CVPR 2019 论文列表

Weakly supervised object detection (WSOD) is a challenging task when provided with image category supervision but required to simultaneously learn object locations and object detectors. Many WSOD approaches adopt multiple instance learning (MIL) and have non-convex loss functions which are prone to get stuck into local minima (falsely localize object parts) while missing full object extent during training. In this paper, we introduce a continuation optimization method into MIL and thereby creating continuation multiple instance learning (C-MIL), with the intention of alleviating the non-convexity problem in a systematic way. We partition instances into spatially related and class related subsets, and approximate the original loss function with a series of smoothed loss functions defined within the subsets. Optimizing smoothed loss functions prevents the training procedure falling prematurely into local minima and facilitates the discovery of Stable Semantic Extremal Regions (SSERs) which indicate full object extent. On the PASCAL VOC 2007 and 2012 datasets, C-MIL improves the state-of-the-art of weakly supervised object detection and weakly supervised object localization with large margins.

Semi-supervised video object segmentation has made significant progress on real and challenging videos in recent years. The current paradigm for segmentation methods and benchmark datasets is to segment objects in video provided a single annotation in the first frame. However, we find that segmentation performance across the entire video varies dramatically when selecting an alternative frame for annotation. This paper addresses the problem of learning to suggest the single best frame across the video for user annotation-this is, in fact, never the first frame of video. We achieve this by introducing BubbleNets, a novel deep sorting network that learns to select frames using a performance-based loss function that enables the conversion of expansive amounts of training examples from already existing datasets. Using BubbleNets, we are able to achieve an 11% relative improvement in segmentation performance on the DAVIS benchmark without any changes to the underlying method of segmentation.

Event-based cameras can measure intensity changes (called 'events') with microsecond accuracy under high-speed motion and challenging lighting conditions. With the active pixel sensor (APS), the event camera allows simultaneous output of the intensity frames. However, the output images are captured at a relatively low frame-rate and often suffer from motion blur. A blurry image can be regarded as the integral of a sequence of latent images, while the events indicate the changes between the latent images. Therefore, we are able to model the blur-generation process by associating event data to a latent image. In this paper, we propose a simple and effective approach, the Event-based Double Integral (EDI) model, to reconstruct a high frame-rate, sharp video from a single blurry frame and its event data. The video generation is based on solving a simple non-convex optimization problem in a single scalar variable. Experimental results on both synthetic and real images demonstrate the superiority of our EDI model and optimization method in comparison to the state-of-the-art.

We present a solution for the goal of extracting a video from a single motion blurred image to sequentially reconstruct the clear views of a scene as beheld by the camera during the time of exposure. We first learn motion representation from sharp videos in an unsupervised manner through training of a convolutional recurrent video autoencoder network that performs a surrogate task of video reconstruction. Once trained, it is employed for guided training of a motion encoder for blurred images. This network extracts embedded motion information from the blurred image to generate a sharp video in conjunction with the trained recurrent video decoder. As an intermediate step, we also design an efficient architecture that enables real-time single image deblurring and outperforms competing methods across all factors: accuracy, speed, and compactness. Experiments on real scenes and standard datasets demonstrate the superiority of our framework over the state-of-the-art and its ability to generate a plausible sequence of temporally consistent sharp frames.

Dense 3D face correspondence is a fundamental and challenging issue in the literature of 3D face analysis. Correspondence between two 3D faces can be viewed as a non-rigid registration problem that one deforms into the other, which is commonly guided by a few facial landmarks in many existing works. However, the current works seldom consider the problem of incoherent deformation caused by landmarks. In this paper, we explicitly formulate the deformation as locally rigid motions guided by some seed points, and the formulated deformation satisfies coherent local motions everywhere on a face. The seed points are initialized by a few landmarks, and are then augmented to boost shape matching between the template and the target face step by step, to finally achieve dense correspondence. In each step, we employ a hierarchical scheme for local shape registration, together with a Gaussian reweighting strategy for accurate matching of local features around the seed points. In our experiments, we evaluate the proposed method extensively on several datasets, including two publicly available ones: FRGC v2.0 and BU-3DFE. The experimental results demonstrate that our method can achieve accurate feature correspondence, coherent local shape motion, and compact data representation. These merits actually settle some important issues for practical applications, such as expressions, noise, and partial data.

Many images shared over the web include overlaid objects, or visual motifs, such as text, symbols or drawings, which add a description or decoration to the image. For example, decorative text that specifies where the image was taken, repeatedly appears across a variety of different images. Often, the reoccurring visual motif, is semantically similar, yet, differs in location, style and content (e.g. text placement, font and letters). This work proposes a deep learning based technique for blind removal of such objects. In the blind setting, the location and exact geometry of the motif are unknown. Our approach simultaneously estimates which pixels contain the visual motif, and synthesizes the underlying latent image. It is applied to a single input image, without any user assistance in specifying the location of the motif, achieving state-of-the-art results for blind removal of both opaque and semi-transparent visual motifs.

(Unsupervised) Domain Adaptation (DA) seeks for classifying target instances when solely provided with source labeled and target unlabeled examples for training. Learning domain-invariant features helps to achieve this goal, whereas it underpins unlabeled samples drawn from a single or multiple explicit target domains (Multi-target DA). In this paper, we consider a more realistic transfer scenario: our target domain is comprised of multiple sub-targets implicitly blended with each other so that learners could not identify which sub-target each unlabeled sample belongs to. This Blending-target Domain Adaptation (BTDA) scenario commonly appears in practice and threatens the validities of existing DA algorithms, due to the presence of domain gaps and categorical misalignments among these hidden sub-targets. To reap the transfer performance gains in this new scenario, we propose Adversarial Meta-Adaptation Network (AMEAN). AMEAN entails two adversarial transfer learning processes. The first is a conventional adversarial transfer to bridge our source and mixed target domains. To circumvent the intra-target category misalignment, the second process presents as "learning to adapt": It deploys an unsupervised meta-learner receiving target data and their ongoing feature-learning feedbacks, to discover target clusters as our "meta-sub-target" domains. This meta-sub-targets auto-design our meta-sub-target adaptation loss, which is capable to progressively eliminate the implicit category mismatching in our mixed target. We evaluate AMEAN and a variety of DA algorithms in three benchmarks under the BTDA setup. Empirical results show that BTDA is a quite challenging transfer setup for most existing DA algorithms, yet AMEAN significantly outperforms these state-of-the-art baselines and effectively restrains the negative transfer effects in BTDA.

Most previous learning-based visual odometry (VO) methods take VO as a pure tracking problem. In contrast, we present a VO framework by incorporating two additional components called Memory and Refining. The Memory component preserves global information by employing an adaptive and efficient selection strategy. The Refining component ameliorates previous results with the contexts stored in the Memory by adopting a spatial-temporal attention mechanism for feature distilling. Experiments on the KITTI and TUM-RGBD benchmark datasets demonstrate that our method outperforms state-of-the-art learning-based methods by a large margin and produces competitive results against classic monocular VO approaches. Especially, our model achieves outstanding performance in challenging scenarios such as texture-less regions and abrupt motions, where classic VO algorithms tend to fail.

Defenses against adversarial examples, when using the ImageNet dataset, are historically easy to defeat. The common understanding is that a combination of simple image transformations and other various defenses are insufficient to provide the necessary protection when the obfuscated gradient is taken into account. In this paper, we explore the idea of stochastically combining a large number of individually weak defenses into a single barrage of randomized transformations to build a strong defense against adversarial attacks. We show that, even after accounting for obfuscated gradients, the Barrage of Random Transforms (BaRT) is a resilient defense against even the most difficult attacks, such as PGD. BaRT achieves up to a 24x improvement in accuracy compared to previous work, and has even extended effectiveness out to a previously untested maximum adversarial perturbation of e=32.

Clustering is an important problem in various machine learning applications, but still a challenging task when dealing with complex real data. The existing clustering algorithms utilize either shallow models with insufficient capacity for capturing the non-linear nature of data, or deep models with large number of parameters prone to overfitting. In this paper, we propose a deep Generative Adversarial Clustering Network (ClusterGAN), which tackles the problems of training of deep clustering models in unsupervised manner. ClusterGAN consists of three networks, a discriminator, a generator and a clusterer (i.e. a clustering network). We employ an adversarial game between these three players to synthesize realistic samples given discriminative latent variables via the generator, and learn the inverse mapping of the real samples to the discriminative embedding space via the clusterer. Moreover, we utilize a conditional entropy minimization loss to increase/decrease the similarity of intra/inter cluster samples. Since the ground-truth similarities are unknown in clustering task, we propose a novel balanced self-paced learning algorithm to gradually include samples into training from easy to difficult, while considering the diversity of selected samples from all clusters. Therefore, our method makes it possible to efficiently train clusterers with large depth by leveraging the proposed adversarial game and balanced self-paced learning algorithm. According our experiments, ClusterGAN achieves competitive results compared to the state-of-the-art clustering and hashing models on several datasets.

A key component of Simultaneous Localization and Mapping (SLAM) systems is the joint optimization of the estimated 3D map and camera trajectory. Bundle adjustment (BA) is the gold standard for this. Due to the large number of variables in dense RGB-D SLAM, previous work has focused on approximating BA. In contrast, in this paper we present a novel, fast direct BA formulation which we implement in a real-time dense RGB-D SLAM algorithm. In addition, we show that direct RGB-D SLAM systems are highly sensitive to rolling shutter, RGB and depth sensor synchronization, and calibration errors. In order to facilitate state-of-the-art research on direct RGB-D SLAM, we propose a novel, well-calibrated benchmark for this task that uses synchronized global shutter RGB and depth cameras. It includes a training set, a test set without public ground truth, and an online evaluation service. We observe that the ranking of methods changes on this dataset compared to existing ones, and our proposed algorithm outperforms all other evaluated SLAM methods. Our benchmark and our open source SLAM algorithm are available at: www.eth3d.net

Data augmentation is an effective technique for improving the accuracy of modern image classifiers. However, current data augmentation implementations are manually designed. In this paper, we describe a simple procedure called AutoAugment to automatically search for improved data augmentation policies. In our implementation, we have designed a search space where a policy consists of many sub-policies, one of which is randomly chosen for each image in each mini-batch. A sub-policy consists of two operations, each operation being an image processing function such as translation, rotation, or shearing, and the probabilities and magnitudes with which the functions are applied. We use a search algorithm to find the best policy such that the neural network yields the highest validation accuracy on a target dataset. Our method achieves state-of-the-art accuracy on CIFAR-10, CIFAR-100, SVHN, and ImageNet (without additional data). On ImageNet, we attain a Top-1 accuracy of 83.5% which is 0.4% better than the previous record of 83.1%. On CIFAR-10, we achieve an error rate of 1.5%, which is 0.6% better than the previous state-of-the-art. Augmentation policies we find are transferable between datasets. The policy learned on ImageNet transfers well to achieve significant improvements on other datasets, such as Oxford Flowers, Caltech-101, Oxford-IIT Pets, FGVC Aircraft, and Stanford Cars.

We propose Scene Graph Auto-Encoder (SGAE) that incorporates the language inductive bias into the encoder-decoder image captioning framework for more human-like captions. Intuitively, we humans use the inductive bias to compose collocations and contextual inference in discourse. For example, when we see the relation "person on bike", it is natural to replace "on" with "ride" and infer "person riding bike on a road" even the "road" is not evident. Therefore, exploiting such bias as a language prior is expected to help the conventional encoder-decoder models less likely to overfit to the dataset bias and focus on reasoning. Specifically, we use the scene graph --- a directed graph (G) where an object node is connected by adjective nodes and relationship nodes --- to represent the complex structural layout of both image (I) and sentence (S). In the textual domain, we use SGAE to learn a dictionary (D) that helps to reconstruct sentences in the S -> G -> D -> S pipeline, where D encodes the desired language prior; in the vision-language domain, we use the shared D to guide the encoder-decoder in the I -> G -> D -> S pipeline. Thanks to the scene graph representation and shared dictionary, the inductive bias is transferred across domains in principle. We validate the effectiveness of SGAE on the challenging MS-COCO image captioning benchmark, e.g., our SGAE-based single-model achieves a new state-of-the-art 127.8 CIDEr-D on the Karpathy split, and a competitive 125.5 CIDEr-D (c40) on the official server even compared to other ensemble models. Code has been made available at: https://github.com/yangxuntu/SGAE.

Recently, Neural Architecture Search (NAS) has successfully identified neural network architectures that exceed human designed ones on large-scale image classification. In this paper, we study NAS for semantic image segmentation. Existing works often focus on searching the repeatable cell structure, while hand-designing the outer network structure that controls the spatial resolution changes. This choice simplifies the search space, but becomes increasingly problematic for dense image prediction which exhibits a lot more network level architectural variations. Therefore, we propose to search the network level structure in addition to the cell level structure, which forms a hierarchical architecture search space. We present a network level search space that includes many popular designs, and develop a formulation that allows efficient gradient-based architecture search (3 P100 GPU days on Cityscapes images). We demonstrate the effectiveness of the proposed method on the challenging Cityscapes, PASCAL VOC 2012, and ADE20K datasets. Auto-DeepLab, our architecture searched specifically for semantic image segmentation, attains state-of-the-art performance without any ImageNet pretraining.

Weakly Supervised Object Localization (WSOL) techniques learn the object location only using image-level labels, without location annotations. A common limitation for these techniques is that they cover only the most discriminative part of the object, not the entire object. To address this problem, we propose an Attention-based Dropout Layer (ADL), which utilizes the self-attention mechanism to process the feature maps of the model. The proposed method is composed of two key components: 1) hiding the most discriminative part from the model for capturing the integral extent of object, and 2) highlighting the informative region for improving the recognition power of the model. Based on extensive experiments, we demonstrate that the proposed method is effective to improve the accuracy of WSOL, achieving a new state-of-the-art localization accuracy in CUB-200-2011 dataset. We also show that the proposed method is much more efficient in terms of both parameter and computation overheads than existing techniques.

Visual explanation enables humans to understand the decision making of deep convolutional neural network (CNN), but it is insufficient to contribute to improving CNN performance. In this paper, we focus on the attention map for visual explanation, which represents a high response value as the attention location in image recognition. This attention region significantly improves the performance of CNN by introducing an attention mechanism that focuses on a specific region in an image. In this work, we propose Attention Branch Network (ABN), which extends a response-based visual explanation model by introducing a branch structure with an attention mechanism. ABN can be applicable to several image recognition tasks by introducing a branch for the attention mechanism and is trainable for visual explanation and image recognition in an end-to-end manner. We evaluate ABN on several image recognition tasks such as image classification, fine-grained recognition, and multiple facial attribute recognition. Experimental results indicate that ABN outperforms the baseline models on these image recognition tasks while generating an attention map for visual explanation. Our code is available.

We present Argoverse, a dataset designed to support autonomous vehicle perception tasks including 3D tracking and motion forecasting. Argoverse includes sensor data collected by a fleet of autonomous vehicles in Pittsburgh and Miami as well as 3D tracking annotations, 300k extracted interesting vehicle trajectories, and rich semantic maps. The sensor data consists of 360 degree images from 7 cameras with overlapping fields of view, forward-facing stereo imagery, 3D point clouds from long range LiDAR, and 6-DOF pose. Our 290km of mapped lanes contain rich geometric and semantic metadata which are not currently available in any public dataset. All data is released under a Creative Commons license at Argoverse.org. In baseline experiments, we use map information such as lane direction, driveable area, and ground height to improve the accuracy of 3D object tracking. We use 3D object tracking to mine for more than 300k interesting vehicle trajectories to create a trajectory forecasting benchmark. Motion forecasting experiments ranging in complexity from classical methods (k-NN) to LSTMs demonstrate that using detailed vector maps with lane-level information substantially reduces prediction error. Our tracking and forecasting experiments represent only a superficial exploration of the potential of rich maps in robotic perception. We hope that Argoverse will enable the research community to explore these problems in greater depth.

One of the main challenges in feature learning using Deep Convolutional Neural Networks (DCNNs) for large-scale face recognition is the design of appropriate loss functions that can enhance the discriminative power. Centre loss penalises the distance between deep features and their corresponding class centres in the Euclidean space to achieve intra-class compactness. SphereFace assumes that the linear transformation matrix in the last fully connected layer can be used as a representation of the class centres in the angular space and therefore penalises the angles between deep features and their corresponding weights in a multiplicative way. Recently, a popular line of research is to incorporate margins in well-established loss functions in order to maximise face class separability. In this paper, we propose an Additive Angular Margin Loss (ArcFace) to obtain highly discriminative features for face recognition. The proposed ArcFace has a clear geometric interpretation due to its exact correspondence to geodesic distance on a hypersphere. We present arguably the most extensive experimental evaluation against all recent state-of-the-art face recognition methods on ten face recognition benchmarks which includes a new large-scale image database with trillions of pairs and a large-scale video dataset. We show that ArcFace consistently outperforms the state of the art and can be easily implemented with negligible computational overhead. To facilitate future research, the code has been made available.

Scene text detection attracts much attention in computer vision, because it can be widely used in many applications such as real-time text translation, automatic information entry, blind person assistance, robot sensing and so on. Though many methods have been proposed for horizontal and oriented texts, detecting irregular shape texts such as curved texts is still a challenging problem. To solve the problem, we propose a robust scene text detection method with adaptive text region representation. Given an input image, a text region proposal network is first used for extracting text proposals. Then, these proposals are verified and refined with a refinement network. Here, recurrent neural network based adaptive text region representation is proposed for text region refinement, where a pair of boundary points are predicted each time step until no new points are found. In this way, text regions of arbitrary shapes are detected and represented with adaptive number of boundary points. This gives more accurate description of text regions. Experimental results on five benchmarks, namely, CTW1500, TotalText, ICDAR2013, ICDAR2015 and MSRA-TD500, show that the proposed method achieves state-of-the-art in scene text detection.