Q-Learned Importance Sampling for Physically Based Light Transport on the GPU
Summary
We present a GPU implementation of Dahm and Keller's Q-learning based importance sampling technique. The method requires a caching scheme to store the radiance distributions that are learned during path tracing. We tested the method on a photon map, a Poisson disk distribution and the Poisson disk distribution with a new addition called light occlusion. We found that the uniformly distributed points of the photon map produces the best results. The method itself was tested on four different scenes. We show that in an optimized GPU path tracer the method can have a positive influence on the performance, depending on the difficulty of the scene. In a comparison to a bidirectional path tracer we see that the method is able to outperform the bidirectional path tracer in a scene that is almost exclusively lit by indirect lighting. We conclude that it can be beneficial to implement the method in a GPU based path tracer.