GPU accelerated isosurface volume rendering using depth-based coherence

With large scientific and medical datasets, visualization tools have trouble maintaining a high enough frame-rate to remain interactive. In this paper, we present a novel GPU based system that permits visualization of isosurfaces in large data sets in real time. In particular, we present a novel use of a depth buffer to speed up the operation of rotating around a volume data set. As the user rotates the viewpoint around the 3D volume data, there is much coherence between depth buffers from two sequential renderings. We utilize this coherence in our novel prediction buffer approach, and achieve a marked increase in speed during rotation. The authors of [Klein et al. 2005] used a depth buffer based approach, but they did not alter their traversal based on the prediction value. Our prediction buffer is a 2D array in which we store a single floating point value for each pixel. If a particular pixel p_ij has some positive depth value d_ij, this indicates that the ray R_ij, which was cast through p_ij on the previous render, intersected an isosurface at depth d_ij. The prediction buffer also handles three special cases. When the ray R_ij misses the isosurface, but hits the bounding box containing the volume data, we store a negative flag value, d_{hitBoxMissSurf} in p_ij. When R_ij misses the bounding box, we store the value d_missBox. Lastly, when we have no prediction stored in the buffer, we store the value d_noInfo.