PG2018 Short Papers and Posters

Permanent URI for this collection

https://diglib.eg.org/handle/10.2312/2632665

Browse

Now showing 1 - 20 of 28

3D VAE-Attention Network: A Parallel System for Single-view 3D Reconstruction
(The Eurographics Association, 2018) Hu, Fei; Yang, Xinyan; Zhong, Wei; Ye, Long; Zhang, Qin; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
3D object reconstruction from single view image is a challenge task. Due to the fact that the information contained in one isolated image is not sufficient for reasonable 3D shape reconstruction, the existing results on single-view 3D reconstruction always lack marginal voxels. To tackle this problem, we propose a parallel system named 3D VAE-attention network (3VAN) for single view 3D reconstruction. Distinct from the common encoder-decoder structure, the proposed network consists of two parallel branches, 3D-VAE and Attention Network. 3D-VAE completes the general shape reconstruction by an extension of standard VAE model, and Attention Network supplements the missing details by a 3D reconstruction attention network. In the experiments, we verify the feasibility of our 3VAN on the ShapeNet and PASCAL 3D+ datasets. By comparing with the state-of-art methods, the proposed 3VAN can produce more precise 3D object models in terms of both qualitative and quantitative evaluation.
Anisotropic Spectral Manifold Wavelet Descriptor for Deformable Shape Analysis and Matching
(The Eurographics Association, 2018) Li, Qinsong; Liu, Shengjun; Hu, Ling; Liu, Xinru; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
In this paper, we present a novel framework termed Anisotropic Spectral Manifold Wavelet Transform (ASMWT) for shape analysis. ASMWT comprehensively analyzes the signals from multiple directions on local manifold regions of the shape with a series of low-pass and band-pass frequency filters in each direction. Using the ASMWT coefficients of a very simple function, we efficiently construct a localizable and discriminative multiscale point descriptor, named as the Anisotropic Spectral Manifold Wavelet Descriptor (ASMWD). Since the filters used in our descriptor are direction-sensitive and able to robustly reconstruct the signals with a finite number of scales, it makes our descriptor be intrinsic-symmetry unambiguous, compact as well as efficient. The extensive experimental results demonstrate that our method achieves significant performance than several state-of-the-art methods when applied in vertex-wise shape matching.
Bottom-up/Top-down Geometric Object Reconstruction with CNN Classification for Mobile Education
(The Eurographics Association, 2018) Guo, Ting; Cui, Rundong; Qin, Xiaoran; Wang, Yongtao; Tang, Zhi; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Geometric objects in educational materials are often illustrated as 2D line drawings, which results in the loss of depth information. To alleviate the problem of fully understanding the 3D structure of geometric objects, we propose a novel method to reconstruct the 3D shape of a geometric object illustrated in a line drawing image. In contrast to most existing methods, ours directly take a single line drawing image as input and generate a valid sketch for reconstruction. Given a single input line drawing image, we first classify the geometric object in the image with convolution neural network (CNN). More specifically, we pre-train the model with simulated images to alleviate the problems of data collection and unbalanced distribution among different classes. Then, we generate the sketch of the geometric object with our proposed bottom-up and top-down scheme. Finally, we finish reconstruction by minimizing an objective function of reconstruction error. Extensive experimental results demonstrate that our method performs significantly better in both accuracy and efficiency compared with the existing methods.
A Deep Learned Method for Video Indexing and Retrieval
(The Eurographics Association, 2018) Men, Xin; Zhou, Feng; Li, Xiaoyong; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
In this paper, we proposed a deep neural network based method for content based video retrieval. Our approach leveraged the deep neural network to generate the semantic information and introduced the graph-based storage structure to establish the video indices. We devised the Inception-Single Shot Multibox Detector (ISSD) and RI3D model to extract spatial semantic information (objects) and extract temporal semantic information (actions). Our ISSD model achieved a mAP of 26.7% on MS COCO dataset, increasing 3.2% over the original SSD model, while the RI3D model achieved a top-1 accuracy of 97.7% on dataset UCF-101. And we also introduced the graph structure to build the video index with the temporal and spatial semantic information. Our experiment results showed that the deep learned semantic information is highly effective for video indexing and retrieval.
Direct Limit Volumes: Constant-Time Limit Evaluation for Catmull-Clark Solids
(The Eurographics Association, 2018) Altenhofen, Christian; Müller, Joel; Weber, Daniel; Stork, André; Fellner, Dieter W.; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
We present a novel approach for efficient limit volume evaluation on Catmull-Clark (CC) subdivision solids. Although several analogies exist between subdivision surfaces and subdivision volumes, extending Stam's limit evaluation technique from 2 to 3 dimensions is not straightforward, as irregularities and boundaries introduce new challenges in the volumetric case. We present new direct evaluation techniques for irregular volumetric topologies and boundary cells, which allow for calculating the limit of CC subdivision solids at arbitrary parameter values in constant time. Evaluation of limit points is a central aspect when using CC solids for applications such as simulation and multi-material additive manufacturing, or as a compact volumetric representation scheme for continuous scalar fields. We demonstrate that our approach is faster than existing evaluation techniques for every topological configuration or target parameter (u, v, w) that requires more than two local subdivision steps.
Effects of Surface Anisotropy on Perception of Car Body Attractiveness
(The Eurographics Association, 2018) Filip, Jiri; Kolafová, Martina; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
In the automotive industry effect coatings are used to introduce customized product design, visually communicating the unique impression of a car. Industrial effect coatings systems achieve primarily a globally isotropic appearance, i.e., surface appearance that does not change when material rotates around its normal. To the contrary, anisotropic appearance exhibits variable behavior due to oriented structural elements. This paper studies to what extent anisotropic appearance improves a visual impression of a car body beyond a standard isotropic one. We ran several psychophysical studies identifying the proper alignment of an anisotropic axis over a car body, showing that regardless of the illumination conditions, subjects always preferred an anisotropy axis orthogonal to car body orientation. The majority of subjects also found the anisotropic appearance more visually appealing than the isotropic one.
Efficient Metropolis Path Sampling for Material Editing and Re-rendering
(The Eurographics Association, 2018) Yamaguchi, Tomoya; Yatagawa, Tatsuya; Morishima, Shigeo; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
This paper proposes efficient path sampling for re-rendering scenes after material editing. The proposed sampling method is based on Metropolis light transport (MLT) and distributes more path samples to pixels whose values have been changed significantly by editing. First, we calculate the difference between images before and after editing to estimate the changes in pixel values. In this step, we render the difference image directly rather than calculating the difference in the images by separately rendering the images before and after editing. Then, we sample more paths for pixels with larger difference values and render the scene after editing by reducing variances of Monte Carlo estimators using the control variates. Thus, we can obtain rendering results with a small amount of noise using only a small number of path samples. We examine the proposed sampling method with a range of scenes and demonstrate that it achieves lower estimation errors and variances over the state-of-the-art methods.
Extreme Feature Regions for Image Matching
(The Eurographics Association, 2018) Fan, Baijiang; Rao, Yunbo; Pu, Jiansu; Deng, Jianhua; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Extreme feature regions are increasingly critical for many image matching applications on affine image-pairs. In this paper, we focus on the time-consumption and accuracy of using extreme feature regions to do the affine-invariant image matching. Specifically, we proposed novel image matching algorithm using three types of critical points in Morse theory to calculate precise extreme feature regions. Furthermore, Random Sample Consensus (RANSAC) method is used to eliminate the features of complex background, and improve the accuracy of the extreme feature regions. Moreover, the saddle regions is used to calculate the covariance matrix for image matching. Extensive experiments on several benchmark image matching databases validate the superiority of the proposed approaches over many recently proposed affine-invariant SIFT algorithms.
Facial-Expression-Aware Emotional Color Transfer Based On Convolutional Neural Network
(The Eurographics Association, 2018) Pei, Min; Liu, Shiguang; Zhang, Xiaoli; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Emotional color transfer aims to change the evoked emotion of the source image to that of the target image by adjusting color distribution. Most of existing emotional color transfer methods ignore the facial expression features in the image. Therefore, we propose a new facial-expression-aware emotional color transfer framework. We firstly predict the emotion label of the image through the emotion classification network. Then, emotion labels are matched with pre-trained emotional models. Finally, we use the matched emotion model to transfer the color of the target image to the input image. Experiments demonstrate that our method outperforms the state-of-the-arts, which can successfully capture and transfer sophisticated emotion features.
Frontmatter: Pacific Graphics 2018 - Short Papers and Posters
(The Eurographics Association, 2018) Fu, Hongbo; Ghosh, Abhijeet; Kopf, Johannes; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Gauss-Seidel Progressive Iterative Approximation (GS-PIA) for Loop Surface Interpolation
(The Eurographics Association, 2018) Wang, Zhihao; Li, Yajuan; Ma, Weiyin; Deng, Chongyang; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
We propose a Gauss-Seidel progressive iterative approximation (GS-PIA) method for Loop subdivision surface interpolation by combining classical Gauss-Seidel iterative method for linear system and progressive iterative approximation (PIA) for data interpolation. We prove that GS-PIA is convergent by applying matrix theory. GS-PIA algorithm retains the good features of the classical PIA method, such as the resemblance with the given mesh and the advantages of both a local method and a global method. Compared with some existed interpolation methods of subdivision surfaces, GS-PIA algorithm has advantages in three aspects. First, it has a faster convergence rate compared with the PIA and WPIA algorithms. Second, compared with WPIA algorithm, GS-PIA algorithm need not to choose weights. Third, GS-PIA need not to modify the mesh topology compared with other methods with fairness measures. Numerical examples for Loop subdivision surfaces interpolation illustrated in this paper show the efficiency and effectiveness of GS-PIA algorithm.
GPU-based Real-time Cloth Simulation for Virtual Try-on
(The Eurographics Association, 2018) Su, Tongkui; Zhang, Yan; Zhou, Yu; Yu, Yao; Du, Sidan; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
We present a novel real-time approach for dynamic detailed clothing simulation on a moving body. The most distinctive feature of our method is that it divides dynamic simulation into two parts: local driving and static cloth simulation. In local driving, feature points of clothing will be handled between two consecutive frames. And then we apply static cloth simulation for a specific frame. Both parts are ecxuted in an entire parallel way. In practice, our system achieves real-time virtual try-on using a depth camera to capture the moving body model and meanwhile, keeps high-fidelity. Experimental results indicate that our method has significant speedups over prior related techniques.
InspireMePosing: Learn Pose and Composition from Portrait Examples
(The Eurographics Association, 2018) Sheng, Bin; Jin, Yuxi; Li, Ping; Wang, Wenxiao; Fu, Hongbo; Wu, Enhua; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Since people tend to build relationship with others by personal photography, capturing high quality photographs on mobile device has become a strong demand. We propose a portrait photography guidance system to guide user's photographing. We consider current scene image as our input and find professional photograph examples with similar aesthetic features for it. Deep residual network is introduced to gather scene classification information and represent common photograph rules by features, and random forest is adopted to establishing mapping relations between extracted features and examples. Besides, we implement our guidance system on a camera application and evaluate it by user study.
Japanese Kanji Font Style Transfer based on GAN with Unpaired Training
(The Eurographics Association, 2018) Sakai, Hiroki; Niino, Daisuke; Ijiri, Takashi; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
To design a whole package of Japanese font is labor consuming, since it usually contains about 30k kanji characters. To support an efficient design process, this poster attempts to adopt a style transfer algorithm for font package completion. Given two font packages where one contains all characters and the other lacks a large part, we train CycleGAN to perform style transfer between the two packages and transfer the style from the former to the latter. To illustrate the feasibility of our technique, we performed style transfer experiments and achieved visually plausible results by using a relatively small training data set.
Light-Field DVR on GPU for Streaming Time-Varying Data
(The Eurographics Association, 2018) Ganter, David; Alain, Martin; Hardman, David; Smolic, Aljosa; Manzke, Michael; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Direct Volume Rendering (DVR) of volume data can be a memory intensive task in terms of footprint and cache-coherency. Rayguided methods may not be the best option to interactively render to light-fields due to feedback loops and sporadic sampling, and pre-computation can rule out time-varying data. We present a pipelined approach to schedule the rendering of sub-regions of streaming time-varying volume data while minimising intermediate sub-buffers needed, sharing the work load between CPU and GPU. We show there is significant advantage to using such an approach.
Mesh Parameterization: a Viewpoint from Constant Mean Curvature Surfaces
(The Eurographics Association, 2018) Zhao, Hui; Su, Kehua; Li, Chenchen; Zhang, Boyu; Liu, Shirao; Yang, Lei; Lei, Na; Gortler, Steven J.; Gu, Xianfeng; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
We present a unified mesh paramterization algorithm for both planar and spheric domains based on mesh deformation. Unlike previous methods, our approach can produce intermediate frames from the original to target meshes. We derive and define a novel geometric flow: unit normal flow(UNF) and prove that if unit normal flow converges, it will deform a surface to a constant mean curvature(CMC) surface, such as plane and sphere. Our method works by deforming meshes of disk topology to planes, meshes of spheric topology to spheres. The unit normal flow we propose also suggests a potential direction for creating CMC surfaces.
Modeling Detailed Cloud Scene from Multi-source Images
(The Eurographics Association, 2018) Cen, Yunchi; Liang, Xiaohui; Chen, Junping; Yang, Bailin; Li, Frederick W. B.; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
Realistic cloud is essential for enhancing the quality of computer graphics applications, such as flight simulation. Data-driven method is an effective way in cloud modeling, but existing methods typically only utilize one data source as input. For example, natural images are usually used to model small-scale cloud with details, and satellite images and WRF data are used to model large scale cloud without details. To construct large-scale cloud scene with details, we propose a novel method to extract relevant cloud information from both satellite and natural images. Experiments show our method can produce more detailed cloud scene comparing with existing methods.
Progressive 3D Scene Understanding with Stacked Neural Networks
(The Eurographics Association, 2018) Song, Youcheng; Sun, Zhengxing; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
3D scene understanding is difficult due to the natural hierarchical structures and complicated contextual relationships in the 3d scenes. In this paper, a progressive 3D scene understanding method is proposed. The scene understanding task is decomposed into several different but related tasks, and semantic objects are progressively separated from coarse to fine. It is achieved by stacking multiple segmentation networks. The former network segments the 3D scene at a coarser level and passes the result as context to the latter one for a finer-grained segmentation. For the network training, we build a connection graph (vertices indicating objects and edges' weights indicating contact area between objects), and calculate a maximum spanning tree to generate coarse-to-fine labels. Then we train the stacked network by hierarchical supervision based on the generated coarseto- fine labels. Finally, using the trained model, we can not only obtain better segmentation accuracy at the finest-grained than directly using the segmentation network, but also obtain a hierarchical understanding result of the 3d scene as a bonus.
Recovering 3D Indoor Floor Plans by Exploiting Low-cost Spherical Photography
(The Eurographics Association, 2018) Pintore, Giovanni; Ganovelli, Fabio; Pintus, Ruggero; Scopigno, Roberto; Gobbetti, Enrico; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
We present a novel approach to automatically recover, from a small set of partially overlapping panoramic images, an indoor structure representation in terms of a 3D floor plan registered with a set of 3D environment maps. Our improvements over previous approaches include a new method for geometric context extraction based on a 3D facets representation, which combines color distribution analysis of individual images with sparse multi-view clues, as well as an efficient method to combine the facets from different point-of-view in the same world space, considering the reliability of the facets contribution. The resulting capture and reconstruction pipeline automatically generates 3D multi-room environments where most of the other previous approaches fail, such as in presence of hidden corners, large clutter and sloped ceilings, even without involving additional dense 3D data or tools. We demonstrate the effectiveness and performance of our approach on different real-world indoor scenes.
Robust and Efficient SPH Simulation for High-speed Fluids with the Dynamic Particle Partitioning Method
(The Eurographics Association, 2018) Zheng, Zhong; Gao, Yang; Li, Shuai; Qin, Hong; Hao, Aimin; Fu, Hongbo and Ghosh, Abhijeet and Kopf, Johannes
In this paper, our research efforts are devoted to the efficiency issue of the SPH simulation when the ratio of velocities among fluid particles is large. Specifically, we introduce a k-means clustering method into the SPH framework to dynamically partition fluid particles into two disjoint groups based on their velocities, we then use a two-scale time step scheme for these two types of particles. The smaller time steps are for particles with higher speed in order to preserve temporal details and guarantee the numerical stability. In contrast, the larger time steps are used for particles with smaller speeds to reduce the computational expense, and both types of particles are tightly coupled in the simulation.We conduct various experiments which have manifested the advantages of our methods over the conventional SPH technique and its new variants in terms of efficiency and stability.

Browse

Browsing PG2018 Short Papers and Posters by Title

Results Per Page

Sort Options