摘要
针对用蒙特卡罗路径跟踪绘制3D场景立体画面的计算开销巨大,而3D warping根据左眼画面生成右眼画面,但无法正确地呈现镜面反射现象,且存在空洞的问题,提出3D场景立体画面绘制的视觉感知驱动光照复用方法.首先,逐像素地跟踪少量路径实现预绘制,通过获取左右眼画面的可视场景点表、显著性图和背景掩膜等辅助信息,完成左眼画面绘制;然后分别进行背景复用和3D warping复用,避免右眼画面部分像素的路径跟踪计算;最后,根据显著性图分配镜面反射区域及剩余空洞区域的路径样本数,在低采样水平下完成右眼画面绘制.在PBRT-V3通用测试场景下的实验表明,该方法能够准确地呈现镜面反射现象,提高绘制结果的视觉感知质量,与基于3Dwarping和路径跟踪的方法相比,其平均PSNR,SSIM和VSNR分别提高5.68%~22.62%,6.38%~19.36%和4.25%~34.53%.
Stereoscopic rendering of 3D scenes directly based on Monte Carlo path tracing is usually expensive and 3D warping techniques result in inaccurate specular reflections and holes.We proposed a visual-perception-driven reuse method of illumination computation in rendering stereoscopic images of 3D scenes.Firstly,we rendered two views with a limited number of path samples to gather auxiliary information such as visible surface points,saliency maps,and background masks and synthesized the left image.Secondly,we applied background reusing and 3D warping to reuse the results of illumination computation thus avoiding some tracing calculations.Finally,we rendered the specular regions and the remaining holes with saliency-assigned paths and synthesized the right image.Experiments with the use of general scenes of PBRT-V3 showed the proposed method produced accurate specular reflections and provided better perceptual quality,and its average PSNR,SSIM and VSNR were improved by 5.68%–22.62%,6.38%–19.36%and 4.25%–34.53%,respectively,compared with the existing methods based on 3D warping and path tracing.
作者
申忠业
陈纯毅
于海洋
张日培
胡小娟
Shen Zhongye;Chen Chunyi;Yu Haiyang;Zhang Ripei;Hu Xiaojuan(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2023年第5期708-718,共11页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(U19A2063)
吉林省科技发展计划(202101ZYTS151,202101ZYTS036,20230201080GX)。
作者简介
申忠业(1991-),男,硕士研究生,CCF会员,主要研究方向为三维场景绘制;通信作者:陈纯毅(1981-),男,博士,教授,博士生导师,CCF高级会员,主要研究方向为三维场景建模与绘制,chenchunyi@cust.edu.cn;于海洋(1989-),男,博士,讲师,主要研究方向为机器学习理论与应用;张日培(1992-),男,博士研究生,主要研究方向为三维场景建模与绘制;胡小娟(1985-),女,博士,讲师,CCF会员,主要研究方向为虚拟现实和机器学习.
