Entropy-based adaptive sampling

Ray tracing techniques need supersampling to reduce aliasing and/or noise in the final image. Since not all the pixels in the image require the same number of rays, supersampling can be implemented by adaptive subdivision of the sampling region, resulting in a refinement tree. In this paper we present a theoretically sound adaptive sampling method based on entropy, the classical measure of information. Our algorithm is orthogonal to the method used for sampling the pixel or for obtaining the radiance of the hitpoint in the scene. Results will be shown for our implementation within the context of stochastic ray tracing and path tracing. We demonstrate that our approach compares well to the ones obtained by using classic strategies based on contrast and variance.