基于再次事件估计的光学厚介质高性能渲染方法
A Fast Rendering Method for Optically Thick Participating Media Based on Next-Next Event Estimation
-
摘要: 参与介质的经典渲染方法中, 对于散射点的采样仅考虑相函数和透光率的影响, 这就使参与介质的经典渲染方法可能会以极小的概率采样到极大的贡献, 进而造成数值上极大的噪点. 为了解决这一问题, 一种参与介质正向渲染方法被提出, 首先通过一个启发式的代理分布采样额外散射点, 利用多重重要性采样方法将这些单次散射贡献进行混合, 规避在光源极小时直接光照上的多重重要性采样方法失效的问题; 然后将单次散射贡献保留在内存中, 在一条路径上之后的散射点中被复用, 提供额外的单次散射贡献, 或结合重采样方法, 无需额外的计算就可以无偏地得到更好的多次散射路径. 得到的实验结果表明, 所提方法在保证无偏性的同时, 可以取得比经典方法更快的渲染收敛速度.Abstract: The traditional rendering technique for participating media solely considers the influence of transmittance and phase function while sampling scattering points. This implies that under certain conditions, a substantial contribution could be sampled with a very tiny probability by the participating medium's traditional rendering methods, resulting in a very huge numerical noise. A novel approach to forward rendering for participating media is put forth in this paper in order to address this issue. It involves sampling extra scattering points using a heuristic proxy distribution and then combining these single scattering contributions using multiple importance sampling techniques. With this approach, the issue of the multiple importance sampling method failing under direct illumination from a tiny light source can be avoided. These single scattering contributions can also be combined with the resampling method to obtain a better multiple scattering path without bias or extra computation. Furthermore, these single scattering contributions can be stored in memory and reused at subsequent scattering points along a path to provide additional single scattering contributions. Experiments demonstrate that the suggested approach can guarantee unbiased estimations while achieving a faster rate of rendering convergence than the traditional method.