Replica exchange light transport on relaxed distributions

Hisanari Otsu, Yonghao Yue, Qiming Hou, Kei Iwasaki, Yoshinori Dobashi, Tomoyuki Nishita

Abstract: Developing a robust method for computing global illumination is a challenging problem. A Markov chain Monte Carlo (MCMC) method, like [Jakob and Marschner 2012], samples the light path space with a probability proportional to the per-path contribution, by successively mutating path samples (e.g., perturbing a reflection direction). In practice, a path sample could get stuck in a high energy peak for multiple mutations, resulting in a bright spot artifact. To resolve this problem, we present a new unbiased rendering framework based on a replica exchange technique [Kitaoka et al. 2009], a variant of MCMC technique. A replica exchange technique incorporates a set of different distributions. We propose to introduce a set of relaxed distributions, which are beneficial for reducing the chance of getting stuck.