面向真实感绘制的光场低秩性及其应用研究

Efficiently rendering the richness of our visual world is a central goal of computer graphics. However, fast and realistic rendering is still a challenging task so far. Motivated by the observation that there exist correlations and coherences among rays used in photorealistic rendering, we propose research to exploit and utilize the low-rank properties of these light fields. Recently, low-rank analysis and matrix recovery techniques have been widely studied in data mining and machine learning. In our research, we plan to adapt low-rank matrix completion, one of matrix rank minimization techniques, in the rendering research and develop new fast and efficient rendering algorithms. Our research will be taken as following steps. First, we plan to explore the parameterization and matrix representation of light fields, and carry out theoretical analysis on the low-rank property of light field. The first step is the foundation of our entire research and provides fundamental theoretical support for new rendering algorithms. Second, based on those low-rank matrix completion methods, we will develop new light field sampling, completion and integration algorithms so as to achieve efficient rendering under low sampling rate. Third, we will use perceptional criteria to validate our studies in rank analysis of light field. User study will be the tool to bridge the perceptional errors and the rank criteria. Finally, based on these progresses, we will develop a new rendering architecture for the low-rank light field. Our research will provide new theoretical analysis and algorithms for photorealistic rendering and move the entire rendering research forward.

快速高效的真实感绘制一直是图形学的重要目标之一。本项目主要针对真实感绘制中光场的低秩性及其在绘制算法中的应用开展研究，旨在通过充分挖掘绘制光场光线本身的连续性与冗余性，利用当前在数据分析领域发展出的基于凸优化的低秩矩阵填充等方法，在对光场光线少量采样的情况下，实现快速的真实感绘制。本项目计划围绕如下几方面开展工作：首先研究面向真实感绘制光场的参数化与矩阵表示、低秩性分析等相关理论，为后续研究奠定理论基础；其次，基于低秩矩阵填充，研究针对低秩光场的采样、填充与积分方法，实现低采样率下的快速光场绘制；再次，计划以基于视觉感知的真实性为依据，采用用户调查的方法对本项目提出的低秩光场理论与光场填充方法进行分析与验证；最后，构建基于光场低秩性的快速绘制框架。通过本项目的研究，将为面向真实感绘制的光场数据计算与使用提供理论与分析，并为实现高效率、高逼真度的绘制技术的研发提供理论参考与数据支撑。

本课题主要针对真实感绘制中光场的低秩性及其在绘制算法中的应用开展研究，在国际上创新提出了光场稀疏表达与高效采样重构的方法：通过充分挖掘绘制光场本身信号所表征出的稀疏性，利用数据分析与挖掘领域中发展出基于凸优化的低秩矩阵填充等方法，实现对绘制光场的高效采样与积分，进而实现快速的真实感绘制。主要研究内容包括：面向真实感绘制光场的矩阵表示、稀疏性的相关理论与分析；通过对基于凸优化的矩阵低秩填充方法的改造与借用，开发了针对不同稀疏光场的低采样率的采样、填充与绘制方法；在真实感绘制之外，进一步将稀疏性分析应用于实时绘制中，研发了多种实时绘制算法，分析了不同视觉感知误差对绘制质量与效果的影响；最后，将上述方法与技术应用于并行绘制算法中，针对性地构建了面向着色器性能/绘制质量、不同绘制平台能耗/绘制性能等不同绘制目标的快速绘制框架。共发表相关研究论文14 篇，其中11篇SCI国际期刊论文。在计算机图形学领域的国际顶级学术会议/期刊ACM SIGGRAPH/SIGGRAPH Asia、ACM Trans. on Graph.（TOG）上发表论文4篇，国际期刊CGF、C&G等上发表论文7篇。授权国家发明专利5项，申请国家发明专利5项。课题共培养/联合培养博士3名，硕士生10名。