ECCV 2018 Papers

Recent CNN based object detectors, either one-stage methods like YOLO, SSD, and RetinaNet, or two-stage detectors like Faster R-CNN, R-FCN and FPN, are usually trying to directly finetune from ImageNet pre-trained models designed for the task of image classification. However, there has been little work discussing the backbone feature extractor specifically designed for the task of object detection. More importantly, there are several differences between the tasks of image classification and object detection. (1) Recent object detectors like FPN and RetinaNet usually involve extra stages against the task of image classification to handle the objects with various scales. (2) Object detection not only needs to recognize the category of the object instances but also spatially locate them. Large downsampling factors bring large valid receptive field, which is good for image classification but compromises the object location ability. Due to the gap between the image classification and object detection, we propose DetNet in this paper, which is a novel backbone network specifically designed for object detection. Moreover, DetNet includes the extra stages against traditional backbone network for image classification, while maintains high spatial resolution in deeper layers. Without any bells and whistles, state-of-the-art results have been obtained for both object detection and instance segmentation on the MSCOCO benchmark based on our DetNet~(4.8G FLOPs) backbone. Codes will be released.

Robust cost optimization is the challenging task of fitting a large number of parameters to data points containing a significant and unknown fraction of outliers. In this work we identify three classes of deterministic second-order algorithms that are able to tackle this type of optimization problem: direct approaches that aim to optimize the robust cost directly with a second order method, lifting-based approaches that add so called lifting variables to embed the given robust cost function into a higher dimensional space, and graduated optimization methods that solve a sequence of smoothed cost functions. We study each of these classes of algorithms and propose improvements either to reduce their computational time or to make them find better local minima. Finally, we experimentally demonstrate the superiority of our improved graduated optimization method over the state of the art algorithms both on synthetic and real data for four different problems.

Convolutional neural networks (CNN) are limited by the lack of capability to handle geometric information due to the fixed grid kernel structure. The availability of depth data enables progress in RGB-D semantic segmentation with CNNs. State-of-the-art methods either use depth as additional images or process spatial information in 3D volumes or point clouds. These methods suffer from high computation and memory cost. To address these issues, we present Depth-aware CNN by introducing two intuitive, flexible and effective operations: depth-aware convolution and depth-aware average pooling. By leveraging depth similarity between pixels in the process of information propagation, geometry is seamlessly incorporated into CNN. Without introducing any additional parameters, both operators can be easily integrated into existing CNNs. Extensive experiments and ablation studies on challenging RGB-D semantic segmentation benchmarks validate the effectiveness and flexibility of our approach.

Depth estimation from a single image is a fundamental problem in computer vision. In this paper, we propose a simple yet effective convolutional spatial propagation network (CSPN) to learn the affinity matrix for depth prediction. Specifically, we adopt an efficient linear propagation model, where the propagation is performed with a manner of recurrent convolutional operation, and the affinity among neighboring pixels is learned through a deep convolutional neural network (CNN). We apply the designed CSPN to two depth estimation tasks given a single image: (1) To refine the depth output from state-of-the-art (SOTA) existing methods; and (2) to convert sparse depth samples to a dense depth map by embedding the depth samples within the propagation procedure. The second task is inspired by the availability of LIDARs that provides sparse but accurate depth measurements. We experimented the proposed CSPN over two popular benchmarks for depth estimation, ie NYU v2 and KITTI, where we show that our proposed approach improves in not only quality (e.g., 30% more reduction in depth error), but also speed (e.g., 2 to 5$ imes$ faster) than prior SOTA methods.

This paper targets the problem of image set-based face verification and identification. Unlike traditional single media (an image or video) setting, we encounter a set of heterogeneous contents containing orderless images and videos. The importance of each image is usually considered either equal or based on their independent quality assessment. How to model the relationship of orderless images within a set remains a challenge. We address this problem by formulating it as a Markov Decision Process (MDP) at the feature level. Specifically, we propose a dependency-aware attention control (DAC) network, which resorts to actor-critic reinforcement learning for sequential attention decision of each image embedding to fully exploit the rich correlation cues among the unordered images. Moreover, its sample-efficient variant with off-policy experience replay is introduced to speed up the learning process. The pose-guided representation scheme can further boost the performance at the extremes of the pose variation. We show that our method leads to the state-of-the-art accuracy on IJB-A dataset and also generalizes well in several video-based face recognition tasks, extit{e.g.}, YTF and Celebrity-1000.

Many previous literatures use landmarks to guide the cor- respondence of 3D faces. However, these landmarks, either manually or automatically annotated, are hard to define consistently across differ- ent faces in many circumstances. We propose a general framework for dense correspondence of 3D faces without landmarks in this paper. The dense correspondence goal is revisited in two perspectives: semantic and topological correspondence. Starting from a template facial mesh, we sequentially perform global alignment, primary correspondence by tem- plate warping, and contextual mesh refinement, to reach the final cor- respondence result. The semantic correspondence is achieved by a local iterative closest point (ICP) algorithm of kernelized version, allowing accurate matching of local features. Then, robust deformation from the template to the target face is formulated as a minimization problem. Fur- thermore, this problem leads to a well-posed sparse linear system such that the solution is unique and efficient. Finally, a contextual mesh re- fining algorithm is applied to ensure topological correspondence. In the experiment, the proposed method is evaluated both qualitatively and quantitatively on two datasets including a publicly available FRGC v2.0 dataset, demonstrating reasonable and reliable correspondence results.

In this work we integrate ideas from surface-based modeling with neural synthesis: we propose a combination of surface-based pose estimation and deep generative models that allows us to perform accurate pose transfer, i.e. synthesize a new image of a person based on a single image of that person and the image of a pose donor. We use a dense pose estimation system that maps pixels from both images to a common surface-based coordinate system, allowing the two images to be brought in correspondence with each other. We inpaint and refine the source image intensities in the surface coordinate system, prior to warping them onto the target pose. These predictions are fused with those of a convolutional predictive module through a neural synthesis module allowing for training the whole pipeline jointly end-to-end, optimizing a combination of adversarial and perceptual losses. We show that dense pose estimation is a substantially more powerful conditioning input than landmark-, or mask-based alternatives, and report systematic improvements over state of the art generators on DeepFashion and MVC datasets.

In this work we introduce the Deforming Autoencoder, a generative model for images that disentangles shape from appearance in a latent representation space that is learned in a fully unsupervised manner. As in the deformable template paradigm, shape is represented as a diffeomorphism between a canonical coordinate system (`template') and an observed image, while appearance is modeled in template coordinates, thus discarding variability due to deformations. We introduce novel techniques that allow this approach to be deployed in the setting of autoencoders and show that this method can be used for unsupervised group-wise image alignment. We show experiments with expression morphing in humans, hands, and digits, face manipulation, such as shape and appearance interpolation, as well as unsupervised landmark localization. A more powerful form of unsupervised disentangling becomes possible in template coordinates, allowing us to successfully decompose face images into shading and albedo, and further manipulate face images.

The representation of 3D pose plays a critical role for 3D body action and hand gesture recognition. Rather than directly representing the 3D pose using its joint locations, in this paper, we propose Deformable Pose Traversal Convolution which applies one-dimensional convolution to traverse the 3D pose to represent it. Instead of fixing the reception field when performing traversal convolution, it optimizes the convolutional kernel for each joint, by considering contextual joints with various weights. This deformable convolution can better utilize contextual joints for action and gesture recognition and is more robust to noisy joints. Moreover, by feeding the learned pose feature to a LSTM, we can perform end-to-end training which jointly optimizes 3D pose representation and temporal sequence recognition. Experiments on three benchmark datasets validate the competitive performance of our proposed method, as well as its efficiency and robustness to handle noisy pose.

Compositional models represent patterns with hierarchies of meaningful parts and subparts. Their ability to characterize high-order relationships among body parts helps resolve low-level ambiguities in human pose estimation (HPE). However, prior compositional models make unrealistic assumptions on subpart-part relationships, making them incapable to characterize complex compositional patterns. Moreover, state spaces of their higher-level parts can be exponentially large, complicating both inference and learning. To address these issues, this paper introduces a novel framework, termed as Deeply Learned Compositional Model (DLCM), for HPE. It exploits deep neural networks to learn the compositionality of human bodies. This results in a network with a hierarchical compositional architecture and bottom-up/top-down inference stages. In addition, we propose a novel bone-based part representation. It not only compactly encodes orientations, scales and shapes of parts, but also avoids their potentially large state spaces. With significantly lower complexities, our approach outperforms state-of-the-art methods on three benchmark datasets.

We present a novel method to generate accurate and realistic clothing deformation from real data capture. Previous methods for realistic cloth modeling mainly rely on intensive computation of physics-based simulation (with numerous heuristic parameters), while models reconstructed from visual observations typically suffer from lack of geometric details. Here, we propose an original framework consisting of two modules that work jointly to represent global shape deformation as well as surface details with high fidelity. Global shape deformations are recovered from a subspace model learned from 3D data of clothed people in motion, while high frequency details are added to normal maps created using a conditional Generative Adversarial Network whose architecture is designed to enforce realism and temporal consistency. This leads to unprecedented high-quality rendering of clothing deformation sequences, where fine wrinkles from (real) high resolution observations can be recovered. In addition, as the model is learned independently from body shape and pose, the framework is suitable for applications that require retargeting (e.g., body animation). Our experiments show original high quality results with a flexible model. We claim an entirely data-driven approach to realistic cloth wrinkle generation is possible.

In this paper, we propose a novel deep learning based video saliency prediction method, named DeepVS. Specifically, we establish a large-scale eye-tracking database of videos (LEDOV), which includes 32 subjects' fixations on 538 videos. We find from LEDOV that human attention is more likely to be attracted by objects, particularly the moving objects or the moving parts of objects. Hence, an object-to-motion convolutional neural network (OM-CNN) is developed to predict the intra-frame saliency for DeepVS, which is composed of the objectness and motion subnets. In OM-CNN, cross-net mask and hierarchical feature normalization are proposed to combine the spatial features of the objectness subnet and the temporal features of the motion subnet. We further find from our database that there exists a temporal correlation of human attention with a smooth saliency transition across video frames. We thus propose saliency-structured convolutional long short-term memory (SS-ConvLSTM) network, using the extracted features from OM-CNN as the input. Consequently, the inter-frame saliency maps of a video can be generated, which consider both structured output with center-bias and cross-frame transitions of human attention maps. Finally, the experimental results show that DeepVS advances the state-of-the-art in video saliency prediction.

We present a system for keyframe-based dense camera tracking and depth map estimation that is entirely learned. For tracking, we estimate small pose increments between the current camera image and a synthetic viewpoint. This significantly simplifies the learning problem and alleviates the dataset bias for camera motions. Further, we show that generating a large number of pose hypotheses leads to more accurate predictions. For mapping, we accumulate information in a cost volume centered at the current depth estimate. The mapping network then combines the cost volume and the keyframe image to update the depth prediction, thereby effectively making use of depth measurements and image-based priors. Our approach yields state-of-the-art results with few images and is robust with respect to noisy camera poses. We demonstrate that the performance of our 6~DOF tracking competes with RGB-D tracking algorithms.We compare favorably against strong classic and deep learning powered dense depth algorithms.

Non-contact video-based physiological measurement has many applications in health care and human-computer interaction. Practical applications require measurements to be accurate even in the presence of large head rotations. We propose the first end-to-end system for video-based measurement of heart and breathing rate using a deep convolutional network. The system features a new motion representation based on a skin reflection model and a new attention mechanism using appearance information to guide motion estimation, both of which enable robust measurement under heterogeneous lighting and major motions. Our approach significantly outperforms all current state-of-the-art methods on both RGB and infrared video datasets. Furthermore, it allows spatial-temporal distributions of physiological signals to be visualized via the attention mechanism.

As a second-order pooled representation, covariance matrix has attracted much attention in visual recognition, and some pioneering works have recently integrated it into deep learning framework to jointly learn this matrix for fine-grained image recognition. A recent study shows that kernel matrix works considerably better than covariance matrix for this kind of representation, by modeling the higher-order, nonlinear relationship among pooled visual descriptors. Nevertheless, in that study neither the descriptors nor the kernel matrix is deeply learned. Worse, they are considered separately, hindering the pursuit of an optimal representation. To improve this situation, this work designs a deep network that jointly learns local descriptors and kernel-matrix-based pooled representation in an end-to-end manner. The derivatives for the mapping from a local descriptor set to this representation are derived to carry out backpropagation. More importantly, we introduce the {Daleckiv{i}-Krev{i}n formula} from Operator theory to give a concise and unified result on differentiating general functions defined on symmetric positive-definite (SPD) matrix, which shows its better numerical stability in conducting backpropagation compared with the existing method when handling the Riemannian geometry of SPD matrix. Experiments on multiple fine-grained image benchmark datasets not only show the superiority of kernel-matrix-based SPD representation with deep local descriptors, but also verify the advantage of the proposed deep network in pursuing better SPD representations. Also, ablation study is provided to explain why and from where these improvements are attained.

In computer vision, one is often confronted with problems of domain shifts, which occur when one applies a classifier trained on a source dataset to target data sharing similar characteristics (e.g. same classes), but also different latent data structures (e.g. different acquisition conditions). In such a situation, the model will perform poorly on the new data, since the classifier is specialized to recognize visual cues specific to the source domain. In this work we explore a solution, named DeepJDOT, to tackle this problem: through a measure of discrepancy on joint deep representations/labels based on optimal transport, we not only learn new data representations aligned between the source and target domain, but also simultaneously preserve the discriminative information used by the classifier. We applied DeepJDOT to a series of visual recognition tasks, where it compares favorably against state-of-the-art deep domain adaptation methods.

Estimating the 6D pose of objects from images is an important problem in various applications such as robot manipulation and virtual reality. While direct regression of images to object poses has limited accuracy, matching rendered images of an object against the input image can produce accurate results. In this work, we propose a novel deep neural network for 6D pose matching named DeepIM. Given an initial pose estimation, our network is able to iteratively refine the pose by matching the rendered image with the input image. The network is trained to predict a relative pose transformation using an untangled representation of 3D location and 3D orientation and an iterative training process. Experiments on two commonly used benchmarks for 6D pose estimation demonstrate that DeepIM achieves large improvements over state-of-the-art methods. We furthermore show that DeepIM is able to match previously unseen objects.

In this paper we address the problem of how to robustly train a ConvNet for regression, or deep robust regression. Traditionally, deep regression employ the L2 loss function, known to be sensitive to outliers, i.e. samples that either lie at an abnormal distance away from the majority of the training samples, or that correspond to wrongly annotated targets. This means that, during back-propagation, outliers may bias the training process due to the high magnitude of their gradient. In this paper, we propose DeepGUM: a deep regression model that is robust to outliers thanks to the use of a Gaussian-uniform mixture model. We derive an optimization algorithm that alternates between the unsupervised detection of outliers using expectation-maximization, and the supervised training with cleaned samples using stochastic gradient descent. DeepGUM is able to adapt to a continuously evolving outlier distribution, avoiding to manually impose any threshold on the proportion of outliers in the training set. Extensive experimental evaluations on four different tasks (facial and fashion landmark detection, age and head pose estimation) lead us to conclude that our novel robust technique provides reliability in the presence of various types of noise and protection against a high percentage of outliers.

We present a deep learning-based volumetric capture approach for performance capture using a passive and highly sparse multi-view capture system. We focus on a template-free, per-frame 3D surface reconstruction from as few as three RGB sensors, where conventional visual hull or multi-view stereo methods would fail. State-of-the-art performance capture systems require either pre-scanned actors, large number of cameras or active sensors. We introduce a novel multi-view Convolutional Neural Network (CNN) that maps 2D images to a 3D volumetric field that encodes the probabilistic distribution of surface points of the captured subject. By querying the resulting field, we can instantiate the clothed human body at arbitrary resolutions. Our approach also scales to different numbers of input images, which yield increased reconstruction quality when more views are used. Though only trained on synthetic data, our network can generalize to real captured performances. Since high-quality temporal surface reconstructions are possible, our method is suitable for low-cost full body volumetric capture solutions for consumers, which are gaining popularity for VR and AR content creation. Experimental results demonstrate that our method is significantly more robust and accurate than existing techniques where only very sparse views are available.