Personal tools
You are here: Home Publications Improved Sampling for Gradient-Domain Metropolis Light Transport

Improved Sampling for Gradient-Domain Metropolis Light Transport

ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia 2014)
 GMLT
Marco Manzi Fabrice
Rousselle
Markus Kettunen Jaakko Lehtinen Matthias Zwicker
 University of Bern University of Bern,
Disney Research 
Aalto University Aalto University
NVIDIA
University of Bern

Abstract

We present a generalized framework for gradient-domain Metropolis rendering, and introduce three techniques to reduce sampling artifacts and variance. The first one is a heuristic weighting strategy that combines several sampling techniques to avoid outliers. The second one is an improved mapping to generate offset paths required for computing gradients. Here we leverage the properties of manifold walks in path space to cancel out singularities. Finally, the third technique introduces generalized screen space gradient kernels. This approach aligns the gradient kernels with image structures such as texture edges and geometric discontinuities to obtain sparser gradients than with the conventional gradient kernel. We implement our framework on top of an existing Metropolis sampler, and we demonstrate significant improvements in visual and numerical quality of our results compared to previous work.

Materials

publicationresults and comparisons, presentation

Document Actions