Importance Sampling of Many Lights with Adaptive Tree Splitting

Abstract

We present a technique to importance sample large collections of lights (including mesh lights as collections of small emitters) in the context of Monte-Carlo path tracing. A bounding volume hierarchy over all emitters is traversed at each shading point using a single random number in a way that importance samples their predicted contribution. The tree aggregates energy, spatial and orientation information from the emitters to enable accurate prediction of the effect of a cluster of lights on any given shading point. We further improve the performance of the algorithm by forcing splitting until the importance of a cluster is sufficiently representative of its contents.