Bounded-Error Interactive Ray Tracing

Ray tracing, which computes radiance, is traditionally regarded as an off-line rendering algorithm that is too slow for interactive use. In this paper, we present an interactive system that uses 4D interpolants to approximate radiance, while providing guaranteed error bounds. Our system exploits the object-space, ray-space, image-space and temporal coherence in radiance to accelerate ray tracing. Our system explicitly decouples the two operations of the ray tracer — shading computation and visibility determination at each pixel, which we call pixel assignment. Rendering is accelerated by approximating the shading computation while guaranteeing correct pixel assignment. Without any pre-processing, the system lazily collects 4D radiance samples, which are quadrilinearly interpolated to approximate radiance. An error predicate conservatively guarantees that the relative error introduced by interpolation is bound by a user-specified . The user can change this parameter to trade off performance for quality. The predicate guarantees the error bound by detecting discontinuities and non-linearities in radiance. If error cannot be guaranteed for an interpolant, the system adaptively refines the interpolant. Pixel assignment is accelerated using a novel algorithm that reprojects 4D interpolants to new viewpoints as the user’s viewpoint changes. Using reprojection, a fast scan-line algorithm achieves high performance without sacrificing image quality. This lazy interpolant system with reprojection speeds up ray tracers substantially for smoothly varying viewpoints. In this paper, we present the design, implementation and results for our system. We expect our techniques to be useful for interactively rendering general scenes as well as for batch off-line rendering processes.