Rethinking multiple importance sampling for general and efficient Monte Carlo rendering
No Thumbnail Available
Date
2024-03-14
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Saarländische Universitäts- und Landesbibliothek
Abstract
Computer generated images are essential for many applications from art to engineering. Unfortunately, rendering such images is costly, with render times easily in the hours, days, or even weeks. On top of that, the demands regarding complexity and visual fidelity are ever rising. Consequently, there is an insatiable need for faster rendering. Efficient render times are often achieved through user intervention. For example, modifying the scene and removing difficult lighting effects can keep render times below an acceptable threshold. Also, algorithm parameters can be tuned manually. For instance, diffuse outdoor scenes are best rendered by unidirectional path tracing, while interiors featuring caustics benefit greatly from bidirectional sampling. Such manual tuning, however, is unfortunate as it puts much burden on the user and poses a hurdle for novices. In this thesis, we pave the way for more universal rendering algorithms with less need of user intervention. For that, we revisit multiple importance sampling (MIS), an essential tool to universalize rendering algorithms by combining diverse sampling techniques. We identify hitherto unknown shortcomings of MIS and propose practical solutions and improvements. As a tangible result, we achieve adaptive bidirectional rendering with performance never worse than unidirectional path tracing.
Description