Detecting Visual Convergence for Stochastic Global Illumination

Photorealistic rendering, based on unbiased stochastic global illumination, is now within reach of any computer artist by using commercially or freely available softwares. One of the drawbacks of these softwares is that they do not provide any tool for detecting when convergence is reached, relying entirely on the user for deciding when stopping the computations. In this paper we detail two methods that aim at finding perceptual convergence thresholds for solving this problem. The first one uses the VDP image quality measurement for providing a global threshold. The second one uses SVM classifiers which are trained and used on small subparts of images and allow to take into account the heterogeneity of convergence through the image area. These two approaches are validated by using experimentations with human