Aggressive Visibility Pre-processing with Adaptive Sampling

At the expense of a small error in visibility classification, we remove all invisible polygons. Thresholding and heuristics allows fine control over the behaviour of this error. Our technique is applicable to both concave and convex polygons. It exhibits sublinear computational complexity in the number of scene polygons and logarithmic complexity in the number of cells, while effectively exploiting graphics hardware. The technique is classified as an aggressive from-region method in that it falsely excludes a small subset of visible polygons and estimates visibility on a per cell basis. A kd-tree hierarchy of visibility cells is built by sampling visibility across their surfaces adaptively. Sampling is guided by a novel error heuristic and produces, by adaptive sub-division, a quad-tree like structure. We have applied our technique to both a standard building scenes and a highly complex natural scene. The results demonstrate that significant culling (94.5% average) with low error rates (0.687% average) can be achieved with such scenes in a reasonable period of time.