VoxLOD: Interactive ray tracing of massive models with indirect lighting using voxels

Just encountered an impressive video of a technology named VoxLOD on Youtube today:

http://www.youtube.com/watch?v=1IwtKLtYc9M

As the name aptly implies, VoxLOD uses a voxel-based LOD scheme to smoothly stream in and visualize the geometry of the massive model. The first part of the video shows direct lighting only (primary and shadow rays), while the second half is much more interesting and demonstrates real-time one bounce diffuse indirect lighting (filtered Monte Carlo GI). From the paper "Interactive ray tracing of large models using voxel hierarchies":

"We cast one shadow ray per primary or diffuse ray, and two random diffuse rays per primary ray. The diffuse rays are used to compute both one bounce of indirect irradiance and environment irradiance, which are processed with a bilateral filter [TM98] to eliminate noise."

From an earlier post on the author's blog:

"There are two light sources: a point light (the Sun) and a hemispherical one (the sky). I use Monte Carlo integration to compute the GI with one bounce of indirect lighting. Nothing is precomputed (except the massive model data structure of course).

I trace only two GI rays per pixel, and therefore, the resulting image must be heavily filtered in order to eliminate the extreme noise. While all the ray tracing is done on the CPU, the noise filter runs on the GPU and is implemented in CUDA. Since diffuse indirect lighting is quite low frequency, it is adequate to use low LODs for the GI rays."

Another interesting tidbit from the paper:

"By using LOD voxels, significantly higher frame rates can be achieved, with minimal loss of image quality, because ray traversals are less deep, memory accesses are more coherent, and intersections with voxels are free, contrary to triangles (the voxel fills its parent node, therefore, the intersection is equal to the already computed intersection with the node). Furthermore, the LOD framework can also reduce the amount of aliasing artifacts, especially in case of highly tessellated models"

The quality of the indirect lighting looks pretty amazing for just 2 (filtered) random samples per pixel and is completely noise-free, as can be seen in this picture from the paper

All the ray tracing is currently CPU based (the GI algorithm runs at 1-2 fps on a quad core cpu), but it would probably run in real-time at much higher framerates when implemented entirely on the GPU.