Image Space Photon Mapping
December 3rd, 2009
CII 3039, 4:00 p.m. to 5:00 p.m.
Refreshments at 3:30 p.m.
The computer-generated 3D graphics in films like Terminator 4 are visually indistinguishable from real objects. They are hard to compute, however; the Photon Mapping algorithm used in most film productions requires about one day to compute every second of film. Our new Image Space Photon Mapping transforms that algorithm to a space where it can be evaluated in a fraction of a second. This enables nearly cinema-quality lighting for interactive appliations like video games. We recently demonstrated experimental results of the image space algorithm on complex scenes with dynamic lights and objects, using assets from real games.
Formally, photorealistic rendering as a sampling and reconstruction and Photon Mapping is one of many non-deterministic approximation algorithms for that problem. A better reconstruction strategy enables our Image Space method to converge with fewer samples than previous methods, and our sampling strategy is better-suited to parallel evaluation on graphics hardware. I will conclude with speculation on how our sampling and reconstruction observations might be applied to other problems in high.
This talk describes joint work with David Luebke at NVIDIA corporation.
Morgan McGuire is an assistant professor of Computer Science at Williams College. He received his PhD from Brown University in 2006. He served as general and papers co-chair of the ACM symposium on Interactive 3D Graphics and Games in 2008 and 2009, is an editor of the Journal of the Graphics, Game, and GPU Tools, and is the lead author of the new textbook, Creating Games: Mechanics, Content, and Technology.
He has contributed to many commercial products including the E-Ink display for the Amazon Kindle, the PeakStream high-performance computing infrastructure acquired by Google, the Titan Quest role playing game, and the Marvel Ultimate Alliance 2 video game for Xbox 360.
Last updated: November 30, 2009