In 2012 I was investigating the highly competetive field of global illumination techniques and breaking my head over the popular question what the best global illumination approach for games would be. Most of the existing techniques degraded the illumination quality significantly, and I was most interested in near-perfect quality at real-time framerates. I had an idea which was based on modifying the illumination based on changes in the scene. This required usage of the highly outdated radiosity family of global illumination techniques. Even though it was a long shot, I wanted to go for it. There was no supervisor at the Utrecht University who could assist me on the technical contents of the thesis, so I decided I’d do without supervision on this part.
As can be read in my thesis, my cross redistribution radiosity technique works quite well. It is able to generate nearly perfect diffuse global illumination in real-time, albeit with some limitations on geometric complexity. The technique has advantages and disadvantages completely different from the state of the art techniques, which makes it a valuable addition to the field of global illumination.
The thesis has won the Science Faculty’s thesis prize. Furthermore it was nominated as one of the three candidates for the general Utrecht University thesis prize.
The radiosity class of techniques is relatively underrepresented in the current studies towards real-time dynamic global illumination, while it provides unique advantages. This thesis builds upon the existing theory for the redistribution of radiosity and the existing progressive refinement adaptation for graphics hardware.
The dynamic radiosity theory is expanded by this thesis to differentiate between types of redistribution based on the emitting or receiving role of the static and dynamic patches. The new theory grounds the proposed cross redistribution radiosity algorithm. The redistribution of radiosity originating from static patches is substituted by different types of redistribution. The advantage of this new technique is that the number of rendered hemicubes no longer depends on the number of static patches but only on the number of dynamic patches, which proves to be a significant performance increase. The novel cross projection function is essential for realizing the reduction of rendered hemicubes.
Several hardware accelerated radiosity adaptations are developed for comparison. The hardware accelerated adaptation of the progressive refinement algorithm is improved significantly, focusing on quality and general applicability. Furthermore, a novel hardware accelerated adaptation of incremental radiosity is introduced, which introduces the concept of reshooting. Finally, the cross redistribution radiosity theory is supplemented with a fast hardware accelerated adaptation, with the cross projection adaptation as main innovation.
A qualitative analysis demonstrates that all implementations are capable of delivering very high quality global illumination, although the scene properties are restricted. In certain situations cross redistribution radiosity generates artifacts due to undersampling. The execution time is benchmarked for all implementations, which proves that the cross redistribution radiosity adaptation performs well within real-time bounds. A comparative analysis of competing real-time high quality global illumination techniques is favorable to our cross redistribution radiosity method, within the imposed scene restrictions.
If you have any questions, remarks or suggestions regarding the thesis, please leave a comment or contact me.
The thesis demo, source code and other data will be made available later.
The video is captured in real-time on an AMD Radeon HD 5770.