17. gradient domain vertex connection and merging

Department: Computer Science & Engineering
Research Institute Affiliation: Agile Center for Visual Computing
Faculty Advisor(s): Ravi Ramamoorthi

Primary Student
Name: Weilun Sun
Email: wes028@ucsd.edu
Phone: 510-693-7675
Grad Year: 2019

Abstract
Recently, gradient-domain rendering techniques have shown great promises in reducing Monte Carlo noise through estimating gradients by correlated samples. Gradient domain bidirectional path tracing (GBDPT) and gradient domain photon mapping (GPM) both outperform their traditional primal domain counterparts. However, these 2 methods have their own strengths and weaknesses. GBDPT performs well on rough surface interactions but fails in handling specular-diffuse-specular (SDS) paths. While GPM robustly handles SDS paths, it suffers from significant noise in low photon density regions. To get the best of both worlds, we propose a gradient counterpart of vertex connection and merging (VCM) called gradient domain vertex connection and merging (GVCM). In this work, we describe how to generalize traditional VCM in the gradient sampling context. Our new method is able to robustly combine both GBDPT and GPM through multiple importance sampling. We demonstrate several scenes where GVCM is able to converge, while both GPM and GBDPT fail in some regions within the same computation time.

« Back to Posters or Search Results