Acceleration techniques for monte carlo ray tracing

Abstract

Monte Carlo (MC) ray tracing has been considered as the most effective technique to produce a variety of realistic visual effects. However, its performance tends to be very slow since a lot of ray samples should be generated until we achieve converged images. In this thesis, we propose three novel techniques to accelerate performance of MC ray tracing. We fist develop a ray reordering framework based on a novel ray ordering measure hit point heuristic to compute a cache-coherent access pattern ofray traversals on acceleration hierarchies. In addition to the cache optimization technique, we propose an efficient and robust image-space denoising method for reducing noise generated by MC ray tracing while preserving image features. Our denoising is built upon a novel edge-stopping function virtual flash image which captures a wide variety of image features without taking additional ray samples. Furthermore, we present a new image-space adaptive rendering method based on locally weighted regression. In our adaptive framework, we locally guide ray budgets on high error regions, and estimate optimal filtering bandwidths for each rendering feature in terms of minimizing filtering errors. We have demonstrated that the proposed acceleration techniques improve the performance of MC ray tracing using differentrealistic benchmarks compared to state-of-the-art methods.