ECCV 2018 Papers

We present an adversarial network for rendering photorealistic hair as an alternative to conventional computer graphics pipelines. Our deep learning approach does not require low-level parameter tuning nor ad-hoc asset design. Our method simply takes a strand-based 3D hair model as input and provides intuitive user-control for color and lighting through reference images. To handle the diversity of hairstyles and its appearance complexity, we disentangle hair structure, color, and illumination properties using a sequential GAN architecture and a semi-supervised training approach. We also introduce an intermediate edge activation map to orientation field conversion step to ensure a successful CG-to-photoreal transition, while preserving the hair structures of the original input data. As we only require a feed-forward pass through the network, our rendering performs in real-time. We demonstrate the synthesis of photorealistic hair images on a wide range of intricate hairstyles and compare our technique with state-of-the-art hair rendering methods.

In this work, we consider the problem of robust gaze estimation in natural environments. Large camera-to-subject distances and high variations in head pose and eye gaze angles are common in such environments. This leads to two main shortfalls in state-of-the-art methods for gaze estimation: hindered ground truth gaze annotation and diminished gaze estimation accuracy as image resolution decreases with distance. We first record a novel dataset of varied gaze and head pose images in a natural environment, addressing the issue of ground truth annotation by measuring head pose using a motion capture system and eye gaze using mobile eyetracking glasses. We apply semantic image inpainting to the area covered by the glasses to bridge the gap between training and testing images by removing the obtrusiveness of the glasses. We also present a new real-time algorithm involving appearance-based deep convolutional neural networks with increased capacity to cope with the diverse images in the new dataset. Experiments with this network architecture are conducted on a number of diverse eye-gaze datasets including our own, and in cross dataset evaluations. We demonstrate state-of-the-art performance in terms of estimation accuracy in all experiments, and the architecture performs well even on lower resolution images.

This paper proposes a novel data-driven approach for inertial navigation, which learns to estimate trajectories of natural human motions just from an inertial measurement unit (IMU) in every smartphone. The key observation is that human motions are repetitive and consist of a few major modes (e.g., standing, walking, or turning). Our algorithm regresses a velocity vector from the history of linear accelerations and angular velocities, then corrects low-frequency bias in the linear accelerations, which are integrated twice to estimate positions. We have acquired training data with ground-truth motions across multiple human subjects and multiple phone placements (e.g., in a bag or a hand-held). The qualitatively and quantitatively evaluations have demonstrated that our algorithm has surprisingly shown comparable results to full Visual Inertial navigation. To our knowledge, this paper is the first to integrate sophisticated machine learning techniques with inertial navigation, potentially opening up a new line of research in the domain of data-driven inertial navigation. We will publicly share our code and data to facilitate further research.

While large datasets have proven to be a key enabler for progress in computer vision, they can have biases that lead to erroneous conclusions. The notion of the representation bias of a dataset is proposed to combat this problem. It captures the fact that representations other than the ground-truth representation can achieve good performance on any given dataset. When this is the case, the dataset is said not to be well calibrated. Dataset calibration is shown to be a necessary condition for the standard state-of-the-art evaluation practice to converge to the ground-truth representation. A procedure, RESOUND, is proposed to quantify and minimize representation bias. Its application to the problem of action recognition shows that current datasets are biased towards static representations (objects, scenes and people). Two versions of RESOUND are studied. An Explicit RESOUND procedure is proposed to assemble new datasets by sampling existing datasets. An implicit RESOUND procedure is used to guide the creation of a new dataset, Diving48, of over 18,000 video clips of competitive diving actions, spanning 48 fine-grained dive classes. Experimental evaluation confirms the effectiveness of RESOUND to reduce the static biases of current datasets.

We aim to search for a target person from a gallery of whole scene images for which the annotations of pedestrian bounding boxes are unavailable. Previous approaches to this problem have relied on a pedestrian proposal net, which may generate redundant proposals and increase the computational burden. In this paper, we address this problem by training relational context-aware agents which learn the actions to localize the target person from the gallery of whole scene images. We incorporate the relational spatial and temporal contexts into the framework. Specifically, we propose to use the target person as the query in the query-dependent relational network. The agent determines the best action to take at each time step by simultaneously considering the local visual information, the relational and temporal contexts, together with the target person. To validate the performance of our approach, we conduct extensive experiments on the large-scale Person Search benchmark dataset and achieve significant improvements over the compared approaches. It is also worth noting that the proposed model even performs better than traditional methods with perfect pedestrian detectors.

We propose a method to forecast a vehicle's ego-motion as a distribution over spatiotemporal paths, conditioned on features (e.g., from LIDAR and images) embedded in an overhead map. The method learns a policy inducing a distribution over simulated trajectories that is both diverse (produces most paths likely under the data) and precise (mostly produces paths likely under the data). This balance is achieved through minimization of a symmetrized cross-entropy between the distribution and demonstration data. By viewing the simulated-outcome distribution as the pushforward of a simple distribution under a simulation operator, we obtain expressions for the cross-entropy metrics that can be efficiently evaluated and differentiated, enabling stochastic-gradient optimization. We propose concrete policy architectures for this model, discuss our evaluation metrics relative to previously-used metrics, and demonstrate the superiority of our method relative to state-of-the-art methods in both the KITTI dataset and a similar but novel and larger real-world dataset explicitly designed for the vehicle forecasting domain.

In this paper, we propose a novel Question-Guided Hybrid Convolution (QGHC) network for Visual Question Answering (VQA). Most state-of-the-art VQA methods fuse the high-level textual and visual features from the neural network and abandon the visual spatial information when learning multi-modal features. To address these problems, question-guided kernels generated from the input question are designed to convolute with visual features for capturing the textual and visual relationship in the early stage. The question-guided convolution can tightly couple the textual and visual information but also introduces more parameters when learning kernels. We apply the group convolution which consists of question-independent kernels and question-dependent kernels, to reduce the parameter size and release the over-fitting. The hybrid convolution can generate discriminative multi-modal features with fewer parameters. Our proposed approach is also complementary to existing bilinear pooling fusion methods and attention methods. Integration with them could further boost the performance. Extensive experiments on public VQA datasets validate the effectiveness of QGHC.

Visual Question Answering (VQA) requires integration of feature maps with drastically different structures and focus of the correct regions. Image descriptors have structures at multiple spatial scales, while lexical inputs inherently follow a temporal sequence and naturally cluster into semantically different question types. A lot of previous works use complex models to extract feature representations but neglect to use high-level information summary such as question types in learning . In this work, we propose Question Type-guided Attention (QTA). It utilizes the information of question type to dynamically balance between bottom-up and top-down visual features, respectively extracted from ResNet and Faster R-CNN networks. We experiment with multiple VQA architectures with extensive input ablation studies over the TDIUC dataset and show that QTA systematically improves the performance by more than 5% across multiple question type categories such as “Activity Recognition”, “Utility” and “Counting” on TDIUC dataset compared to the state-of-art. By adding QTA on the state-of-art model MCB, we achieve 3% improvement for overall accuracy. Finally, we propose a multitask extension to predict question types which generalizes QTA to applications that lack of question type, with minimal performance loss.

Neural networks in the real domain have been studied for a long time and achieved promising results in many vision tasks for recent years. However, the extensions of the neural network models in other number fields and their potential applications are not fully-investigated yet. Focusing on color images, which can be naturally represented as quaternion matrices, we propose a quaternion convolutional neural network (QCNN) model to obtain more representative features. In particular, we re-design the basic modules like convolution layer and fully-connected layer in the quaternion domain, which can be used to establish fully-quaternion convolutional neural networks. Moreover, these modules are compatible with almost all deep learning techniques and can be plugged into traditional CNNs easily. We test our QCNN models in both color image classification and denoising tasks. Experimental results show that they outperform the real-valued CNNs with same structures.

In this paper, we propose quantized densely connected U-Nets for efficient visual landmark localization. The idea is that features of the same semantic meanings are globally reused across the stacked U-Nets. This dense connectivity largely improves the information flow, yielding improved localization accuracy. However, a vanilla dense design would suffer from critical efficiency issue in both training and testing. To solve this problem, we first propose order-K dense connectivity to trim off long-distance shortcuts; then, we use a memory-efficient implementation to significantly boost the training efficiency and investigate an iterative refinement that may slice the model size in half. Finally, to reduce the memory consumption and high precision operations both in training and testing, we further quantize weights, inputs, and gradients of our localization network to low bit-width numbers. We validate our approach in two tasks: human pose estimation and face alignment. The results show that our approach achieves state-of-the-art localization accuracy, but using ~70% fewer parameters, ~98% less model size and saving ~32x training memory compared with other benchmark localizers.

In this paper, we propose a simple and general framework for training very tiny CNNs for object detection. Due to limited representation ability, it is challenging to train very tiny networks for complicated tasks like detection. To the best of our knowledge, our method, called Quantization Mimic, is the first one focusing on very tiny networks. We utilize two types of acceleration methods: mimic and quantization. Mimic improves the performance of a student network by transfering knowledge from a teacher network. Quantization converts a full-precision network to a quantized one without large degradation of performance. If the teacher network is quantized, the search scope of the student network will be smaller. Using this feature of the quantization, we propose Quantization Mimic. It first quantizes the large network, then mimic a quantized small network. The quantization operation can help student network to better match the feature maps from teacher network. To evaluate our approach, we carry out experiments on various popular CNNs including VGG and Resnet, as well as different detection frameworks including Faster R-CNN and R-FCN. Experiments on Pascal VOC and WIDER FACE verify that our Quantization Mimic algorithm can be applied on various settings and outperforms state-of-the-art model acceleration methods given limited computing resouces.

This paper presents a Quadtree Convolutional Neural Network (QCNN) for efficiently learning from image datasets representing sparse data such as handwriting, pen strokes, freehand sketches, etc. Instead of storing the sparse sketches in regular dense tensors, our method decomposes and represents the image as a linear quadtree that is only refined in the non-empty portions of the image. The actual image data corresponding to non-zero pixels is stored in the finest nodes of the quadtree. Convolution and pooling operations are restricted to the sparse pixels, leading to better efficiency in computation time as well as memory usage. Specifically, the computational and memory costs in QCNN grow linearly in the number of non-zero pixels, as opposed to traditional CNNs where the costs are quadratic in the number of pixels. This enables QCNN to learn from sparse images much faster and process high resolution images without the memory constraints faced by traditional CNNs. We study QCNN on four sparse image datasets for classification and sketch simplification tasks. The results show that QCNN can obtain comparable accuracy with large reduction in computational and memory costs.

Face detection has been well studied for many years and one of remaining challenges is to detect small, blurred and partially occluded faces in uncontrolled environment. This paper proposes a novel context-assisted single shot face detector, named emph{PyramidBox} to handle the hard face detection problem. Observing the importance of the context, we improve the utilization of contextual information in the following three aspects. First, we design a novel context anchor to supervise high-level contextual feature learning by a semi-supervised method, which we call it PyramidAnchors. Second, we propose the Low-level Feature Pyramid Network to combine adequate high-level context semantic feature and Low-level facial feature together, which also allows the PyramidBox to predict faces of all scales in a single shot. Third, we introduce a context-sensitive structure to increase the capacity of prediction network to improve the final accuracy of output. In addition, we use the method of Data-anchor-sampling to augment the training samples across different scales, which increases the diversity of training data for smaller faces. By exploiting the value of context, PyramidBox achieves superior performance among the state-of-the-art over the two common face detection benchmarks, FDDB and WIDER FACE. Our code is available in PaddlePaddle: href{https://github.com/PaddlePaddle/models/tree/develop/fluid/face_detection}{url{https://github.com/PaddlePaddle/models/tree/develop/fluid/face_detection}}.

This paper proposes a fast video salient object detection model, based on a novel recurrent network architecture, named Pyramid Dilated Bidirectional ConvLSTM (PDB-ConvLSTM). A Pyramid Dilated Convolution (PDC) module is first designed for simultaneously extracting spatial features at multiple scales. These spatial features are then concatenated and fed into an extended Deeper Bidirectional ConvLSTM (DB-ConvLSTM) to learn spatiotemporal information. Forward and backward ConvLSTM units are placed in two layers and connected in a cascaded way, encouraging information flow between the bi-directional streams and leading to deeper feature extraction. We further augment DB-ConvLSTM with a PDC-like structure, by adopting several dilated DB-ConvLSTMs to extract multi-scale spatiotemporal information. Extensive experimental results show that our method outperforms previous video saliency models in a large margin, with a real-time speed of 20 fps on a single GPU. With unsupervised video object segmentation as an example application, the proposed model (with a CRF-based post-process) achieves state-of-the-art results on two popular benchmarks, well demonstrating its superior performance and high applicability.

We propose a new multiplanar superstructure for unified real-time processing of RGB-D data. Modern RGB-D sensors are widely used for indoor 3D capture, with applications ranging from modeling to robotics, through augmented reality. Nevertheless, their use is limited by their low resolution, with frames often corrupted with noise, missing data and temporal inconsistencies. Our approach, named Proxy Clouds, consists in generating and updating through time a single set of compact local statistics parameterized over detected planar proxies, which are fed from raw RGB-D data. Proxy Clouds provide several processing primitives, which improve the quality of the RGB-D stream on-the-fly or lighten further operations. Experimental results confirm that our light weight analysis framework copes well with embedded execution as well as moderate memory and computational capabilities compared to state-of-the-art methods. Processing of RGB-D data with Proxy Clouds includes noise and temporal flickering removal, hole filling and resampling. As a substitute of the observed scene, our proxy cloud can additionally be applied to compression and scene reconstruction. We present experiments performed with our framework in indoor scenes of different natures within a recent open RGB-D dataset.

Photos taken in hazy weather are usually covered with white masks and often lose important details. In this paper, we propose a novel deep learning approach for single image dehazing by learning dark channel and transmission priors. First, we build an energy model for dehazing using dark channel and transmission priors and design an iterative optimization algorithm using proximal operators for these two priors. Second, we unfold the iterative algorithm to be a deep network, dubbed as extit{proximal dehaze-net}, by learning the proximal operators using convolutional neural networks. Our network combines the advantages of traditional prior-based dehazing methods and deep learning methods by incorporating haze-related prior learning into deep network. Experiments show that our method achieves state-of-the-art performance for single image dehazing.

We present a novel 3D pose estimation method based on joint interdependency (JI) for acquiring 3D joints from the human pose of an RGB image. The JI incorporates the body part based structural connectivity of joints to learn the high spatial correlation of human posture on our method. Towards this goal, we propose a new long short-term memory (LSTM)-based deep learning architecture named propagating LSTM networks (p-LSTMs), where each LSTM is connected sequentially to reconstruct 3D depth from the centroid to edge joints through learning the intrinsic JI. In the first LSTM, the seed joints of 3D pose are created and reconstructed into the whole-body joints through the connected LSTMs. Utilizing the p-LSTMs, we achieve the higher accuracy of about 11.2% than state-of-the-art methods on the largest publicly available database. Importantly, we demonstrate that the JI drastically reduces the structural errors at body edges, thereby leads to a significant improvement.

Structure from Motion or the sparse 3D reconstruction out of individual photos is a long studied topic in computer vision. Yet none of the existing reconstruction pipelines fully addresses a progressive scenario where images are only getting available during the reconstruction process and intermediate results are delivered to the user. Incremental pipelines are capable of growing a 3D model but often get stuck in local minima due to wrong (binding) decisions taken based on incomplete information. Global pipelines on the other hand need the access to the complete viewgraph and are not capable of delivering intermediate results. In this paper we propose a new reconstruction pipeline working in a progressive manner rather than in a batch processing scheme. The pipeline is able to recover from failed reconstructions in early stages, avoids to take binding decisions, delivers a progressive output and yet maintains the capabilities of existing pipelines. We demonstrate and evaluate our method on diverse challenging public and dedicated datasets including those with highly symmetric structures and compare to the state of the art.

We propose a new method for learning the structure of convolutional neural networks (CNNs) that is more efficient than recent state-of-the-art methods based on reinforcement learning and evolutionary algorithms. Our approach uses a sequential model-based optimization (SMBO) strategy, in which we search for structures in order of increasing complexity, while simultaneously learning a surrogate model to guide the search through structure space. Direct comparison under the same search space shows that our method is up to 5 times more efficient than the RL method of Zoph et al. (2018) in terms of number of models evaluated, and 8 times faster in terms of total compute. The structures we discover in this way achieve state of the art classification accuracies on CIFAR-10 and ImageNet.