Real-time Path Traced Global Illumination for Games by Path Space Filtering

Abstract

This thesis is about approximating global illumination for games in real-time using ray tracing. We have found that not pre-multiplying the albedo during tracing to avoid having to divide it out is numerically more stable due to floating point arithmetic and avoids a special case where information is lost. Furthermore, we investigate the benefits of using world-space neighbourhood search in path space filtering with a screen-space search, called screen-space path space filtering (SSPSF). In addition, we implement and test multiple extensions of the hashed path space filtering (HPSF) method that were proposed in the original report “Fast Path Space Filtering by Jittered Spatial Hashing” by Binder et al. [1]. Based on HPSF, we develop a novel technique called recursive hashed path space filtering (nHPSF), which filters at multiple selected bounces along a light transport path, which allows for reducing variance. It can transport information between hash cells during scattering as well, which further increases efficiency. Our main result is an algorithm that amortizes the cost of tracing long light transport paths over multiple frames and combines it with the nHPSF technique. It hence is named Amortized nHPSF (A-nHPSF). The thesis includes an extensive survey of the state of the art in the industry and reviews important related literature. It closes with a set of important and promising directions for future research.