After some more testing and re-reading the lightcuts paper I have come to the conclusion that the calculation of the total radiance estimate (used to determine what errors are allowed) in my implementation differs from the one used by the authors of the lightcuts paper. As described in my previous post, I used the maximum total illumination as an estimate of the total illumination. While the authors of the lightcuts papers do not state so explicitly they most likely used the current lightcut approximation as an estimate for the total illumination. This estimate is by definition always less then the maximum total illumination and will thus result in a lower allowed error per cluster.
Using the current lightcut approximation generally improves the estimate but is especially important in highly occluded regions because the difference with the maximum total illumination is largest here. I re-rendered the cornell box scene using this new total illumination estimate:
- resolution: 512x512
- samples/pixel: 4
- pixel filter: gaussian
- area light approx.: 1024 omnidirectional lights
- render-machine: Core2 Duo T7100 (single core used) with 2GB ram
As you can see I used a different machine so the timings are not comparable to my previous test. The first picture shows the reference solution (795s). The following 2 pictures show the lightcuts approximation with a respective maximum relative error of 2% (182s) and 5% (107s). Both of the lightcuts pictures used a maximum cut size of 1024 (i.e. all lights) which is always used in the completely occluded regions due to the new total illumination estimate being 0. The average cut sizes where 119 and 73.
Like in my previous test, the final 2 pictures show the relative error. It is clear that the total illumination estimate used by the authors of the lightcuts paper is a lot better then using the maximum total illumination: the relative error is even in the highly occluded places very close to the target (avg: 0.36% and 0.95%; max: 2.1% and 3.5%).
No comments:
Post a Comment