CVPR 2019 论文列表

This paper is about learning deep representations of images such that images belonging to the same class have more similar representations than those belonging to different classes. For this goal, prior work typically uses the triplet or N-pair loss, specified in terms of either l2-distances or dot-products between deep features. However, such formulations seem poorly suited to the highly non-Euclidean deep feature space. Our first contribution is in specifying the N-pair loss in terms of manifold similarities between deep features. We introduce a new time- and memory-efficient method for estimating the manifold similarities by using a closed-form convergence solution of the Random Walk algorithm. Our efficiency comes, in part, from following the recent work that randomly partitions the deep feature space, and expresses image distances via representatives of the resulting subspaces, a.k.a. proxies. Our second contribution is aimed at reducing overfitting by estimating hard proxies that are as close to one another as possible, but remain in their respective subspaces. Our evaluation demonstrates that we outperform the state of the art in both image retrieval and clustering on the benchmark CUB-200-2011, Cars196, and Stanford Online Products datasets.

Learning class-conditional data distributions is crucial for Generative Adversarial Networks (GAN) in semi-supervised learning. To improve both instance synthesis and classification in this setting, we propose an enhanced TripleGAN (EnhancedTGAN) model in this work. We follow the adversarial training scheme of the original TripleGAN, but completely re-design the training targets of the generator and classifier. Specifically, we adopt feature-semantics matching to enhance the generator in learning class-conditional distributions from both the aspects of statistics in the latent space and semantics consistency with respect to the generator and classifier. Since a limited amount of labeled data is not sufficient to determine satisfactory decision boundaries, we include two classifiers, and incorporate collaborative learning into our model to provide better guidance for generator training. The synthesized high-fidelity data can in turn be used for improving classifier training. In the experiments, the superior performance of our approach on multiple benchmark datasets demonstrates the effectiveness of the mutual reinforcement between the generator and classifiers in facilitating semi-supervised instance synthesis and classification.

In this paper, we reduce the image dehazing problem to an image-to-image translation problem, and propose Enhanced Pix2pix Dehazing Network (EPDN), which generates a haze-free image without relying on the physical scattering model. EPDN is embedded by a generative adversarial network, which is followed by a well-designed enhancer. Inspired by visual perception global-first theory, the discriminator guides the generator to create a pseudo realistic image on a coarse scale, while the enhancer following the generator is required to produce a realistic dehazing image on the fine scale. The enhancer contains two enhancing blocks based on the receptive field model, which reinforces the dehazing effect in both color and details. The embedded GAN is jointly trained with the enhancer. Extensive experiment results on synthetic datasets and real-world datasets show that the proposed EPDN is superior to the state-of-the-art methods in terms of PSNR, SSIM, PI, and subjective visual effect.

In this paper, we propose an Enhanced Bayesian Compression method to flexibly compress the deep networks via reinforcement learning. Unlike the existing Bayesian compression method which cannot explicitly enforce quantization weights during training, our method learns flexible codebooks in each layer for an optimal network quantization. To dynamically adjust the state of codebooks, we employ an Actor-Critic network to collaborate with the original deep network. Different from most existing network quantization methods, our EBC does not require re-training procedures after the quantization. Experimental results show that our method obtains low-bit precision with acceptable accuracy drop on MNIST, CIFAR and ImageNet.

Standard image captioning tasks such as COCO and Flickr30k are factual, neutral in tone and (to a human) state the obvious (e.g., "a man playing a guitar"). While such tasks are useful to verify that a machine understands the content of an image, they are not engaging to humans as captions. With this in mind we define a new task, PERSONALITY-CAPTIONS, where the goal is to be as engaging to humans as possible by incorporating controllable style and personality traits. We collect and release a large dataset of 241,858 of such captions conditioned over 215 possible traits. We build models that combine existing work from (i) sentence representations [36] with Transformers trained on 1.7 billion dialogue examples; and (ii) image representations [32] with ResNets trained on 3.5 billion social media images. We obtain state-of-the-art performance on Flickr30k and COCO, and strong performance on our new task. Finally, online evaluations validate that our task and models are engaging to humans, with our best model close to human performance.

Time-lapse videos usually contain visually appealing content but are often difficult and costly to create. In this paper, we present an end-to-end solution to synthesize a time-lapse video from a single outdoor image using deep neural networks. Our key idea is to train a conditional generative adversarial network based on existing datasets of time-lapse videos and image sequences. We propose a multi-frame joint conditional generation framework to effectively learn the correlation between the illumination change of an outdoor scene and the time of the day. We further present a multi-domain training scheme for robust training of our generative models from two datasets with different distributions and missing timestamp labels. Compared to alternative time-lapse video synthesis algorithms, our method uses the timestamp as the control variable and does not require a reference video to guide the synthesis of the final output. We conduct ablation studies to validate our algorithm and compare with state-of-the-art techniques both qualitatively and quantitatively.

Product Quantization, a dictionary based hashing method, is one of the leading unsupervised hashing techniques. While it ignores the labels, it harnesses the features to construct look up tables that can approximate the feature space. In recent years, several works have achieved state of the art results on hashing benchmarks by learning binary representations in a supervised manner. This work presents Deep Product Quantization (DPQ), a technique that leads to more accurate retrieval and classification than the latest state of the art methods, while having similar computational complexity and memory footprint as the Product Quantization method. To our knowledge, this is the first work to introduce a dictionary-based representation that is inspired by Product Quantization and which is learned end-to-end, and thus benefits from the supervised signal. DPQ explicitly learns soft and hard representations to enable an efficient and accurate asymmetric search, by using a straight-through estimator. Our method obtains state of the art results on an extensive array of retrieval and classification experiments.

We propose a novel multi-task learning architecture, which allows learning of task-specific feature-level attention. Our design, the Multi-Task Attention Network (MTAN), consists of a single shared network containing a global feature pool, together with a soft-attention module for each task. These modules allow for learning of task-specific features from the global features, whilst simultaneously allowing for features to be shared across different tasks. The architecture can be trained end-to-end and can be built upon any feed-forward neural network, is simple to implement, and is parameter efficient. We evaluate our approach on a variety of datasets, across both image-to-image predictions and image classification tasks. We show that our architecture is state-of-the-art in multi-task learning compared to existing methods, and is also less sensitive to various weighting schemes in the multi-task loss function. Code is available at https://github.com/lorenmt/mtan.

We present an end-to-end learned algorithm for seeded segmentation. Our method is based on the Random Walker algorithm, where we predict the edge weights of the un- derlying graph using a convolutional neural network. This can be interpreted as learning context-dependent diffusiv- ities for a linear diffusion process. After calculating the exact gradient for optimizing these diffusivities, we pro- pose simplifications that sparsely sample the gradient while still maintaining competitive results. The proposed method achieves the currently best results on the seeded CREMI neuron segmentation challenge.

We develop hierarchically quantized efficient embedding representations for similarity-based search and show that this representation provides not only the state of the art performance on the search accuracy but also provides several orders of speed up during inference. The idea is to hierarchically quantize the representation so that the quantization granularity is greatly increased while maintaining the accuracy and keeping the computational complexity low. We also show that the problem of finding the optimal sparse compound hash code respecting the hierarchical structure can be optimized in polynomial time via minimum cost flow in an equivalent flow network. This allows us to train the method end-to-end in a mini-batch stochastic gradient descent setting. Our experiments on Cifar100 and ImageNet datasets show the state of the art search accuracy while providing several orders of magnitude search speedup respectively over exhaustive linear search over the dataset.

Despite the recent success in face recognition and object classification, in the field of human gaze prediction, computer models are still struggling to accurately mimic human attention. One main reason is that visual attention is a complex human behavior influenced by multiple factors, ranging from low-level features (e.g., color, contrast) to high-level human perception (e.g., objects interactions, object sentiment), making it difficult to model computationally. In this work, we investigate the relation between object sentiment and human attention. We first introduce a new evaluation metric (AttI) for measuring human attention that focuses on human fixation consensus. A series of empirical data analyses with AttI indicate that emotion-evoking objects receive attention favor, especially when they co-occur with emotionally-neutral objects, and this favor varies with different image complexity. Based on the empirical analyses, we design a deep neural network for human attention prediction which allows the attention bias on emotion-evoking objects to be encoded in its feature space. Experiments on two benchmark datasets demonstrate its superior performance, especially on metrics that evaluate relative importance of salient regions. This research provides the clearest picture to date on how object sentiments influence human attention, and it makes one of the first attempts to model this phenomenon computationally.

In this paper, a deep embedding algorithm is developed to achieve higher accuracy rates on large-scale image classification. By adapting the importance of the object classes to their error rates, our deep embedding algorithm can train multiple complementary deep networks sequentially, where each of them focuses on achieving higher accuracy rates for different subsets of object classes in an easy-to-hard way. By integrating such complementary deep networks to generate an ensemble network, our deep embedding algorithm can improve the accuracy rates for the hard object classes (which initially have higher error rates) at certain degrees while effectively preserving high accuracy rates for the easy object classes. Our deep embedding algorithm has achieved higher overall accuracy rates on large scale image classification.

We focus on an important yet challenging problem: using a 2D deep network to deal with 3D segmentation for medical image analysis. Existing approaches either applied multi-view planar (2D) networks or directly used volumetric (3D) networks for this purpose, but both of them are not ideal: 2D networks cannot capture 3D contexts effectively, and 3D networks are both memory-consuming and less stable arguably due to the lack of pre-trained models. In this paper, we bridge the gap between 2D and 3D using a novel approach named Elastic Boundary Projection (EBP). The key observation is that, although the object is a 3D volume, what we really need in segmentation is to find its boundary which is a 2D surface. Therefore, we place a number of pivot points in the 3D space, and for each pivot, we determine its distance to the object boundary along a dense set of directions. This creates an elastic shell around each pivot which is initialized as a perfect sphere. We train a 2D deep network to determine whether each ending point falls within the object, and gradually adjust the shell so that it gradually converges to the actual shape of the boundary and thus achieves the goal of segmentation. EBP allows boundary-based segmentation without cutting a 3D volume into slices or patches, which stands out from conventional 2D and 3D approaches. EBP achieves promising accuracy in abdominal organ segmentation. Our code will be released on https://github.com/twni2016/Elastic-Boundary-Projection .

Network compression reduces the computational complexity and memory consumption of deep neural networks by reducing the number of parameters. In SVD-based network compression the right rank needs to be decided for every layer of the network. In this paper we propose an efficient method for obtaining the rank configuration of the whole network. Unlike previous methods which consider each layer separately, our method considers the whole network to choose the right rank configuration. We propose novel accuracy metrics to represent the accuracy and complexity relationship for a given neural network. We use these metrics in a non-iterative fashion to obtain the right rank configuration which satisfies the constraints on FLOPs and memory while maintaining sufficient accuracy. Experiments show that our method provides better compromise between accuracy and computational complexity/memory consumption while performing compression at much higher speed. For VGG-16 our network can reduce the FLOPs by 25% and improve accuracy by 0.7% compared to the baseline, while requiring only 3 minutes on a CPU to search for the right rank configuration. Previously, similar results were achieved in 4 hours with 8 GPUs. The proposed method can be used for lossless compression of a neural network as well. The better accuracy and complexity compromise, as well as the extremely fast speed of our method make it suitable for neural network compression.

The problem of multi-domain learning of deep networks is considered. An adaptive layer is induced per target domain and a novel procedure, denoted covariance normalization (CovNorm), proposed to reduce its parameters. CovNorm is a data driven method of fairly simple implementation, requiring two principal component analyzes (PCA) and fine-tuning of a mini-adaptation layer. Nevertheless, it is shown, both theoretically and experimentally, to have several advantages over previous approaches, such as batch normalization or geometric matrix approximations. Furthermore, CovNorm can be deployed both when target datasets are available sequentially or simultaneously. Experiments show that, in both cases, it has performance comparable to a fully fine-tuned network, using as few as 0.13% of the corresponding parameters per target domain.

Single-stage object detectors have recently gained popularity due to their combined advantage of high detection accuracy and real-time speed. However, while promising results have been achieved by these detectors on standard-sized objects, their performance on small objects is far from satisfactory. To detect very small/large objects, classical pyramid representation can be exploited, where an image pyramid is used to build a feature pyramid (featurized image pyramid), enabling detection across a range of scales. Existing single-stage detectors avoid such a featurized image pyramid representation due to its memory and time complexity. In this paper, we introduce a light-weight architecture to efficiently produce featurized image pyramid in a single-stage detection framework. The resulting multi-scale features are then injected into the prediction layers of the detector using an attention module. The performance of our detector is validated on two benchmarks: PASCAL VOC and MS COCO. For a 300x300 input, our detector operates at 111 frames per second (FPS) on a Titan X GPU, providing state-of-the-art detection accuracy on PASCAL VOC 2007 testset. On the MS COCO testset, our detector achieves state-of-the-art results surpassing all existing single-stage methods in the case of single-scale inference.

Face recognition has obtained remarkable progress in recent years due to the great improvement of deep convolutional neural networks (CNNs). However, deep CNNs are vulnerable to adversarial examples, which can cause fateful consequences in real-world face recognition applications with security-sensitive purposes. Adversarial attacks are widely studied as they can identify the vulnerability of the models before they are deployed. In this paper, we evaluate the robustness of state-of-the-art face recognition models in the decision-based black-box attack setting, where the attackers have no access to the model parameters and gradients, but can only acquire hard-label predictions by sending queries to the target model. This attack setting is more practical in real-world face recognition systems. To improve the efficiency of previous methods, we propose an evolutionary attack algorithm, which can model the local geometry of the search directions and reduce the dimension of the search space. Extensive experiments demonstrate the effectiveness of the proposed method that induces a minimum perturbation to an input face image with fewer queries. We also apply the proposed method to attack a real-world face recognition system successfully.

We propose an effective deep learning approach to aesthetics quality assessment that relies on a new type of pre-trained features, and apply it to the AVA data set, the currently largest aesthetics database. While previous approaches miss some of the information in the original images, due to taking small crops, down-scaling or warping the originals during training, we propose the first method that efficiently supports full resolution images as an input, and can be trained on variable input sizes. This allows us to significantly improve upon the state of the art, increasing the Spearman rank-order correlation coefficient (SRCC) of ground-truth mean opinion scores (MOS) from the existing best reported of 0.612 to 0.756. To achieve this performance, we extract multi-level spatially pooled (MLSP) features from all convolutional blocks of a pre-trained InceptionResNet-v2 network, and train a custom shallow Convolutional Neural Network (CNN) architecture on these new features.

In this paper, we introduce a new type of sensing modality, the Dynamic Vision Sensors (Event Cameras), for the task of gait recognition. Compared with the traditional RGB sensors, the event cameras have many unique advantages such as ultra low resources consumption, high temporal resolution and much larger dynamic range. However, those cameras only produce noisy and asynchronous events of intensity changes rather than frames, where conventional vision-based gait recognition algorithms can't be directly applied. To address this, we propose a new Event-based Gait Recognition (EV-Gait) approach, which exploits motion consistency to effectively remove noise, and uses a deep neural network to recognise gait from the event streams. To evaluate the performance of EV-Gait, we collect two event-based gait datasets, one from real-world experiments and the other by converting the publicly available RGB gait recognition benchmark CASIA-B. Extensive experiments show that EV-Gait can get nearly 96% recognition accuracy in the real-world settings, while on the CASIA-B benchmark it achieves comparable performance with state-of-the-art RGB-based gait recognition approaches.