Real-time filtering of Monte Carlo noise on GPU

Producing photo-realistic images, hardly distinguishable from the real photos, is one of the most important problems in computer graphics. Physically based rendering and particularly Monte Carlo path tracing is able to produce images of such quality by performing excessive sampling during Monte Carlo integration. The problem of Monte Carlo Integration is a high variance at low sampling rate. This variance appears as a noise in final image. In order to address such problem high-dimensional filtering is used. In this thesis we inspect the applicability of the Genetic Programming for the search of new high-dimensional filtering expressions and present three novel expressions generated by our method. Our method consists of iterative random changes of initial expressions until the finishing criterion is met and the comparison of the filtered pixel values, obtained with newly generated expressions, with the ground truth of the training scenes. The resulting expressions perform better than cross-bilateral filter with constant parameters. Additionally, our GPU implementation of identified expressions allows fast filtering of Monte Carlo noise with computational time of less than a second.