Stanford University
This work will be presented at SIGGRAPH '95.
This talk presents a powerful new tool for reducing the variance of MC integration, allowing much higher quality images for the same computational effort. At the same time we show how these calculations can be made more robust, by constructing estimators that have low variance for a broad class of integrands.
Our basic approach is to take samples using a set of sampling techniques, where each technique is designed to sample some difficult aspect of the integrand. The key innovation is how we combine the samples to minimize variance. Rather than partitioning the domain, we explore strategies that compute weighted combinations of all the samples. We present specific strategies that are unbiased, simple to compute, and provably close to optimal.
We show that our methods can reduce variance significantly at little additional cost, using experiments from several areas in rendering: calculation of glossy highlights from area light sources, the "final gather" pass of some radiosity algorithms, and direct solution of the rendering equation using path tracing.
The full paper is also available.