Hardware Accelerated Visibility Preprocessing using Adaptive Sampling

Abstract

We present a novel aggressive visibility preprocessing technique for general 3D scenes. Our technique exploits commodity graphics hardware and is faster than most conservative solutions, while simultaneously not overestimating the set of visible polygons. The cost of this benefit is that of potential image error. In order to reduce image error, we have developed an effective error minimization heuristic. We present results showing the application of our technique to highly complex scenes, consisting of many small polygons. We give performance results, an in depth error analysis using various metrics, and an empirical analysis showing a high degree of scalability. We show that our technique can rapidly compute from-region visibility (1hr 19min for a 5 million polygon forest), with minimal error (0.3% of image). On average 91.3% of the scene is culled.