On-line Learning of Parametric Mixture Models for Light Transport Simulation

摘要

Monte Carlo techniques for light transport simulation rely on importancernsampling when constructing light transport paths. Previousrnwork has shown that suitable sampling distributions can bernrecovered from particles distributed in the scene prior to rendering.rnWe propose to represent the distributions by a parametric mixturernmodel trained in an on-line (i.e. progressive) manner from a potentiallyrninfinite stream of particles. This enables recovering goodrnsampling distributions in scenes with complex lighting, where thernnecessary number of particles may exceed available memory. Usingrnthese distributions for sampling scattering directions and lightrnemission significantly improves the performance of state-of-the-artrnlight transport simulation algorithms when dealing with complexrnlighting.