Abstract:

Bibtex:

@PHDTHESIS{w-pbmrr-04,
   AUTHOR = {Michael Wand},
   TITLE = {Point-Based Multi-Resolution Rendering},
   SCHOOL = {Department of computer science and cognitive science, University of T{\"u}bingen}
   INSTITUTION = {Wilhelm Schickard Institute for Computer Science, Graphical-Interactive Systems (WSI/GRIS), University of T{\"u}bingen},
   YEAR = {2004},
   KEYWORDS = {GRIS},
   ABSTRACT = {This thesis describes a new rendering paradigm for handling complex scenes, point-based multi-resolution rendering. The basic idea is to approximate the appearance of complex scenes using a small set of surface sample points. Using hierarchical data structures, the sampling process can be performed in time mostly independent of the scene complexity. This allows an efficient display of highly complex scenes. The thesis proposes different variants of sampling data structures that are useful in different application scenarios, including a variant for handling animated scenes (general keyframe animations). Two different rendering approaches are described: The first approach is a real-time forward mapping algorithm, being a point-based generalization of z-buffer rendering. In contrast to conventional z-buffer rendering, the point-based multi-resolution algorithm can render scenes consisting of trillions of primitives at real-time frame rates while maintaining a comparable rendering quality. The second approach is a backward mapping (i. e. raytracing) algorithm that aims at offline rendering. It is able to compute shadows, reflections, and refractions. It uses a hierarchy of prefiltered sample points to provide efficient antialiasing. Additionally, classic distributed raytracing effects such as soft shadows, depth-of-field, or blurry reflections can be approximated efficiently. In comparison with classic stochastic raytracing techniques, the new algorithm provides noise-free renderings at lower costs than stochastic oversampling for scenes of high variance. The image quality is roughly comparable to that of the classic approach; only a small bias is observed. The thesis provides a theoretical analysis of the sampling and rendering process. Upper bounds for the rendering time are established. For the randomized components of some of the algorithms, analytical lower bounds for the failure probability are derived, showing that arbitrarily high confidence probabilities can be achieved at a small increase of computational costs. An analysis of oversampling properties of different sampling and stratification strategies allows a quantitative comparison, needed to choose the best technique for a certain application. A prototype implementation is presented. The influence of different algorithmic parameters is evaluated empirically and compared to theoretical predictions. Practical applications of the proposed algorithms comprise real-time walkthroughs of highly complex static scenes as well as real-time visualizations of large crowd animations such as a herd of animals or a football stadium with ten thousands of animated football fans. In addition, dynamic modifications of the data structure as needed for interactive editing is examined. Finally, extensions to volume rendering, out-of-core rendering, sound rendering, and simulation of caustics from area light sources are discussed briefly. Overall, the presented techniques extend the possibilities for rendering of highly complex scenes to areas that could not be treated before with comparable efficiency.}
}