Improving Global Exploration of MCMC Light Transport Simulation
Equal-time comparison of the original MMLT (d) to three variants of our algorithm – Neighbor swapping (a),
Equi-energy moves (b), Importance-sampled permutations (c). Reference is shown in the image (e). Due to insufficient global exploration in the original MMLT, some of the transport (e.g. reflected caustics) is missing in the image (d), while our algorithm is able to recover it.
Markov Chain Monte Carlo (MCMC) has recently received a lot of attention in light transport simulation research [Hanika et al. 2015; Hachisuka et al. 2014]. While these methods aim at high quality sampling of local extremes of the path space (so called local exploration), the other issue – discovering these extremes – has been so far neglected. Poor global exploration results in oversampling some parts of the paths space, while undersampling or completely missing other parts (see Fig. 1). Such behavior of MCMC-based light transport algorithms limits their use in practice, since we can never tell for sure whether the image has already converged.
Outside of computer graphics, the problem of global exploration has received much attention. One of the most popular methods for improving global exploration of MCMC is parallel tempering (PT) [Swendsen and Wang 1986]. While PT has already been applied in light transport simulation [Kitaoka et al. 2009; Hachisuka and Jensen 2011], we believe its full potential is yet to be exploited. In our work we have developed a parallel tempering algorithm specifically suited for light transport simulation. We demonstrate its three variants and show that they improve global exploration of MCMC algorithms in light transport simulation.
Martin Šik and Jaroslav Křivánek. 2016. Improving global exploration of MCMC light transport simulation. In ACM SIGGRAPH 2016 Posters (SIGGRAPH '16). ACM, New York, NY, USA, , Article 50 , 2 pages. DOI: https://doi.org/10.1145/2945078.2945128
Links and Downloads
The work was supported by Charles University in Prague, project GA UK 164815, by the grant SVV–2016–260332, and by the Czech Science Foundation grant 16–18964S.