Noise Reduction for Monte Carlo Rendering: Image Space and Gradient-Domain Techniques
On 06/10/2015 at 10:30 to 12:00. Salle C3, Bâtiment Nautibus, Université Lyon I
URL : https://liris.cnrs.fr/seminaire/seminaires-mensuels/seminaires-mensuels
Informations contact : G. Damiand. guillaume.damiand@liris.cnrs.fr. +33 (0)4.72.43.14.34.
With the ongoing path tracing revolution in the movie industry, there has been a renewed interest in noise reduction for realistic image synthesis using Monte Carlo rendering. Monte Carlo algorithms are attractive because they are conceptually simple, general, and based on a physical model of light transport. Noise artifacts, however, have remained a challenge for real world applications. In this talk, I will present two strategies that we have been pursuing in our research to address this problem. First, I will discuss recent advances in image space adaptive sampling and reconstruction. I will describe an approach that leverages auxiliary image features and cross-bilateral filtering in combination with statistical error estimation to obtain surprisingly effective results. This technique has already made an impact in practical movie production. Second, I will present gradient-domain rendering, which is a form of correlated sampling for Monte Carlo rendering, followed by Poisson reconstruction of the final image. I will outline a theoretical analysis that shows how this approach reduces variance in the final output compared to conventional Monte Carlo sampling.