Symmetry Improved Photon Maps

Abstract

In recent years there has been a tremendous effort in improving the performance of various global illumination algorithms. Most of this effort has been focused on maximizing the utilization of available CPU and RAM resources, as well as improving the parallel performance of the algorithms. This thesis provides an alternative approach towards improving global illumination by looking at similar areas within the image. A lot of synthetically generated images used in practice exhibit a large amount of repetitive or smooth areas. Combined with the recursive nature of global illumination this suggests that a reuse strategy might, in fact, prove very effective. This thesis proposes a novel approach towards tackling the global illumination problem. We compute an approximate global illumination solution using photon maps and reuse the computed illumination in different areas of the scene to improve the solution. The scene is separated into symmetrical segments and parts of the scene, that share similar radiance are identified. By using the symmetrical relationships between the segments, these areas are then filtered with a version of the non-local means filter that works on photon maps. An analysis of the visual quality of the resulting images as well as a study of convergence behaviour using mean square error is provided. We show that a reuse strategy, based on local geometric symmetry, can be a viable approach toward improving global illumination algorithms as well as provide a stepping stone for further research in the area.