ECCV 2018 论文列表

We present a method for estimating detailed scene illumination using human faces in a single image. In contrast to previous works that estimate lighting in terms of low-order basis functions or distant point lights, our technique estimates illumination at a higher precision in the form of a non-parametric environment map. Based on the observation that faces can exhibit strong highlight reflections from a broad range of lighting directions, we propose a deep neural network for extracting highlights from faces, and then trace these reflections back to the scene to acquire the environment map. Since real training data for highlight extraction is very limited, we introduce an unsupervised scheme for finetuning the network on real images, based on the consistent diffuse chromaticity of a given face seen in multiple real images. In tracing the estimated highlights to the environment, we reduce the blurring effect of skin reflectance on reflected light through a deconvolution determined by prior knowledge on face material properties. Comparisons to previous techniques for highlight extraction and illumination estimation show the state-of-the-art performance of this approach on a variety of indoor and outdoor scenes.

State-of-the-art face super-resolution methods use deep convolutional neural networks to learn a mapping between low-resolution (LR) facial patterns and their corresponding high-resolution (HR) counterparts by exploring local information. However, most of them do not account for face structure and suffer from degradations due to large pose variations and misalignments of faces. Our method incorporates structural information of faces explicitly into face super-resolution by using a multi-task convolutional neural network (CNN). Our CNN has two branches: one for super-resolving face images and the other branch for predicting salient regions of a face coined facial component heatmaps. These heatmaps guide the up-sampling stream for generating better super-resolved faces with high-quality details. Our method uses not only the low-level information (ie intensity similarity), but also middle-level information (ie face structure) to further explore spatial constraints of facial components from LR inputs images. Therefore, we are able to super-resolve very small unaligned face images (16$ imes$16 pixels) with a large upscaling factor of 8$ imes$ while preserving face structure. Extensive experiments demonstrate that our network achieves superior face hallucination results and outperforms the state-of-the-art.

In current face recognition approaches with convolutional neural network (CNN), a pair of faces to compare are independently fed into the CNN for feature extraction. For both faces the same kernels are applied and hence the representation of a face stays fixed regardless of who it is compared with. As for us humans, however, one generally focuses on varied characteristics of a face when comparing it with distinct persons. Inspired, we propose a novel CNN structure with what we referred to as contrastive convolution, which specifically focuses on the distinct characteristics between the two faces to compare, i.e., those contrastive characteristics. Extensive experiments on the challenging LFW, and IJB-A show that our proposed contrastive convolution significantly improves the vanilla CNN and achieves quite promising performance in face verification task.

Many prior face anti-spoofing works develop discriminative models for recognizing the subtle differences between live and spoof faces. Those approaches often regard the image as an indivisible unit, and process it holistically, without explicit modeling of the spoofing process. In this work, motivated by the noise modeling and denoising algorithms, we identify a new problem of face de-spoofing, for the purpose of anti-spoofing: inversely decomposing a spoof face into a spoof noise and a live face, and then utilizing the spoof noise for classification. A CNN architecture with proper constraints and supervisions is proposed to overcome the problem of having no ground truth for the decomposition. We evaluate the proposed method on multiple face anti-spoofing databases. The results show promising improvements due to our spoof noise modeling. Moreover, the estimated spoof noise provides a visualization which helps to understand the added spoof noise by each spoof medium.

In this paper we propose a novel decomposition method based on filter group approximation, which can significantly reduce the redundancy of deep convolutional neural networks (CNNs) while maintaining the majority of feature representation. Unlike other low-rank decomposition algorithms which operate on spatial or channel dimension of filters, our proposed method mainly focuses on exploiting the filter group structure for each layer. For several commonly used CNN models, including VGG and ResNet, our method can reduce over 80% floating-point operations (FLOPs) with less accuracy drop than state-of-the-art methods on various image classification datasets. Besides, experiments demonstrate that our method is conducive to alleviating degeneracy of the compressed network, which hurts the convergence and performance of the network.

We consider the problem of inferring a layered representa-tion, its depth ordering and motion segmentation from a video in whichobjects may undergo 3D non-planar motion relative to the camera. Wegeneralize layered inference to the aforementioned case and correspond-ing self-occlusion phenomena. We accomplish this by introducing a flat-tened 3D object representation, which is a compact representation of anobject that contains all visible portions of the object seen in the video,including parts of an object that are self-occluded (as well as occluded)in one frame but seen in another. We formulate the inference of such flat-tened representations and motion segmentation, and derive an optimiza-tion scheme. We also introduce a new depth ordering scheme, which isindependent of layered inference and addresses the case of self-occlusion.It requires almost no computation given the flattened representations.Experiments on benchmark datasets show the advantage of our methodcompared to existing layered methods, which do not model 3D motionand self-occlusion.

State-of-the-art visual perception models for a wide range of tasks rely on supervised pretraining. ImageNet classification is the de facto pretraining task for these models. Yet, ImageNet is now nearly ten years old and is by modern standards "small". Even so, relatively little is known about the behavior of pretraining with datasets that are multiple orders of magnitude larger. The reasons are obvious: such datasets are difficult to collect and annotate. In this paper, we present a unique study of transfer learning with large convolutional networks trained to predict hashtags on billions of social media images. Our experiments demonstrate that training for large-scale hashtag prediction leads to excellent results. We show improvements on several image classification and object detection tasks, and report the highest ImageNet-1k single-crop, top-1 accuracy to date: 85.4% (97.6% top-5). We also perform extensive experiments that provide novel empirical data on the relationship between large-scale pretraining and transfer learning performance.

It is always well believed that modeling relationships between objects would be helpful for representing and eventually describing an image. Nevertheless, there has not been evidence in support of the idea on image description generation. In this paper, we introduce a new design to explore the connections between objects for image captioning under the umbrella of attention-based encoder-decoder framework. Specifically, we present Graph Convolutional Networks plus Long Short-Term Memory (dubbed as GCN-LSTM) architecture that novelly integrates both semantic and spatial object relationships into image encoder. Technically, we build graphs over the detected objects in an image based on their spatial and semantic connections. The representations of each region proposed on objects are then refined by leveraging graph structure through GCN. With the learnt region-level features, our GCN-LSTM capitalizes on LSTM-based captioning framework with attention mechanism for sentence generation. Extensive experiments are conducted on COCO image captioning dataset, and superior results are reported when comparing to state-of-the-art approaches. More remarkably, GCN-LSTM increases CIDEr-D performance from 120.1% to 128.7% on COCO testing set.

In this work, we address the problem of 3D human pose estimation from a sequence of 2D human poses. Although the recent success of deep networks has led many state-of-the-art methods for 3D pose estimation to train deep networks end-to-end to predict from images directly, the top-performing approaches have shown the effectiveness of dividing the task of 3D pose estimation into two steps: using a state-of-the-art 2D pose estimator to estimate the 2D pose from images and then mapping them into 3D space. They also showed that a low-dimensional representation like 2D locations of a set of joints can be discriminative enough to estimate 3D pose with high accuracy. However, estimation of 3D pose for individual frames leads to temporally incoherent estimates due to independent error in each frame causing jitter. Therefore, in this work we utilize the temporal information across a sequence of 2D joint locations to estimate a sequence of 3D poses. We designed a sequence-to-sequence network composed of layer-normalized LSTM units with shortcut connections connecting the input to the output on the decoder side and imposed temporal smoothness constraint during training. We found that the knowledge of temporal consistency improves the best reported result on Human3.6M dataset by approximately $12.2%$ and helps our network to recover temporally consistent 3D poses over a sequence of images even when the 2D pose detector fails.

This paper proposes a segment-free method for geometric rectification of a distorted document image captured by a hand-held camera. The method can recover the 3D page shape by exploiting the intrinsic vector fields of the image. Based on the assumption that the curled page shape is a general cylindrical surface, we estimate the parameters related to the camera and the 3D shape model through weighted majority voting on the vector fields. Then the spatial directrix of the surface is recovered by solving an ordinary differential equation (ODE) through the Euler method. Finally, the geometric distortions in images can be rectified by flattening the estimated 3D page surface onto a plane. Our method can exploit diverse types of visual cues available in a distorted document image to estimate its vector fields for 3D page shape recovery. In comparison to the state-of-the-art methods, the great advantage is that it is a segment-free method and does not have to extract curved text lines or textual blocks, which is still a very challenging problem especially for a distorted document image. Our method can therefore be freely applied to document images with extremely complicated page layouts and severe image quality degradation. Extensive experiments are implemented to demonstrate the effectiveness of the proposed method.

In complex inferential tasks like question answering, machine learning models must confront two challenges: the need to implement a compositional reasoning process, and, in many applications, the need for this reasoning process to be interpretable to assist users in both development and prediction. Existing models designed to produce interpretable traces of their decision-making process typically require these traces to be supervised at training time. In this paper, we present a novel neural modular approach that performs compositional reasoning by automatically inducing a desired sub-task decomposition without relying on strong supervision. Our model allows linking different reasoning tasks though shared modules that handle common routines across tasks. Experiments show that the model is more interpretable to human evaluators compared to other state-of-the-art models: users can better understand the model's underlying reasoning procedure and predict when it will succeed or fail based on observing its intermediate outputs.

We introduce a new method for interpreting computer vision models: visually perceptible, decision-boundary crossing transformations. Our goal is to answer a simple question: why did a model classify an image as being of class A instead of class B? Existing approaches to model interpretation, including saliency and explanation-by-nearest neighbor, fail to visually illustrate examples of transformations required for a specific input to alter a model's prediction. On the other hand, algorithms for creating decision-boundary crossing transformations (e.g., adversarial examples) produce differences that are visually imperceptible and do not enable insightful explanation. To address this we introduce ExplainGAN, a generative model that produces visually perceptible decision-boundary crossing transformations. These transformations provide high-level conceptual insights which illustrate how a model makes decisions. We validate our model using both traditional quantitative interpretation metrics and introduce a new validation scheme for our approach and generative models more generally.

Modern semantic segmentation frameworks usually combine low-level and high-level features from pre-trained backbone convolutional models to boost performance. In this paper, we first point out that a simple fusion of low-level and high-level features could be less effective because of the gap in semantic levels and spatial resolution. We find that introducing semantic information into low-level features and high-resolution details into high-level features are more effective for the later fusion. Based on this observation, we propose a new framework, named ExFuse, to bridge the gap between low-level and high-level features thus significantly improve the segmentation quality by 4.0% in total. Furthermore, we evaluate our approach on the challenging PASCAL VOC 2012 segmentation benchmark and achieve 87.9% mean IoU, which outperforms the previous state-of-the-art results.

Inspired by the pioneering work of information bottleneck principle for Deep Neural Networks (DNNs) analysis, we design an information plane based framework to evaluate the capability of DNNs for image classification tasks, which not only helps understand the capability of DNNs, but also helps us choose a neural network which leads to higher classification accuracy more efficiently. Further, with experiments, the relationship among the model accuracy, I(X;T) and I(T;Y) are analyzed, where I(X;T) and I(T;Y) are the mutual information of DNN's output T with input X and label Y. We also show the information plane is more informative than loss curve and apply mutual information to infer the model's capability for recognizing objects of each class. Our studies would facilitate a better understanding of DNNs.

While in supervised learning, the validation error is an unbiased estimator of the generalization (test) error and complexity-based generalization bounds are abundant, no such bounds exist for learning a mapping in an unsupervised way. As a result, when training GANs and specifically when using GANs for learning to map between domains in a completely unsupervised way, one is forced to select the hyperparameters and the stopping epoch by subjectively examining multiple options. We propose a novel bound for predicting the success of unsupervised cross domain mapping methods, which is motivated by the recently proposed simplicity hypothesis. The bound can be applied both in expectation, for comparing hyperparameters, or per sample, in order to predict the success of a specific cross-domain translation. The utility of the bound is demonstrated in an extensive set of experiments employing multiple recent algorithms. Our code is available at https://github.com/sagiebenaim/gan_bound.

We present a deep model that can accurately produce dense depth maps given an RGB image with known depth at a very sparse set of pixels. The model works *simultaneously* for both indoor/outdoor scenes and produces state-of-the-art dense depth maps at nearly real-time speeds on both the NYUv2 and KITTI datasets. We surpass the state-of-the-art for monocular depth estimation even with depth values for only 1 out of every ~10000 image pixels, and we outperform other sparse-to-dense depth methods at all sparsity levels. With depth values for 1/256 of the image pixels, we achieve a mean error of less than 1% of actual depth on indoor scenes, comparable to the performance of consumer-grade depth sensor hardware. Our experiments demonstrate that it would indeed be possible to efficiently transform sparse depth measurements obtained using e.g. lower-power depth sensors or SLAM systems into high-quality dense depth maps.

Generative adversarial networks (GANs) often suffer from unpredictable mode-collapsing during training. We study the issue of mode collapse of Boundary Equilibrium Generative Adversarial Network (BEGAN), which is one of the state-of-the-art generative models. Despite its potential of generating high-quality images, we find that BEGAN tends to collapse at some modes after a period of training. We propose a new model, called BEGAN with a Constrained Space (BEGAN-CS), which includes a latent-space constraint in the loss function. We show that BEGAN-CS can significantly improve training stability and suppress mode collapse without either increasing the model complexity or degrading the image quality. Further, we visualize the distribution of latent vectors to elucidate the effect of latent-space constraint. The experimental results show that our method has additional advantages of being able to train on small datasets and to generate images similar to a given real image but with variations of designated attributes on-the-fly.

Limited angular resolution has become the main bottleneck of microlens-based plenoptic cameras towards practical vision applications. Existing view synthesis methods mainly break the task into two steps, i.e. depth estimating and view warping, which are usually inefficient and produce artifacts over depth ambiguities. In this paper, an end-to-end deep learning framework is proposed to solve these problems by exploring Pseudo 4DCNN. Specifically, 2D strided convolutions operated on stacked EPIs and detail-restoration 3D CNNs connected with angular conversion are assembled to build the Pseudo 4DCNN. The key advantage is to efficiently synthesize dense 4D light fields from a sparse set of input views. The learning framework is well formulated as an entirely trainable problem, and all the weights can be recursively updated with standard backpropagation. The proposed framework is compared with state-of-the-art approaches on both genuine and synthetic light field databases, which achieves significant improvements of both image quality (+2dB higher) and computational efficiency (over 10X faster). Furthermore, the proposed framework shows good performances in real-world applications such as biometrics and depth estimation.

For human drivers, having rear and side-view mirrors is vital for safe driving. They deliver a more complete view of what is happening around the car. Human drivers also heavily exploit their mental map for navigation. Nonetheless, several methods have been published that learn driving models with only a front-facing camera and without a route planner. This lack of information renders the self-driving task quite intractable. We investigate the problem in a more realistic setting, which consists of a surround-view camera system with eight cameras, a route planner, and a CAN bus reader. In particular, we develop a sensor setup that provides data for a 360-degree view of the area surrounding the vehicle, the driving route to the destination, and low-level driving maneuvers (e.g. steering angle and speed) by human drivers. With such a sensor setup we collect a new driving dataset, covering diverse driving scenarios and varying weather/illumination conditions. Finally, we learn a novel driving model by integrating information from the surround-view cameras and the route planner. Two route planners are exploited: 1) by representing the planned routes on OpenStreetMap as a stack of GPS coordinates, and 2) by rendering the planned routes on TomTom Go Mobile and recording the progression into a video. Our experiments show that: 1) 360-degree surround-view cameras help avoid failures made with a single front-view camera, in particular for city driving and intersection scenarios; and 2) route planners help the driving task significantly, especially for steering angle prediction.