Iterative adaptive sampling for accurate direct illumination.

Master's thesis, Cornell University, August 2004.

This thesis introduces a new multipass algorithm, Iterative Adaptive Sampling, for effciently computing the direct illumination in scenes with many lights, including area lights that cause realistic soft shadows. Real world architectural scenes frequently contain large numbers of lights; however many current algorithms do not scale well in performance when rendering these types of scenes. Our algorithm is based upon an observation that although many hundreds of lights may contribute to the illumination of a single image, much lower lighting complexity typically exists on a localized basis within subsections of the image. Since the predominant cost of computing the direct illumination at a point is the testing of light source visibility, our algorithm works to exploit this observation of low localized lighting complexity to reduce the number of visibility tests (shadow rays) needed to accurately render each pixel. This reduction of shadow rays is made possible by sampling light sources in proportion to their actual contribution to a pixel's luminance value. We do this by iteratively modifying a probability density function (PDF) until it adaptively captures the local lighting configuration. We use sample data collected during rendering as feedback to drive the optimization of the PDF. Our algorithm takes advantage of coherence in image space by aggregating sample data on both a perpixel and per-block level as well as coherence in world space by aggregating sample data on light clusters. We have tested this algorithm on several complex lighting environments and demonstrated roughly an order of magnitude improvement over standard procedures.

This paper is available as a PDF file Don04.pdf ( 21M).