HIERARCHICALLY ACCELERATED RAY CASTING FOR VOLUME RENDERING WITH CONTROLLED ERROR

Ray casting for volume rendering can be accelerated by taking large steps over regions where data need not be processed at a fine resolution. A new implementation is described that utilizes both a user-specified data importance and a high-level data model in appropriate regions to achieve acceleration. Previously reported work exploited high accumulated opacity, transparency, and nearly constant valued regions. This work generalizes homogeneity from nearly constant values to least-squares fits with a set of basis functions. The implemented set supports trilinear functions, and the framework supports other functions. Experience is reported with sequential and small-scale parallel runs. Comparisons are made with other methods on the bases of time, image quality, and sensitivity to error tolerance. Ray casting with zero error provides a reference image with which others are compared quantitatively. The implementation lives within a general system for multi-dimensional trees. For a given error tolerance, this ensures that the same resolution is used in all regions, by all hierarchical rendering methods, whether based on projection or ray-casting. Data-importance functions were found to be a significant source of acceleration. Image comparisons led to identification of the main source of image degradation in projection methods as being color interpolation rather than data interpolation. An improvement to projection methods is suggested, based on hardware texture maps.