KD-tree based parallel adaptive rendering

Abstract

Multidimensional adaptive sampling technique is crucial for generating high quality images with effects such as motion blur, depth-of-field and soft shadows, but it costs a lot of memory and computation time. We propose a novel kd-tree based parallel adaptive rendering approach. First, a two-level framework for adaptive sampling in parallel is introduced to reduce the computation time and control the memory cost: in the prepare stage, we coarsely sample the entire multidimensional space and use kd-tree structure to separate it into several multidimensional subspaces; in the main stage, each subspace is refined by a sub kd-tree and rendered in parallel. Second, novel kd-tree based strategies are introduced to measure space’s error value and generate anisotropic Poisson disk samples. The experimental results show that our algorithm produces better quality images than previous ones.