Abstract
This paper describes algorithms for predictive database management used in The UC Berkeley Building Walkthrough System. The algorithms forecast a range of possible observer viewpoints during upcoming frames and use precomputed cell visibility information to determine a set of objects likely to be visible to the observer in the near future. For each of these objects, detail elision techniques determine which levels of detail must be stored in a memory resident cache. Cache management algorithms determine which objects to load into memory from disk, and which to replace when the cache is full. Using these algorithms, the system is able to maintain real-time frame rates during interactive visualization of large building models with furniture and radiosity illumination.
