(90 pages, 95 figures.)
Abstract
This report presents an extensive empirical comparison of matrix, progressive, and wavelet radiosity algorithms for simulating diffuse interreflection in three-dimensional scenes. The algorithms are tested in their basic forms, without advanced variations such as clustering, discontinuity meshing, or Monte Carlo techniques. The three algorithms were implemented in a common code base to facilitate direct empirical comparison. A number of parameterized scenes were designed to test the basic methods' ability to deal with such issues as singularities, occlusion, high reflectance, and scene complexity. Each algorithm was run on the set of scenes at several parameter settings, and results were examined in terms of their error, speed, and memory consumption.
For the basic algorithms as we implemented them, our results show: Progressive radiosity with substructuring is best for simple scenes, but for moderately complex scenes it is outperformed by wavelet radiosity using the Haar basis. Wavelet methods use an immense amount of memory; without clustering they become totally impractical for complex scenes. The problem is particularly severe for higher order bases, less so for Haar. Visibility handling was also found to be a critical problem with higher order wavelets.
This study also provides a general framework for comparisons of global illumination techniques.
The report is online in the following formats:
(Note that the table of contents, section numbers, and figure numbers are all clickable links in the PDF file, which makes it easier to browse online.)
See also the short version.
ajw@cs.cmu.edu, ph@cs.cmu.edu.