First render

Last weekend I implemented a light tree construction algorithm for omnidirectional lights, bounds for omnidirectional lights and a trivial bound for the cos term (i.e. 1) and the diffuse brdf (this brdf is already a constant). Using this basic implementation I rendered some pictures of the famous cornell box. The setup for all pictures was as follows:

• resolution: 512x512
• samples/pixel: 4
• pixel filter: gaussian
• area light approx.: 256 omnidirectional lights arranged in a regular grid
• render-machine: AMD X2 3800+ (single core used) with 1GB ram

The first picture shows a reference solution (i.e. sampling all lights) and rendered for for 234s. The next 2 pictures were rendered using the basic implementation with a maximum relative error of respectively 1% and 2%. These images rendered for 204s and 101s with a average cutsizes of 157 and 72 respectively.



The final 2 pictures show the relative error between the reference picture and the pictures rendered using the lightcuts algorithm. There are a few strangely colored pixel due to a division by zero in some (but not all) of the channels. Although the relative error on most pixels is far below the maximum relative error, some pixels (especially in dark regions) have considerably higher relative errors (up to 100% for the 1% relative error image and up to 300% for the 2% relative error image). It is not clear yet whether these are caused by a flaw in the implementation, some numerical issue or something else.