CVPR 2019 论文列表

Many DNN-enabled vision applications constantly operate under severe energy constraints such as unmanned aerial vehicles, Augmented Reality headsets, and smartphones. Designing DNNs that can meet a stringent energy budget is becoming increasingly important. This paper proposes ECC, a framework that compresses DNNs to meet a given energy constraint while minimizing accuracy loss. The key idea of ECC is to model the DNN energy consumption via a novel bilinear regression function. The energy estimate model allows us to formulate DNN compression as a constrained optimization that minimizes the DNN loss function over the energy constraint. The optimization problem, however, has nontrivial constraints. Therefore, existing deep learning solvers do not apply directly. We propose an optimization algorithm that combines the essence of the Alternating Direction Method of Multipliers (ADMM) framework with gradient-based learning algorithms. The algorithm decomposes the original constrained optimization into several subproblems that are solved iteratively and efficiently. ECC is also portable across different hardware platforms without requiring hardware knowledge. Experiments show that ECC achieves higher accuracy under the same or lower energy budget compared to state-of-the-art resource-constrained DNN compression techniques.

A prerequisite to successfully alleviate pain in animals is to recognize it, which is a great challenge in non-verbal species. Furthermore, prey animals such as horses tend to hide their pain. In this study, we propose a deep recurrent two-stream architecture for the task of distinguishing pain from non-pain in videos of horses. Different models are evaluated on a unique dataset showing horses under controlled trials with moderate pain induction, which has been presented in earlier work. Sequential models are experimentally compared to single-frame models, showing the importance of the temporal dimension of the data, and are benchmarked against a veterinary expert classification of the data. We additionally perform baseline comparisons with generalized versions of state-of-the-art human pain recognition methods. While equine pain detection in machine learning is a novel field, our results surpass veterinary expert performance and outperform pain detection results reported for other larger non-human species.

Dynamic Scene deblurring is a challenging low-level vision task where spatially variant blur is caused by many factors, e.g., camera shake and object motion. Recent study has made significant progress. Compared with the parameter independence scheme [19] and parameter sharing scheme [33], we develop the general principle for constraining the deblurring network structure by proposing the generic and effective selective sharing scheme. Inside the subnetwork of each scale, we propose a nested skip connection structure for the nonlinear transformation modules to replace stacked convolution layers or residual blocks. Besides, we build a new large dataset of blurred/sharp image pairs towards better restoration quality. Comprehensive experimental results show that our parameter selective sharing scheme, nested skip connection structure, and the new dataset are all significant to set a new state-of-the-art in dynamic scene deblurring.

This paper proposes the dynamic recursive neural network (DRNN), which simplifies the duplicated building blocks in deep neural network. Different from forwarding through different blocks sequentially in previous networks, we demonstrate that the DRNN can achieve better performance with fewer blocks by employing block recursively. We further add a gate structure to each block, which can adaptively decide the loop times of recursive blocks to reduce the computational cost. Since the recursive networks are hard to train, we propose the Loopy Variable Batch Normalization (LVBN) to stabilize the volatile gradient. Further, we improve the LVBN to correct statistical bias caused by the gate structure. Experiments show that the DRNN reduces the parameters and computational cost and while outperforms the original model in term of the accuracy consistently on CIFAR-10 and ImageNet-1k. Lastly we visualize and discuss the relation between image saliency and the number of loop time.

Learning effective fusion of multi-modality features is at the heart of visual question answering. We propose a novel method of dynamically fuse multi-modal features with intra- and inter-modality information flow, which alternatively pass dynamic information between and across the visual and language modalities. It can robustly capture the high-level interactions between language and vision domains, thus significantly improves the performance of visual question answering. We also show that, the proposed dynamic intra modality attention flow conditioned on the other modality can dynamically modulate the intra-modality attention of the current modality, which is vital for multimodality feature fusion. Experimental evaluations on the VQA 2.0 dataset show that the proposed method achieves the state-of-the-art VQA performance. Extensive ablation studies are carried out for the comprehensive analysis of the proposed method.

In this paper, we present a novel approach for dynamic text effects transfer by using example-based texture synthesis. In contrast to previous works that require an input video of the target to provide motion guidance, we aim to animate a still image of the target text by transferring the desired dynamic effects from an observed exemplar. Due to the simplicity of target guidance and complexity of realistic effects, it is prone to producing temporal artifacts such as flickers and pulsations. To address the problem, our core idea is to find a common Nearest-neighbor Field (NNF) that would optimize the textural coherence across all keyframes simultaneously. With the static NNF for video sequences, we implicitly transfer motion properties from source to target. We also introduce a guided NNF search by employing the distance-based weight map and Simulated Annealing (SA) for deep direction-guided propagation to allow intense dynamic effects to be completely transferred with no semantic guidance provided. Experimental results demonstrate the effectiveness and superiority of our method in dynamic text effects transfer through extensive comparisons with state-of-the-art algorithms. We also show the potentiality of our method via multiple experiments for various application domains.

In this paper, we study design of deep neural networks for tasks of image restoration. We propose a novel style of residual connections dubbed "dual residual connection", which exploits the potential of paired operations, e.g., up- and down-sampling or convolution with large- and small-size kernels. We design a modular block implementing this connection style; it is equipped with two containers to which arbitrary paired operations are inserted. Adopting the "unraveled" view of the residual networks proposed by Veit et al., we point out that a stack of the proposed modular blocks allows the first operation in a block interact with the second operation in any subsequent blocks. Specifying the two operations in each of the stacked blocks, we build a complete network for each individual task of image restoration. We experimentally evaluate the proposed approach on five image restoration tasks using nine datasets. The results show that the proposed networks with properly chosen paired operations outperform previous methods on almost all of the tasks and datasets.

This paper attacks the challenging problem of zero-example video retrieval. In such a retrieval paradigm, an end user searches for unlabeled videos by ad-hoc queries described in natural language text with no visual example provided. Given videos as sequences of frames and queries as sequences of words, an effective sequence-to-sequence cross-modal matching is required. The majority of existing methods are concept based, extracting relevant concepts from queries and videos and accordingly establishing associations between the two modalities. In contrast, this paper takes a concept-free approach, proposing a dual deep encoding network that encodes videos and queries into powerful dense representations of their own. Dual encoding is conceptually simple, practically effective and end-to-end. As experiments on three benchmarks, i.e. MSR-VTT, TRECVID 2016 and 2017 Ad-hoc Video Search show, the proposed solution establishes a new state-of-the-art for zero-example video retrieval.

In this paper, we address the scene segmentation task by capturing rich contextual dependencies based on the self-attention mechanism. Unlike previous works that capture contexts by multi-scale features fusion, we propose a Dual Attention Networks (DANet) to adaptively integrate local features with their global dependencies. Specifically, we append two types of attention modules on top of traditional dilated FCN, which model the semantic interdependencies in spatial and channel dimensions respectively. The position attention module selectively aggregates the features at each position by a weighted sum of the features at all positions. Similar features would be related to each other regardless of their distances. Meanwhile, the channel attention module selectively emphasizes interdependent channel maps by integrating associated features among all channel maps. We sum the outputs of the two attention modules to further improve feature representation which contributes to more precise segmentation results. We achieve new state-of-the-art segmentation performance on three challenging scene segmentation datasets, i.e., Cityscapes, PASCAL Context and COCO Stuff dataset. In particular, a Mean IoU score of 81.5% on Cityscapes test set is achieved without using coarse data.

We present a deep learning framework, called DuLa-Net, to predict Manhattan-world 3D room layouts from a single RGB panorama. To achieve better prediction accuracy, our method leverages two projections of the panorama at once, namely the equirectangular panorama-view and the perspective ceiling-view, that each contains different clues about the room layouts. Our network architecture consists of two encoder-decoder branches for analyzing each of the two views. In addition, a novel feature fusion structure is proposed to connect the two branches, which are then jointly trained to predict the 2D floor plans and layout heights. To learn more complex room layouts, we introduce the Realtor360 dataset that contains panoramas of Manhattan-world room layouts with different numbers of corners. Experimental results show that our work outperforms recent state-of-the-art in prediction accuracy and performance, especially in the rooms with non-cuboid layouts.

Computed tomography (CT) is an imaging modality widely used for medical diagnosis and treatment. CT images are often corrupted by undesirable artifacts when metallic implants are carried by patients, which creates the problem of metal artifact reduction (MAR). Existing methods for reducing the artifacts due to metallic implants are inadequate for two main reasons. First, metal artifacts are structured and non-local so that simple image domain enhancement approaches would not suffice. Second, the MAR approaches which attempt to reduce metal artifacts in the X-ray projection (sinogram) domain inevitably lead to severe secondary artifact due to sinogram inconsistency. To overcome these difficulties, we propose an end-to-end trainable Dual Domain Network (DuDoNet) to simultaneously restore sinogram consistency and enhance CT images. The linkage between the sigogram and image domains is a novel Radon inversion layer that allows the gradients to back-propagate from the image domain to the sinogram domain during training. Extensive experiments show that our method achieves significant improvements over other single domain MAR approaches. To the best of our knowledge, it is the first end-to-end dual-domain network for MAR.

Great progress has been made on estimating disparity maps from stereo images. However, with the limited stereo data available in the existing datasets and unstable ranging precision of current stereo methods, industry-level stereo matching in autonomous driving remains challenging. In this paper, we construct a novel large-scale stereo dataset named DrivingStereo. It contains over 180k images covering a diverse set of driving scenarios, which is hundreds of times larger than the KITTI Stereo dataset. High-quality labels of disparity are produced by a model-guided filtering strategy from multi-frame LiDAR points. For better evaluations, we present two new metrics for stereo matching in the driving scenes, i.e. a distance-aware metric and a semantic-aware metric. Extensive experiments show that compared with the models trained on FlyingThings3D or Cityscapes, the models trained on our DrivingStereo achieve higher generalization accuracy in real-world driving scenes, while the proposed metrics better evaluate the stereo methods on all-range distances and across different classes. Our dataset and code are available at https://drivingstereo-dataset.github.io.

Blind deconvolution problems are heavily ill-posed where the specific blurring kernel is not known. Recovering these images typically requires estimates of the kernel. In this paper, we present a method called Dr-Net, which does not require any such estimate and is further able to invert the effects of the blurring in blind image recovery tasks. These image recovery problems typically have two terms, the data fidelity term (for faithful reconstruction) and the image prior (for realistic looking reconstructions). We use the Douglas-Rachford iterations to solve this problem since it is a more generally applicable optimization procedure than methods such as the proximal gradient descent algorithm. Two proximal operators originate from these iterations, one from the data fidelity term and the second from the image prior. It is non-trivial to design a hand-crafted function to represent these proximal operators for the data fidelity and the image prior terms which would work with real-world image distributions. We therefore approximate both these proximal operators using deep networks. This provides a sound motivation for the final architecture for Dr-Net which we find outperforms the state-of-the-art on two mainstream blind deconvolution benchmarks. We also find that Dr-Net is one of the fastest algorithms according to wall-clock times while doing so.

Unsupervised clustering for high-dimension data (such as imageset or video) is a hard issue in data processing and data mining area since these data always lie on a manifold (such as Grassmann manifold). Inspired of Low Rank representation theory, researchers proposed a series of effective clustering methods for high-dimension data with non-linear metric. However, most of these methods adopt the traditional single nuclear norm as the relaxation of the rank function, which would lead to suboptimal solution deviated from the original one. In this paper, we propose a new low rank model for high-dimension data clustering task on Grassmann manifold based on the Double Nuclear norm which is used to better approximate the rank minimization of matrix. Further, to consider the inner geometry or structure of data space, we integrated the adaptive Laplacian regularization to construct the local relationship of data samples. The proposed models have been assessed on several public datasets for imageset clustering. The experimental results show that the proposed models outperform the state-of-the-art clustering ones.

In this paper, we investigate the problem of zero-shot sketch-based image retrieval (ZS-SBIR), where human sketches are used as queries to conduct retrieval of photos from unseen categories. We importantly advance prior arts by proposing a novel ZS-SBIR scenario that represents a firm step forward in its practical application. The new setting uniquely recognizes two important yet often neglected challenges of practical ZS-SBIR, (i) the large domain gap between amateur sketch and photo, and (ii) the necessity for moving towards large-scale retrieval. We first contribute to the community a novel ZS-SBIR dataset, QuickDraw-Extended, that consists of 330,000 sketches and 204,000 photos spanning across 110 categories. Highly abstract amateur human sketches are purposefully sourced to maximize the domain gap, instead of ones included in existing datasets that can often be semi-photorealistic. We then formulate a ZS-SBIR framework to jointly model sketches and photos into a common embedding space. A novel strategy to mine the mutual information among domains is specifically engineered to alleviate the domain gap. External semantic knowledge is further embedded to aid semantic transfer. We show that, rather surprisingly, retrieval performance significantly outperforms that of state-of-the-art on existing datasets that can already be achieved using a reduced version of our model. We further demonstrate the superior performance of our full model by comparing with a number of alternatives on the newly proposed dataset. The new dataset, plus all training and testing code of our model, will be publicly released to facilitate future research.

Unsupervised domain adaptation aims to learn a model of classifier for unlabeled samples on the target domain, given training data of labeled samples on the source domain. Impressive progress is made recently by learning invariant features via domain-adversarial training of deep networks. In spite of the recent progress, domain adaptation is still limited in achieving the invariance of feature distributions at a finer category level. To this end, we propose in this paper a new domain adaptation method called Domain-Symmetric Networks (SymNets). The proposed SymNet is based on a symmetric design of source and target task classifiers, based on which we also construct an additional classifier that shares with them its layer neurons. To train the SymNet, we propose a novel adversarial learning objective whose key design is based on a two-level domain confusion scheme, where the category-level confusion loss improves over the domain-level one by driving the learning of intermediate network features to be invariant at the corresponding categories of the two domains. Both domain discrimination and domain confusion are implemented based on the constructed additional classifier. Since target samples are unlabeled, we also propose a scheme of cross-domain training to help learn the target classifier. Careful ablation studies show the efficacy of our proposed method. In particular, based on commonly used base networks, our SymNets achieve the new state of the art on three benchmark domain adaptation datasets.

We propose a novel unsupervised domain adaptation framework based on domain-specific batch normalization in deep neural networks. We aim to adapt to both domains by specializing batch normalization layers in convolutional neural networks while allowing them to share all other model parameters, which is realized by a two-stage algorithm. In the first stage, we estimate pseudo-labels for the examples in the target domain using an external unsupervised domain adaptation algorithm---for example, MSTN or CPUA---integrating the proposed domain-specific batch normalization. The second stage learns the final models using a multi-task classification loss for the source and target domains. Note that the two domains have separate batch normalization layers in both stages. Our framework can be easily incorporated into the domain adaptation techniques based on deep neural networks with batch normalization layers. We also present that our approach can be extended to the problem with multiple source domains. The proposed algorithm is evaluated on multiple benchmark datasets and achieves the state-of-the-art accuracy in the standard setting and the multi-source domain adaption scenario.

Human adaptability relies crucially on the ability to learn and merge knowledge both from supervised and unsupervised learning: the parents point out few important concepts, but then the children fill in the gaps on their own. This is particularly effective, because supervised learning can never be exhaustive and thus learning autonomously allows to discover invariances and regularities that help to generalize. In this paper we propose to apply a similar approach to the task of object recognition across domains: our model learns the semantic labels in a supervised fashion, and broadens its understanding of the data by learning from self-supervised signals how to solve a jigsaw puzzle on the same images. This secondary task helps the network to learn the concepts of spatial correlation while acting as a regularizer for the classification task. Multiple experiments on the PACS, VLCS, Office-Home and digits datasets confirm our intuition and show that this simple method outperforms previous domain generalization and adaptation solutions. An ablation study further illustrates the inner workings of our approach.

Human pose estimation (HPE) is inherently a homogeneous multi-task learning problem, with the localization of each body part as a different task. Recent HPE approaches universally learn a shared representation for all parts, from which their locations are linearly regressed. However, our statistical analysis indicates not all parts are related to each other. As a result, such a sharing mechanism can lead to negative transfer and deteriorate the performance. This potential issue drives us to raise an interesting question. Can we identify related parts and learn specific features for them to improve pose estimation? Since unrelated tasks no longer share a high-level representation, we expect to avoid the adverse effect of negative transfer. In addition, more explicit structural knowledge, e.g., ankles and knees are highly related, is incorporated into the model, which helps resolve ambiguities in HPE. To answer this question, we first propose a data-driven approach to group related parts based on how much information they share. Then a part-based branching network (PBN) is introduced to learn representations specific to each part group. We further present a multi-stage version of this network to repeatedly refine intermediate features and pose estimates. Ablation experiments indicate learning specific features significantly improves the localization of occluded parts and thus benefits HPE. Our approach also outperforms all state-of-the-art methods on two benchmark datasets, with an outstanding advantage when occlusion occurs.