Levels of Detail: The Missing Link in Global Illumination

However in presence of complex or curved surfaces these techniques sometimes lacks precision. On one hand, using coarse representation for each surface may be very inaccurate; but on the other hand, using fine representation affects performance.
The first technique (clustering) reduces numbers of links but may lack precision since it does not account precisly for the orientations of the patches within the emitting cluster relative to the location of the receiving one. For scenes containing small objects represented by a large collection of polygons, clustering algorithms are not particularly well-behaved.
As for the second approach (partitioning), it is efficient since it partitions large environments into 3D cells and needs only a subset of cells for computing illumination at each step of the resolution process. Nevertheless, the computational complexity still remains high when the cells contain many objects with highly detailed geometry.
The new HR algorithms described in this talk, present solutions to these problems by making use of Levels of Detail (LOD) that approximate the geometry of the object surfaces in the scene.
Our methods are not only a good alternative to clustering and partitioning techniques but are also a perfect complement to them. We will first present the use of LODs for the diffuse case (Hierarchical Radiosity) and then the general case with arbitrary reflectances (3-Point-Transport Algorithm).