Object Space Adaptive Sampling for Global Illumination

Complex scenes and radiance distributions are common in realistic image synthesis. The variance of Monte Carlo sampling is large in these situations. Therefore adaptive method is needed to sample efficiently. We present an object space adaptive sampling method to handle complex radiance distributions in global illumination. The scene is segmented into sub-regions with a 5D tree, and the incident radiance distributions within each sub-region are approximated with spherical 2D trees. The spherical 2D trees is used together with BRDF and light source sampling in the Rao-Blackwellized D-kernel population Monte Carlo framework. Significant efficiency improvements are achieved over the existing methods.