available. That reaffirms the need to jointly model these two interconnected tasks.

The recent work of Gatys et al., who characterized the style of an image by the statistics of convolutional neural network filters, ignited a renewed interest in the texture generation and image stylization problems. While their image generation technique uses a slow optimization process, recently several authors have proposed to learn generator neural networks that can produce similar outputs in one quick forward pass. While generator networks are promising, they are still inferior in visual quality and diversity compared to generation-by-optimization. In this work, we advance them in two significant ways. First, we introduce an instance normalization module to replace batch normalization with significant improvements to the quality of image stylization. Second, we improve diversity by introducing a new learning formulation that encourages generators to sample unbiasedly from the Julesz texture ensemble, which is the equivalence class of all images characterized by certain filter responses. Together, these two improvements take feed forward texture synthesis and image stylization much closer to the quality of generation-via-optimization, while retaining the speed advantage.

We present an improved three-step pipeline for the stereo matching problem and introduce multiple novelties at each stage. We propose a new highway network architecture for computing the matching cost at each possible disparity, based on multilevel weighted residual shortcuts, trained with a hybrid loss that supports multilevel comparison of image patches. A novel post-processing step is then introduced, which employs a second deep convolutional neural network for pooling global information from multiple disparities. This network outputs both the image disparity map, which replaces the conventional "winner takes all" strategy, and a confidence in the prediction. The confidence score is achieved by training the network with a new technique that we call the reflective loss. Lastly, the learned confidence is employed in order to better detect outliers in the refinement step. The proposed pipeline achieves state of the art accuracy on the largest and most competitive stereo benchmarks, and the learned confidence is shown to outperform all existing alternatives.

We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. Moreover, since the release of the pix2pix software associated with this paper, hundreds of twitter users have posted their own artistic experiments using our system. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without handengineering our loss functions either.

Recently, Convolutional Neural Network (CNN) based models have achieved great success in Single Image Super-Resolution (SISR). Owing to the strength of deep networks, these CNN models learn an effective nonlinear mapping from the low-resolution input image to the high-resolution target image, at the cost of requiring enormous parameters. This paper proposes a very deep CNN model (up to 52 convolutional layers) named Deep Recursive Residual Network (DRRN) that strives for deep yet concise networks. Specifically, residual learning is adopted, both in global and local manners, to mitigate the difficulty of training very deep networks; recursive learning is used to control the model parameters while increasing the depth. Extensive benchmark evaluation shows that DRRN significantly outperforms state of the art in SISR, while utilizing far fewer parameters. Code is available at https://github.com/tyshiwo/DRRN_CVPR17.

Recent advances on image manipulation techniques have made image forgery detection increasingly more challenging. An important component in such tools is to fake motion and/or defocus blurs through boundary splicing and copy-move operators, to emulate wide aperture and slow shutter effects. In this paper, we present a new technique based on the analysis of the camera response functions (CRF) for efficient and robust splicing and copy-move forgery detection and localization. We first analyze how non-linear CRFs affect edges in terms of the intensity-gradient bivariable histograms. We show distinguishable shape differences on real vs. forged blurs near edges after a splicing operation. Based on our analysis, we introduce a deep-learning framework to detect and localize forged edges. In particular, we show the problem can be transformed to a handwriting recognition problem an resolved by using a convolutional neural network. We generate a large dataset of forged images produced by splicing followed by retouching and comprehensive experiments show our proposed method outperforms the state-of-the-art techniques in accuracy and robustness.

Camera motion introduces motion blur, affecting many computer vision tasks. Dark Channel Prior (DCP) helps the blind deblurring on scenes including natural, face, text, and low-illumination images. However, it has limitations and is less likely to support the kernel estimation while bright pixels dominate the input image. We observe that the bright pixels in the clear images are not likely to be bright after the blur process. Based on this observation, we first illustrate this phenomenon mathematically and define it as the Bright Channel Prior (BCP). Then, we propose a technique for deblurring such images which elevates the performance of existing motion deblurring algorithms. The proposed method takes advantage of both Bright and Dark Channel Prior. This joint prior is named as extreme channels prior and is crucial for achieving efficient restorations by leveraging both the bright and dark information. Extensive experimental results demonstrate that the proposed method is more robust and performs favorably against the state-of-the-art image deblurring methods on both synthesized and natural images.

We consider scenarios in which we wish to perform joint scene understanding, object tracking, activity recognition, and other tasks in scenarios in which multiple people are wearing body-worn cameras while a third-person static camera also captures the scene. To do this, we need to establish person-level correspondences across first- and third-person videos, which is challenging because the camera wearer is not visible from his/her own egocentric video, preventing the use of direct feature matching. In this paper, we propose a new semi-Siamese Convolutional Neural Network architecture to address this novel challenge. We formulate the problem as learning a joint embedding space for first- and third-person videos that considers both spatial- and motion-domain cues. A new triplet loss function is designed to minimize the distance between correct first- and third-person matches while maximizing the distance between incorrect ones. This end-to-end approach performs significantly better than several baselines, in part by learning the first- and third-person features optimized for matching jointly with the distance measure itself.

The effectiveness of current iris recognition systems depends on the accurate segmentation and parameterisation of the iris boundaries, as failures at this point misalign the coefficients of the biometric signatures. This paper describes IRINA, an algorithm for Iris Recognition that is robust against INAccurately segmented samples, which makes it a good candidate to work in poor-quality data. The process is based in the concept of "corresponding" patch between pairs of images, that is used to estimate the posterior probabilities that patches regard the same biological region, even in case of segmentation errors and non-linear texture deformations. Such information enables to infer a free-form deformation field (2D registration vectors) between images, whose first and second-order statistics provide effective biometric discriminating power. Extensive experiments were carried out in four datasets (CASIA-IrisV3-Lamp, CASIA-IrisV4-Lamp, CASIA-IrisV4-Thousand and WVU) and show that IRINA not only achieves state-of-the-art performance in good quality data, but also handles effectively severe segmentation errors and large differences in pupillary dilation / constriction.

Hyperspectral image(HSI)super-resolution, which fuses a low-resolution (LR) HSI with a high-resolution (HR) multispectral image (MSI), has recently attracted much attention. Most of the current HSI super-resolution approaches are based on matrix factorization, which unfolds the three-dimensional HSI as a matrix before processing. In general, the matrix data representation obtained after the matrix unfolding operation makes it hard to fully exploit the inherent HSI spatial-spectral structures. In this paper, a novel HSI super-resolution method based on non-local sparse tensor factorization (called as the NLSTF) is proposed. The sparse tensor factorization can directly decompose each cube of the HSI as a sparse core tensor and dictionaries of three modes, which reformulates the HSI super-resolution problem as the estimation of sparse core tensor and dictionaries for each cube. To further exploit the non-local spatial self-similarities of the HSI, similar cubes are grouped together, and they are assumed to share the same dictionaries. The dictionaries are learned from the LR-HSI and HR-MSI for each group, and corresponding sparse core tensors are estimated by spare coding on the learned dictionaries for each cube. Experimental results demonstrate the superiority of the proposed NLSTF approach over several state-of-the-art HSI super-resolution approaches.

Recent low-rank based matrix/tensor recovery methods have been widely explored in multispectral images (MSI) denoising. These methods, however, ignore the difference of the intrinsic structure correlation along spatial sparsity, spectral correlation and non-local self-similarity mode. In this paper, we go further by giving a detailed analysis about the rank properties both in matrix and tensor cases, and figure out the non-local self-similarity is the key ingredient, while the low-rank assumption of others may not hold. This motivates us to design a simple yet effective unidirectional low-rank tensor recovery model that is capable of truthfully capturing the intrinsic structure correlation with reduced computational burden. However, the low-rank models suffer from the ringing artifacts, due to the aggregation of overlapped patches/cubics. While previous methods resort to spatial information, we offer a new perspective by utilizing the exclusively spectral information in MSIs to address the issue. The analysis-based hyper-Laplacian prior is introduced to model the global spectral structures, so as to indirectly alleviate the ringing artifacts in spatial domain. The advantages of the proposed method over the existing ones are multi-fold: more reasonably structure correlation representability, less processing time, and less artifacts in the overlapped regions. The proposed method is extensively evaluated on several benchmarks, and significantly outperforms state-of-the-art MSI denoising methods.

Recent advances in deep learning have shown exciting promise in filling large holes in natural images with semantically plausible and context aware details, impacting fundamental image manipulation tasks such as object removal. While these learning-based methods are significantly more effective in capturing high-level features than prior techniques, they can only handle very low-resolution inputs due to memory limitations and difficulty in training. Even for slightly larger images, the inpainted regions would appear blurry and unpleasant boundaries become visible. We propose a multi-scale neural patch synthesis approach based on joint optimization of image content and texture constraints, which not only preserves contextual structures but also produces high-frequency details by matching and adapting patches with the most similar mid-layer feature correlations of a deep classification network. We evaluate our method on the ImageNet and Paris Streetview datasets and achieved state-of-the-art inpainting accuracy. We show our approach produces sharper and more coherent results than prior methods, especially for high-resolution images.

Multimodal classification arises in many computer vision tasks such as object classification and image retrieval. The idea is to utilize multiple sources (modalities) measuring the same instance to improve the overall performance compared to using a single source (modality). The varying characteristics exhibited by multiple modalities make it necessary to simultaneously learn the corresponding metrics. In this paper, we propose a multiple metrics learning algorithm for multimodal data. Metric of each modality is a product of two matrices: one matrix is modality specific, the other is enforced to be shared by all the modalities. The learned metrics can improve multimodal classification accuracy and experimental results on four datasets show that the proposed algorithm outperforms existing learning algorithms based on multiple metrics as well as other approaches tested on these datasets. Specifically, we report 95.0% object instance recognition accuracy, 89.2% object category recognition accuracy on the multi-view RGB-D dataset and 52.3% scene category recognition accuracy on SUN RGB-D dataset.

The use of Recurrent Neural Networks for video captioning has recently gained a lot of attention, since they can be used both to encode the input video and to generate the corresponding description. In this paper, we present a recurrent video encoding scheme which can discover and leverage the hierarchical structure of the video. Unlike the classical encoder-decoder approach, in which a video is encoded continuously by a recurrent layer, we propose a novel LSTM cell which can identify discontinuity points between frames or segments and modify the temporal connections of the encoding layer accordingly. We evaluate our approach on three large-scale datasets: the Montreal Video Annotation dataset, the MPII Movie Description dataset and the Microsoft Video Description Corpus. Experiments show that our approach can discover appropriate hierarchical representations of input videos and improve the state of the art results on movie description datasets.

Translating or rotating an input image should not affect the results of many computer vision tasks. Convolutional neural networks (CNNs) are already translation equivariant: input image translations produce proportionate feature map translations. This is not the case for rotations. Global rotation equivariance is typically sought through data augmentation, but patch-wise equivariance is more difficult. We present Harmonic Networks or H-Nets, a CNN exhibiting equivariance to patch-wise translation and 360-rotation. We achieve this by replacing regular CNN filters with circular harmonics, returning a maximal response and orientation for every receptive field patch. H-Nets use a rich, parameter-efficient and fixed computational complexity representation, and we show that deep feature maps within the network encode complicated rotational invariants. We demonstrate that our layers are general enough to be used in conjunction with the latest architectures and techniques, such as deep supervision and batch normalization. We also achieve state-of-the-art classification on rotated-MNIST, and competitive results on other benchmark challenges.