Statistical segmentation for computer graphics

Abstract

Data segmentation involves the analysis and grouping of conceptually meaningful sections from the input, so that the individual segments contribute to an higher order understanding of the data as a whole. Usually, segmentation is a crucial first step towards further data processing. This dissertation discusses two segmentation techniques, both based on a hierarchical statistical analysis. In the first part a technique is developed to segment motion vectors generated from moving images into independently moving entities. This segmentation is then used to automatically regenerate missing frames. This is especially useful for historical movies, as these often have been the subject of severe degradations, due to the susceptibility of celluloid to several degradations. The second part deals with the planar segmentation of three dimensional point clouds. To this end, a statistical method is developed, based on principal components analysis and graph theory. The planar segmentation is then used in a number of industrial inspection problems, motivated by examples from the sheet metal bending industry. The dissertation is concluded with a hybrid rendering paradigm, combining both points and planar geometry into a single rendering framework.