WACV 2020 论文列表

Despite the advances in Human Activity Recognition, the ability to exploit the dynamics of human body motion in videos has yet to be achieved. In numerous recent works, researchers have used appearance and motion as independent inputs to infer the action that is taking place in a specific video. In this paper, we highlight that while using a novel representation of human body motion, we can benefit from appearance and motion simultaneously. As a result, better performance of action recognition can be achieved. We start with a pose estimator to extract the location and heat-map of body joints in each frame. We use a dynamic encoder to generate a fixed size representation from these body joint heat-maps. Our experimental results show that training a convolutional neural network with the dynamic motion representation outperforms state-of-the-art action recognition models. By modeling distinguishable activities as distinct dynamical systems and with the help of two stream networks, we obtain the best performance on HMDB, JHMDB, UCF-101, and AVA datasets.

Dealing properly with different viewing conditions remains a key challenge for computer vision in autonomous driving. Domain adaptation has opened new possibilities for data augmentation, translating arbitrary road scene images into different environmental conditions. Although multimodal concepts have demonstrated the capability to separate content and style, we find that existing methods fail to reproduce scenes in the exact appearance given by a reference image. In this paper, we address the aforementioned problem by introducing a style alignment loss between output and reference image. We integrate this concept into a multimodal unsupervised image-to-image translation model with a novel dual-mode training process and additional adversarial losses. Focusing on road scene images, we evaluate our model in various aspects including visual quality and feature matching. Our experiments reveal that we are able to significantly improve both style alignment and image quality in different viewing conditions. Adapting concepts from neural style transfer, our new training approach allows to control the output of multimodal domain adaptation, making it possible to generate arbitrary scenes and viewing conditions for data augmentation.

In video captioning, the description of a video usually relies on the domain to which the video belongs. Typically, the videos belong to wide range domains such as sports, music, news, cooking, etc. In many cases, a video can be associated with more than one domain. In this paper, we propose an approach to video captioning that uses domain-specific decoders. We build a domain classifier to obtain the estimates of probabilities of a video belonging to different domains. For each video, we identify the top-k domains based on the estimated probabilities. Each video in the training data set is shared in training the domain-specific decoders of top-k labels obtained from the domain classifier. The domain-specific decoders use the domain-specific semantic tags for generating captions. The proposed approach uses the Temporal VLAD for preprocessing the features extracted from 2D-CNN and 3D-CNN features. The preprocessed features provide better feature representation of the videos. The effectiveness of the proposed approach is demonstrated through the results of experimental studies on Microsoft Video Description (MSVD) corpus and MSR-VTT dataset.

Image-to-image translation architectures may have limited effectiveness in some circumstances. For example, while generating rainy scenarios, they may fail to model typical traits of rain as water drops, and this ultimately impacts the synthetic images realism. With our method, called domain bridge, web-crawled data are exploited to reduce the domain gap, leading to the inclusion of previously ignored elements in the generated images. We make use of a network for clear to rain translation trained with the domain bridge to extend our work to Unsupervised Domain Adaptation (UDA). In that context, we introduce an online multimodal style-sampling strategy, where image translation multimodality is exploited at training time to improve performances. Finally, a novel approach for self-supervised learning is presented, and used to further align the domains. With our contributions, we simultaneously increase the realism of the generated images, while reaching on par performances with respect to the UDA state-of-the-art, with a simpler approach.

Previous studies generally agree that face recognition accuracy is higher for older persons than for younger persons. But most previous studies were before the wave of deep learning matchers, and most considered accuracy only in terms of the verification rate for genuine pairs. This paper investigates accuracy for age groups 16-29, 30-49 and 50-70, using three modern deep CNN matchers, and considers differences in the impostor and genuine distributions as well as verification rates and ROC curves. We find thataccuracy is lower for older persons and higher for younger persons. In contrast, a pre deep learning matcher on the same dataset shows the traditional result of higher accuracy for older persons, although its overall accuracy is much lower than that of the deep learning matchers. Comparing the impostor and genuine distributions, we conclude that impostor scores have a larger effect than genuine scores in causing lower accuracy for the older age group. We also investigate the effects of training data across the age groups. Our results show that fine-tuning the deep CNN models on additional images of older persons actually lowers accuracy for the older age group. Also, we fine-tune and train from scratch two models using age-balanced training datasets, and these results also show lower accuracy for older age group. These results argue that the lower accuracy for the older age group is not due to imbalance in the original training data.

Creating plausible virtual actors from images of real actors remains one of the key challenges in computer vision and computer graphics. Marker-less human motion estimation and shape modeling from images in the wild bring this challenge to the fore. Although the recent advances on view synthesis and image-to-image translation, currently available formulations are limited to transfer solely style and do not take into account the character's motion and shape, which are by nature intermingled to produce plausible human forms. In this paper, we propose a unifying formulation for transferring appearance and retargeting human motion from monocular videos that regards all these aspects. Our method synthesizes new videos of people in a different context where they were initially recorded. Differently from recent appearance transferring methods, our approach takes into account body shape, appearance, and motion constraints. The evaluation is performed with several experiments using publicly available real videos containing hard conditions. Our method is able to transfer both human motion and appearance outperforming state-of-the-art methods, while preserving specific features of the motion that must be maintained (e.g., feet touching the floor, hands touching a particular object) and holding the best visual quality and appearance metrics such as Structural Similarity (SSIM) and Learned Perceptual Image Patch Similarity (LPIPS)

In this paper, we present a canonical structure for controlling information flow in neural networks with an efficient feedback routing mechanism based on a strategy of Distributed Iterative Gating (DIGNet). The structure of this mechanism derives from a strong conceptual foundation, and presents a light-weight mechanism for adaptive control of computation similar to recurrent convolutional neural networks by integrating feedback signals with a feed forward architecture. In contrast to other RNN formulations, DIGNet generates feedback signals in a cascaded manner that implicitly carries information from all the layers above. This cascaded feedback propagation by means of the propagator gates is found to be more effective compared to other feedback mechanisms that use feedback from output of either the corresponding stage or from the previous stage. Experiments reveal the high degree of capability that this recurrent approach with cascaded feedback presents over feed-forward baselines and other recurrent models for pixel-wise labeling problems on three challenging datasets, PASCAL VOC 2012, COCO-Stuff, and ADE20K.

We tackle the problem of Human Locomotion Forecasting, a task for jointly predicting the spatial positions of several keypoints on human body in the near future under an egocentric setting. In contrast to the previous work that aims to solve either the task of pose prediction or trajectory forecasting in isolation, we propose a framework to unify these two problems and address the practically useful task of pedestrian locomotion prediction in the wild. Among the major challenges in solving this task is the scarcity of annotated egocentric video datasets with dense annotations for pose, depth, or egomotion. To surmount this difficulty, we use state-of-the-art models to generate (noisy) annotations and propose robust forecasting models that can learn from this noisy supervision. We present a method to disentangle the overall pedestrian motion into easier to learn subparts by utilizing a pose completion and a decomposition module. The completion module fills in the missing key-point annotations and the decomposition module breaks the cleaned locomotion down to global (trajectory) and local (pose keypoint movements). Further, with Quasi RNN as our backbone, we propose a novel hierarchical trajectory forecasting network that utilizes low-level vision domain specific signals like egomotion and depth to predict the global trajectory. Our method leads to state-of-the-art results for the prediction of human locomotion in the egocentric view

Reasoning about complex visual scenes involves perception of entities and their relations. Scene Graphs (SGs) provide a natural representation for reasoning tasks, by assigning labels to both entities (nodes) and relations (edges). Reasoning systems based on SGs are typically trained in a two-step procedure: first, a model is trained to predict SGs from images, and next a separate model is trained to reason based on the predicted SGs. However, it would seem preferable to train such systems in an end-to-end manner. The challenge, which we address here is that scene-graph representations are non-differentiable and therefore it isn't clear how to use them as intermediate components. Here we propose Differentiable Scene Graphs (DSGs), an image representation that is amenable to differentiable end-to-end optimization, and requires supervision only from the downstream tasks. DSGs provide a dense representation for all regions and pairs of regions, and do not spend modelling capacity on regions of the image that do not contain objects or relations of interest. We evaluate our model on the challenging task of identifying referring relationships (RR) in three benchmark datasets: Visual Genome, VRD and CLEVR. Using DSGs as an intermediate representation leads to new state-of-the-art performance. The full code is available at https://github.com/shikorab/DSG.

State-of-the-art neural network models estimate large displacement optical flow in multi-resolution and use warping to propagate the estimation between two resolutions. Despite their impressive results, it is known that there are two problems with the approach. First, the multi-resolution estimation of optical flow fails in situations where small objects move fast. Second, warping creates artifacts when occlusion or dis-occlusion happens. In this paper, we propose a new neural network module, Deformable Cost Volume, which alleviates the two problems. Based on this module, we designed the Deformable Volume Network (Devon) which can estimate multi-scale optical flow in a single high resolution. Experiments show Devon is more suitable in handling small objects moving fast and achieves comparable results to the state-of-the-art methods in public benchmarks.

In this work, we address the problem of precisely localizing key frames of an action, for example, the precise time that a pitcher releases a baseball, or the precise time that a crowd begins to applaud. Key frame localization is a largely overlooked and important action-recognition problem, for example in the field of neuroscience, in which we would like to understand the neural activity that produces the start of a bout of an action. To address this problem, we introduce a novel structured loss function that properly weights the types of errors that matter in such applications: it more heavily penalizes extra and missed action start detections over small misalignments. Our structured loss is based on the best matching between predicted and labeled action starts. We train recurrent neural networks (RNNs) to minimize differentiable approximations of this loss. To evaluate these methods, we introduce the Mouse Reach Dataset, a large, annotated video dataset of mice performing a sequence of actions. The dataset was collected and labeled by experts for the purpose of neuroscience research. On this dataset, we demonstrate that our method outperforms related approaches and baseline methods using an unstructured loss.

The evolving deployment of face recognition system has raised concerns regarding the vulnerability of those systems to various attacks. The morphed face attack involves two different face images via morphing to obtain an attack image face image similar to both original images. The obtained morphed image can easily be verified against both subjects visually and by Face Recognition Systems (FRS). The face morphing attack thus raises a severe concern to various security applications like border control and e-passport unless such attacks are detected and mitigated. In this work, we propose a novel method to reliably detect the morphed face attacks using a new denoising framework. Considering the time complexity and parameterization efforts, we realize the proposed method using a new deep Multi-scale Context Aggregation Network (MS-CAN). Extensive experiments are carried out on three different morphed face image datasets. The Morphed Attack Detection (MAD) performance of the proposed method is also benchmarked against 13 different state-of-the-art techniques using the ISO IEC 30107-3 evaluation metrics. Based on the obtained quantitative results reported, the proposed method has indicated the best performance.

Visual content has become the primary source of information, as evident in the billions of images and videos, shared and uploaded every single day. This has led to an increase in alterations in images and videos to make them more informative and eye-catching for the viewers worldwide. Some of these alterations are simple, like copy-move, and are easily detectable, while other sophisticated alterations like reenactment are hard to detect. Reenactment alterations allow the source to change the target expressions and create photo-realistic images and videos where such modifications are hard to detect. Significant work has been done towards creating such images and videos. However, the detection of such alterations still requires research. This research proposes a learning-based algorithm for detecting reenactment based alterations. The proposed algorithm uses a multi-stream network that learns regional artifacts and provides a robust performance at various compression levels. We also propose a loss function for the balanced learning of the streams for the proposed network. The performance is evaluated on the publicly available FaceForensics dataset. The results show state-of-the-art classification accuracy of 99.96%, 99.10%, and 91.20% for no, easy, and hard compression factors, respectively.

In this paper we consider the task of image-guided depth completion where our system must infer the depth at every pixel of an input image based on the image content and a sparse set of depth measurements. We propose a novel approach that builds upon the strengths of modern deep learning techniques and classical fitting algorithms and significantly improves performance. The proposed method replaces the final 1-by-1 convolutional layer employed in most depth completion networks with a least squares fitting module which computes weights by fitting the implicit depth bases to the given sparse depth measurements. In addition, we show how our proposed method can be naturally extended to a multi-scale formulation for improved self-supervised training. We demonstrate through extensive experiments on various datasets that our approach achieves consistent improvements over a state-of-the-art baseline method with minimal computational overhead.

This paper proposes a Deep Learning based edge detector, which is inspired on both HED (Holistically-Nested Edge Detection) and Xception networks. The proposed approach generates thin edge-maps that are plausible for human eyes; it can be used in any edge detection task without previous training or fine tuning process. As a second contribution, a large dataset with carefully annotated edges, has been generated. This dataset has been used for training the proposed approach as well the state-of-the-art algorithms for comparisons. Quantitative and qualitative evaluations have been performed on different benchmarks showing improvements with the proposed method when F-measure of ODS and OIS are considered.

Rotating and zooming cameras, also called PTZ (Pan-Tilt-Zoom) cameras, are widely used in modern surveillance systems. While their zooming ability allows acquiring detailed images of the scene, it also makes their calibration more challenging since any zooming action results in a modification of their intrinsic parameters. Therefore, such camera calibration has to be computed online; this process is called self-calibration. In this paper, given an image pair captured by a PTZ camera, we propose a deep learning based approach to automatically estimate the focal length and distortion parameters of both images as well as the rotation angles between them. The proposed approach relies on a dual-Siamese structure, imposing bidirectional constraints. The proposed network is trained on a large-scale dataset automatically generated from a set of panoramas. Empirically, we demonstrate that our proposed approach achieves competitive performance with respect to both deep learning-based and traditional state-of-the art methods. Our code and model will be publicly available at https://github.com/ChaoningZhang/DeepPTZ.

In this paper, we propose a two-stage fully 3D network, namely DeepFuse, to estimate human pose in 3D space by fusing body-worn Inertial Measurement Unit (IMU) data and multi-view images deeply. The first stage is designed for pure vision estimation. To preserve data primitiveness of multi-view inputs, the vision stage uses multi-channel volume as data representation and 3D soft-argmax as activation layer. The second one is the IMU refinement stage which introduces an IMU-bone layer to fuse the IMU and vision data earlier at data level. without requiring a given skeleton model a priori, we can achieve a mean joint error of 28.9mm on TotalCapture dataset and 13.4mm on Human3.6M dataset under protocol 1, improving the SOTA result by a large margin. Finally, we discuss the effectiveness of a fully 3D network for 3D pose estimation experimentally which may benefit future research.

Paper-intensive industries like insurance, law, and government have long leveraged optical character recognition (OCR) to automatically transcribe hordes of scanned documents into text strings for downstream processing. Even today, there are still many scanned documents and mail that come into businesses in non-digital format. Text to be extracted from real world documents is often nestled inside rich formatting, such as tabular structures or forms with fill-in-the-blank boxes or underlines whose ink often touches or even strikes through the ink of the text itself. Further, the text region could have random ink smudges or spurious strokes. Such ink artifacts can severely interfere with the performance of recognition algorithms or other downstream processing tasks. In this work, we propose DeepErase, a neural-based preprocessor to erase ink artifacts from text images. We devise a method to programmatically assemble real text images and real artifacts into realistic-looking "dirty" text images, and use them to train an artifact segmentation network in a weakly supervised manner, since pixel-level annotations are automatically obtained during the assembly process. In addition to high segmentation accuracy, we show that our cleansed images achieve a significant boost in recognition accuracy by popular OCR software such as Tesseract 4.0. Finally, we test DeepErase on out-of-distribution datasets (NIST SDB) of scanned IRS tax return forms and achieve double-digit improvements in accuracy. All experiments are performed on both printed and handwritten text.

Effective analysis of hyperspectral imagery is essential for gathering fast and actionable information of large areas affected by atmospheric and green house gases. Existing methods, which process hyperspectral data to detect amorphous gases such as CH4 require manual inspection from domain experts and annotation of massive datasets. These methods do not scale well and are prone to human errors due to the plumes' small pixel-footprint signature. The proposed Hyperspectral Mask-RCNN (H-mrcnn) uses principled statistics, signal processing, and deep neural networks to address these limitations. H-mrcnn introduces fast algorithms to analyze large-area hyper-spectral information and methods to autonomously represent and detect CH4 plumes. H-mrcnn processes information by match-filtering sliding windows of hyperspectral data across the spectral bands. This process produces information-rich features that are both effective plume representations and gas concentration analogs. The optimized matched-filtering stage processes spectral data, which is spatially sampled to train an ensemble of gas detectors. The ensemble outputs are fused to estimate a natural and accurate plume mask. Thorough evaluation demonstrates that H-mrcnn matches the manual and experience-dependent annotation process of experts by 85% (IOU). H-mrcnn scales to larger datasets, reduces the manual data processing and labeling time (12 times), and produces rapid actionable information about gas plumes.