CVPR 2019 论文列表

The superior performance of Deformable Convolutional Networks arises from its ability to adapt to the geometric variations of objects. Through an examination of its adaptive behavior, we observe that while the spatial support for its neural features conforms more closely than regular ConvNets to object structure, this support may nevertheless extend well beyond the region of interest, causing features to be influenced by irrelevant image content. To address this problem, we present a reformulation of Deformable ConvNets that improves its ability to focus on pertinent image regions, through increased modeling power and stronger training. The modeling power is enhanced through a more comprehensive integration of deformable convolution within the network, and by introducing a modulation mechanism that expands the scope of deformation modeling. To effectively harness this enriched modeling capability, we guide network training via a proposed feature mimicking scheme that helps the network to learn features that reflect the object focus and classification power of R-CNN features. With the proposed contributions, this new version of Deformable ConvNets yields significant performance gains over the original model and produces leading results on the COCO benchmark for object detection and instance segmentation.

A plethora of recent work has shown that convolutional networks are not robust to adversarial images: images that are created by perturbing a sample from the data distribution as to maximize the loss on the perturbed example. In this work, we hypothesize that adversarial perturbations move the image away from the image manifold in the sense that there exists no physical process that could have produced the adversarial image. This hypothesis suggests that a successful defense mechanism against adversarial images should aim to project the images back onto the image manifold. We study such defense mechanisms, which approximate the projection onto the unknown image manifold by a nearest-neighbor search against a web-scale image database containing tens of billions of images. Empirical evaluations of this defense strategy on ImageNet suggest that it very effective in attack settings in which the adversary does not have access to the image database. We also propose two novel attack methods to break nearest-neighbor defense settings and show conditions under which nearest-neighbor defense fails. We perform a series of ablation experiments, which suggest that there is a trade-off between robustness and accuracy between as we use features from deeper in the network, that a large index size (hundreds of millions) is crucial to get good performance, and that careful construction of database is crucial for robustness against nearest-neighbor attacks.

The vulnerability of machine learning systems to adversarial attacks questions their usage in many applications. In this paper, we propose a randomized diversification as a defense strategy. We introduce a multi-channel architecture in a gray-box scenario, which assumes that the architecture of the classifier and the training data set are known to the attacker. The attacker does not only have access to a secret key and to the internal states of the system at the test time. The defender processes an input in multiple channels. Each channel introduces its own randomization in a special transform domain based on a secret key shared between the training and testing stages. Such a transform based randomization with a shared key preserves the gradients in key-defined sub-spaces for the defender but it prevents gradient back propagation and the creation of various bypass systems for the attacker. An additional benefit of multi-channel randomization is the aggregation that fuses soft-outputs from all channels, thus increasing the reliability of the final score. The sharing of a secret key creates an information advantage to the defender. Experimental evaluation demonstrates an increased robustness of the proposed method to a number of known state-of-the-art attacks.

Convolutional Neural Networks (CNNs) have become deeper and more complicated compared with the pioneering AlexNet. However, current prevailing training scheme follows the previous way of adding supervision to the last layer of the network only and propagating error information up layer-by-layer. In this paper, we propose Deeply-supervised Knowledge Synergy (DKS), a new method aiming to train CNNs with improved generalization ability for image classification tasks without introducing extra computational cost during inference. Inspired by the deeply-supervised learning scheme, we first append auxiliary supervision branches on top of certain intermediate network layers. While properly using auxiliary supervision can improve model accuracy to some degree, we go one step further to explore the possibility of utilizing the probabilistic knowledge dynamically learnt by the classifiers connected to the backbone network as a new regularization to improve the training. A novel synergy loss, which considers pairwise knowledge matching among all supervision branches, is presented. Intriguingly, it enables dense pairwise knowledge matching operations in both top-down and bottom-up directions at each training iteration, resembling a dynamic synergy process for the same task. We evaluate DKS on image classification datasets using state-of-the-art CNN architectures, and show that the models trained with it are consistently better than the corresponding counterparts. For instance, on the ImageNet classification benchmark, our ResNet-152 model outperforms the baseline model with a 1.47% margin in Top-1 accuracy. Code is available at https://github.com/sundw2014/DKS.

Siamese networks have drawn great attention in visual tracking because of their balanced accuracy and speed. However, the backbone networks used in Siamese trackers are relatively shallow, such as AlexNet, which does not fully take advantage of the capability of modern deep neural networks. In this paper, we investigate how to leverage deeper and wider convolutional neural networks to enhance tracking robustness and accuracy. We observe that direct replacement of backbones with existing powerful architectures, such as ResNet and Inception, does not bring improvements. The main reasons are that 1) large increases in the receptive field of neurons lead to reduced feature discriminability and localization precision; and 2) the network padding for convolutions induces a positional bias in learning. To address these issues, we propose new residual modules to eliminate the negative impact of padding, and further design new architectures using these modules with controlled receptive field size and network stride. The designed architectures are lightweight and guarantee real-time tracking speed when applied to SiamFC and SiamRPN. Experiments show that solely due to the proposed network architectures, our SiamFC+ and SiamRPN+ obtain up to 9.8%/6.3% (AUC), 23.3%/8.8% (EAO) and 24.4%/25.0% (EAO) relative improvements over the original versions on the OTB-15, VOT-16 and VOT-17 datasets, respectively.

In this work, we address the lack of 3D understanding of generative neural networks by introducing a persistent 3D feature embedding for view synthesis. To this end, we propose DeepVoxels, a learned representation that encodes the view-dependent appearance of a 3D scene without having to explicitly model its geometry. At its core, our approach is based on a Cartesian 3D grid of persistent embedded features that learn to make use of the underlying 3D scene structure. Our approach combines insights from 3D geometric computer vision with recent advances in learning image-to-image mappings based on adversarial loss functions. DeepVoxels is supervised, without requiring a 3D reconstruction of the scene, using a 2D re-rendering loss and enforces perspective and multi-view geometry in a principled manner. We apply our persistent 3D scene representation to the problem of novel view synthesis demonstrating high-quality results for a variety of challenging scenes.

We present a novel approach to view synthesis using multiplane images (MPIs). Building on recent advances in learned gradient descent, our algorithm generates an MPI from a set of sparse camera viewpoints. The resulting method incorporates occlusion reasoning, improving performance on challenging scene features such as object boundaries, lighting reflections, thin structures, and scenes with high depth complexity. We show that our method achieves high-quality, state-of-the-art results on two datasets: the Kalantari light field dataset, and a new camera array dataset, Spaces, which we make publicly available.

Computer graphics, 3D computer vision and robotics communities have produced multiple approaches to representing 3D geometry for rendering and reconstruction. These provide trade-offs across fidelity, efficiency and compression capabilities. In this work, we introduce DeepSDF, a learned continuous Signed Distance Function (SDF) representation of a class of shapes that enables high quality shape representation, interpolation and completion from partial and noisy 3D input data. DeepSDF, like its classical counterpart, represents a shape's surface by a continuous volumetric field: the magnitude of a point in the field represents the distance to the surface boundary and the sign indicates whether the region is inside (-) or outside (+) of the shape, hence our representation implicitly encodes a shape's boundary as the zero-level-set of the learned function while explicitly representing the classification of space as being part of the shapes interior or not. While classical SDF's both in analytical or discretized voxel form typically represent the surface of a single shape, DeepSDF can represent an entire class of shapes. Furthermore, we show state-of-the-art performance for learned 3D shape representation and completion while reducing the model size by an order of magnitude compared with previous work.

We propose DeepMapping, a novel registration framework using deep neural networks (DNNs) as auxiliary functions to align multiple point clouds from scratch to a globally consistent frame. We use DNNs to model the highly non-convex mapping process that traditionally involves hand-crafted data association, sensor pose initialization, and global refinement. Our key novelty is that "training" these DNNs with properly defined unsupervised losses is equivalent to solving the underlying registration problem, but less sensitive to good initialization than ICP. Our framework contains two DNNs: a localization network that estimates the poses for input point clouds, and a map network that models the scene structure by estimating the occupancy status of global coordinates. This allows us to convert the registration problem to a binary occupancy classification, which can be solved efficiently using gradient-based optimization. We further show that DeepMapping can be readily extended to address the problem of Lidar SLAM by imposing geometric constraints between consecutive point clouds. Experiments are conducted on both simulated and real datasets. Qualitative and quantitative comparisons demonstrate that DeepMapping often enables more robust and accurate global registration of multiple point clouds than existing techniques. Our code is available at https://ai4ce.github.io/DeepMapping/.

We present a learning-based method to infer plausible high dynamic range (HDR), omnidirectional illumination given an unconstrained, low dynamic range (LDR) image from a mobile phone camera with a limited field of view (FOV). For training data, we collect videos of various reflective spheres placed within the camera's FOV, leaving most of the background unoccluded, leveraging that materials with diverse reflectance functions reveal different lighting cues in a single exposure. We train a deep neural network to regress from the LDR background image to HDR lighting by matching the LDR ground truth sphere images to those rendered with the predicted illumination using image-based relighting, which is differentiable. Our inference runs at interactive frame rates on a mobile device, enabling realistic rendering of virtual objects into real scenes for mobile mixed reality. Training on automatically exposed and white-balanced videos, we improve the realism of rendered objects compared to the state-of-the art methods for both indoor and outdoor scenes.

In this paper, we propose a deep learning architecture that produces accurate dense depth for the outdoor scene from a single color image and a sparse depth. Inspired by the indoor depth completion, our network estimates surface normals as the intermediate representation to produce dense depth, and can be trained end-to-end. With a modified encoder-decoder structure, our network effectively fuses the dense color image and the sparse LiDAR depth. To address outdoor specific challenges, our network predicts a confidence mask to handle mixed LiDAR signals near foreground boundaries due to occlusion, and combines estimates from the color image and surface normals with learned attention maps to improve the depth accuracy especially for distant areas. Extensive experiments demonstrate that our model improves upon the state-of-the-art performance on KITTI depth completion benchmark. Ablation study shows the positive impact of each model components to the final performance, and comprehensive analysis shows that our model generalizes well to the input with higher sparsity or from indoor scenes.

Computing object skeletons in natural images is challenging, owing to large variations in object appearance and scale, and the complexity of handling background clutter. Many recent methods frame object skeleton detection as a binary pixel classification problem, which is similar in spirit to learning-based edge detection, as well as to semantic segmentation methods. In the present article, we depart from this strategy by training a CNN to predict a two-dimensional vector field, which maps each scene point to a candidate skeleton pixel, in the spirit of flux-based skeletonization algorithms. This "image context flux" representation has two major advantages over previous approaches. First, it explicitly encodes the relative position of skeletal pixels to semantically meaningful entities, such as the image points in their spatial context, and hence also the implied object boundaries. Second, since the skeleton detection context is a region-based vector field, it is better able to cope with object parts of large width. We evaluate the proposed method on three benchmark datasets for skeleton detection and two for symmetry detection, achieving consistently superior performance over state-of-the-art methods.

Understanding fashion images has been advanced by benchmarks with rich annotations such as DeepFashion, whose labels include clothing categories, landmarks, and consumer-commercial image pairs. However, DeepFashion has nonnegligible issues such as single clothing-item per image, sparse landmarks (4 8 only), and no per-pixel masks, making it had significant gap from real-world scenarios. We fill in the gap by presenting DeepFashion2 to address these issues. It is a versatile benchmark of four tasks including clothes detection, pose estimation, segmentation, and retrieval. It has 801K clothing items where each item has rich annotations such as style, scale, view- point, occlusion, bounding box, dense landmarks (e.g. 39 for 'long sleeve outwear' and 15 for 'vest'), and masks. There are also 873K Commercial-Consumer clothes pairs. The annotations of DeepFashion2 are much larger than its counterparts such as 8x of FashionAI Global Challenge. A strong baseline is proposed, called Match R- CNN, which builds upon Mask R-CNN to solve the above four tasks in an end-to-end manner. Extensive evaluations are conducted with different criterions in Deep- Fashion2. DeepFashion2 Dataset will be released at : https://github.com/switchablenorms/DeepFashion2

Capsule Network is a promising concept in deep learning, yet its true potential is not fully realized thus far, providing sub-par performance on several key benchmark datasets with complex data. Drawing intuition from the success achieved by Convolutional Neural Networks (CNNs) by going deeper, we introduce DeepCaps, a deep capsule network architecture which uses a novel 3D convolution based dynamic routing algorithm. With DeepCaps, we surpass the state-of-the-art capsule domain networks results on CIFAR10, SVHN and Fashion MNIST, while achieving a 68% reduction in the number of parameters. Further, we propose a class independent decoder network, which strengthens the use of reconstruction loss as a regularization term. This leads to an interesting property of the decoder, which allows us to identify and control the physical attributes of the images represented by the instantiation parameters.

In this paper, we address a new task called instance co-segmentation. Given a set of images jointly covering object instances of a specific category, instance co-segmentation aims to identify all of these instances and segment each of them, i.e. generating one mask for each instance. This task is important since instance-level segmentation is preferable for humans and many vision applications. It is also challenging because no pixel-wise annotated training data are available and the number of instances in each image is unknown. We solve this task by dividing it into two sub-tasks, co-peak search and instance mask segmentation. In the former sub-task, we develop a CNN-based network to detect the co-peaks as well as co-saliency maps for a pair of images. A co-peak has two endpoints, one in each image, that are local maxima in the response maps and similar to each other. Thereby, the two endpoints are potentially covered by a pair of instances of the same category. In the latter subtask, we design a ranking function that takes the detected co-peaks and co-saliency maps as inputs and can select the object proposals to produce the final results. Our method for instance co-segmentation and its variant for object colocalization are evaluated on four datasets, and achieve favorable performance against the state-of-the-art methods. The source codes and the collected datasets are available at https://github.com/KuangJuiHsu/DeepCO3/

Deep networks consume a large amount of memory by their nature. A natural question arises can we reduce that memory requirement whilst maintaining performance. In particular, in this work we address the problem of memory efficient learning for multiple tasks. To this end, we propose a novel network architecture producing multiple networks of different configurations, termed deep virtual networks (DVNs), for different tasks. Each DVN is specialized for a single task and structured hierarchically. The hierarchical structure, which contains multiple levels of hierarchy corresponding to different numbers of parameters, enables multiple inference for different memory budgets. The building block of a deep virtual network is based on a disjoint collection of parameters of a network, which we call a unit. The lowest level of hierarchy in a deep virtual network is a unit, and higher levels of hierarchy contain lower levels' units and other additional units. Given a budget on the number of parameters, a different level of a deep virtual network can be chosen to perform the task. A unit can be shared by different DVNs, allowing multiple DVNs in a single network. In addition, shared units provide assistance to the target task with additional knowledge learned from another tasks. This cooperative configuration of DVNs makes it possible to handle different tasks in a memory-aware manner. Our experiments show that the proposed method outperforms existing approaches for multiple tasks. Notably, ours is more efficient than others as it allows memory-aware inference for all tasks.

Video inpainting aims to fill spatio-temporal holes with plausible content in a video. Despite tremendous progress of deep neural networks for image inpainting, it is challenging to extend these methods to the video domain due to the additional time dimension. In this work, we propose a novel deep network architecture for fast video inpainting. Built upon an image-based encoder-decoder model, our framework is designed to collect and refine information from neighbor frames and synthesize still-unknown regions. At the same time, the output is enforced to be temporally consistent by a recurrent feedback and a temporal memory module. Compared with the state-of-the-art image inpainting algorithm, our method produces videos that are much more semantically correct and temporally smooth. In contrast to the prior video completion method which relies on time-consuming optimization, our method runs in near real-time while generating competitive video results. Finally, we applied our framework to video retargeting task, and obtain visually pleasing results.

Face anti-spoofing is designed to keep face recognition systems from recognizing fake faces as the genuine users. While advanced face anti-spoofing methods are developed, new types of spoof attacks are also being created and becoming a threat to all existing systems. We define the detection of unknown spoof attacks as Zero-Shot Face Anti-spoofing (ZSFA). Previous works of ZSFA only study 1-2 types of spoof attacks, such as print/replay attacks, which limits the insight of this problem. In this work, we expand the ZSFA problem to a wide range of 13 types of spoof attacks, including print attack, replay attack, 3D mask attacks, and so on. A novel Deep Tree Network (DTN) is proposed to tackle the ZSFA. The tree is learned to partition the spoof samples into semantic sub-groups in an unsupervised fashion. When a data sample arrives, being know or unknown attacks, DTN routes it to the most similar spoof cluster, and make the binary decision. In addition, to enable the study of ZSFA, we introduce the first face anti-spoofing database that contains diverse types of spoof attacks. Experiments show that our proposed method achieves the state of the art on multiple testing protocols of ZSFA.

We propose a transfer learning-based solution for the problem of multiple class novelty detection. In particular, we propose an end-to-end deep-learning based approach in which we investigate how the knowledge contained in an external, out-of-distributional dataset can be used to improve the performance of a deep network for visual novelty detection. Our solution differs from the standard deep classification networks on two accounts. First, we use a novel loss function, membership loss, in addition to the classical cross-entropy loss for training networks. Secondly, we use the knowledge from the external dataset more effectively to learn globally negative filters, filters that respond to generic objects outside the known class set. We show that thresholding the maximal activation of the proposed network can be used to identify novel objects effectively. Extensive experiments on four publicly available novelty detection datasets show that the proposed method achieves significant improvements over the state-of-the-art methods.