Ph.D Thesis, Department of Computer Science, KU Leuven, Celestijnenlaan 200A, 3001 Heverlee, Belgium, September 2002.
Computer-generated realistic images are being used increasingly in applications such as architecture, lighting design and visual effects in the movie industry. Therefore, realistic image synthesis, the subject of this dissertation, has remained one of the most important research areas in computer graphics for many years. Realistic image synthesis takes as input a description of a three-dimensional scene and a virtual camera looking upon that scene. By simulating the light transport, the intensity of each pixel in the image as seen by the camera can be computed. Light transport can be described by an integral equation, often called the rendering equation. Algorithms that compute all possible interreflections of light in the scene and thus provide a full solution to the rendering equation, are called global illumination algorithms. Such algorithms produce images that are virtually indistinguishable from a real photograph, but unfortunately they are slow. In this dissertation we present several new techniques to construct more robust and more efficient global illumination algorithms. Our work focuses on Monte Carlo algorithms and multi-pass methods. Monte Carlo algorithms statistically estimate integrals and are very well suited for solving transport problems. Multi-pass methods combine several rendering algorithms into a single method. A good multi-pass configuration will preserve the strengths of each individual component, while avoiding its weaknesses. First, new techniques for designing better multi-pass methods are presented, including weighted multi- pass methods that weight the contributions of different algorithms automatically in a provably good manner. Second, path differentials are introduced. They allow to augment light transport paths with neighborhood information which is useful in many global illumination algorithms. Finally a technique is presented to reduce the memory requirements and to control the error in photon mapping, an efficient and popular global illumination algorithm. All these techniques result in more robust and efficient global illumination algorithms and contribute to the increasing use of these algorithms in computer graphics applications.
