Adaptive sampling of intersectable models exploiting image and object-space coherence

We present a sampling strategy and rendering framework for intersectable models, whose surface is implicitly defined by a black box intersection test that provides the location and normal of the closest intersection of a ray with the surface. To speed up image generation despite potentially slow intersection tests, our method exploits spatial coherence by adjusting the sampling resolution in image space to the surface variation in object space. The result is a set of small, view-dependent bilinear surface approximations, which are rendered as quads using conventional graphics hardware. The advantage of this temporary rendering representation is two-fold: First, rendering is performed on the GPU, leaving CPU time for ray intersection computation. As the number of primitives is typically small, complex per vertex or per fragment programs can be used to achieve a variety of rendering effects. Second, bilinear surface approximations are derived from the geometry and can be reused in other views. Here, graphics hardware is exploited to determine the subset of image space in need of re-sampling. We demonstrate our system by ray casting an implicit surface defined from point samples, for which current ray-surface intersection computations are usually too slow to generate images at interactive rates.