An empirically derived system for high-speed shadow rendering

Abstract

Shadows have captivated humanity since the dawn of time; with the current age being no exception – shadows are core to realism and ambience, be it to invoke a classic Baroque interplay of lights, darks and colours as the case in Rembrandt van Rijn’s Militia Company of Captain Frans Banning Cocq or to create a sense of mystery as found in film noir and expressionist cinematography. Shadows, in this traditional sense, are regions of blocked light – the combined effect of placing an object between a light source and surface. This dissertation focuses on real-time shadow generation as a subset of 3D computer graphics. Its main focus is the critical analysis of numerous real-time shadow rendering algorithms and the construction of an empirically derived system for the high-speed rendering of shadows. This critical analysis allows us to assess the relationship between shadow rendering quality and performance. It also allows for the isolation of key algorithmic weaknesses and possible bottleneck areas. Focusing on these bottleneck areas, we investigate several possibilities of improving the performance and quality of shadow rendering; both on a hardware and software level. Primary performance benefits are seen through effective culling, clipping, the use of hardware extensions and by managing the polygonal complexity and silhouette detection of shadow casting meshes. Additional performance gains are achieved by combining the depth-fail stencil shadow volume algorithm with dynamic spatial subdivision. Using this performance data gathered during the analysis of various shadow rendering algorithms, we are able to define a fuzzy logic-based expert system to control the real-time selection of shadow rendering algorithms based on environmental conditions. This system ensures the following: nearby shadows are always of high-quality, distant shadows are, under certain conditions, rendered at a lower quality and the frames per second rendering performance is always maximised.