Local Quasi-Monte Carlo Exploration

In physically-based image synthesis, the path space of light transport paths is usually explored by stochastic sampling. The two main families of algorithms are Monte Carlo/quasi-Monte Carlo sampling and Markov chain Monte Carlo. While the former is known for good uniform discovery of important regions, the latter facilitates efficient exploration of local effects. We introduce a hybrid sampling technique which uses quasi-Monte Carlo points to achieve good stratification in both stages: we use the Halton sequence to generate initial seed paths and rank-1 lattices for local exploration. This method avoids the hard problem of introducing QMC sequences into MCMC while still stratifying samples both globally and locally. We propose perturbation strategies that prefer dimensions close to the camera, facilitating efficient reuse of transport path suffixes. This framework provides maximum control of the sampling scheme by the programmer, which can be hard to achieve with Markov chain-based methods. We show that local QMC exploration can generate results on par with state of the art light transport sampling methods, while providing more uniform convergence, improving temporal consistency.