Face Cluster Radiosity

Abstract

An algorithm for simulating diffuse interreflection in complex three di-mensional scenes is described. It combines techniques from hierarchical radiosity and multiresolution modelling. A new face clustering technique for automatically partitioning polygonal models is used. The face clusters produced group adjacent triangles with similar normal vectors. They are used during radiosity solution to represent the light reflected by a complex object at multiple levels of detail. Also, the radiosity method is reformulated in terms of vector irradiance and power. To-gether, face clustering and the vector formulation of radiosity permit large sav-ings. Excessively fine levels of detail are not accessed by the algorithm during the bulk of the solution phase, greatly reducing its memory requirements relative to previous methods. Consequently, the costliest steps in the simulation can be made sub-linear in scene complexity. Using this algorithm, radiosity simulations on scenes of one million input polygons can be computed on a standard workstation. 1 Introduction The hierarchical radiosity algorithm in its various forms is probably the most promising radiosity method in existence. The best hierarchical radiosity methods, using clustering, permit scenes of moderate complexity (several hundred thousand input polygons) to be simulated in a few hours. Unfortunately, current radiosity techniques, even with clus-tering, use excessive memory and their speeds are not competitive with other, less real-istic rendering methods. We would like to be able to apply radiosity methods to the complex scenes common in special effects. Such scenes routinely use objects each em-ploying 100,000 polygons or more. We therefore seek an enhancement to the hierarchi-cal radiosity algorithm that will permit very complex scenes — scenes with millions of input polygons — to be economically simulated on a standard computer. One of the greatest difficulties with existing radiosity methods is that their memory use is at least linear in the number of input polygons. This is not a problem if the scene is small, but if the input polygons cannot fit in physical memory, the algorithm will thrash and performance will degrade dramatically. To deal with very complex scenes, we need methods which in practice have memory and time cost that is sub-linear in the number of input polygons. In this paper we describe the face cluster radiosity algorithm, a technique that achieves this goal. Its three main phases are preprocessing, solution, and postprocess-ing. Preprocessing converts the scene description into a multiresolution, hierarchical model. The time cost of this is super-linear in the number of input polygons, but pre-processing can be done on an off-line, object-by-object basis, so its memory costs are 1. {ajw|ph|garland}@cs.cmu.edu