Structure Analysis Of Low-Quality Geometric Data Based On Prior Knowledge

Digital geometry processing is one of the main research area in computer graphics.Its traditional process is to acquire point cloud by high-precision scanning devices andthen perform mesh reconstruction, followed by shape analysis, manipulation and so on.With the development and popularization of scanning devices, rapid and large-scale ge-ometric data has emerged. However, it causes the generation of low-quality geometricdata, typically corrupted with severe noise and missing data. For such kind of data, tra-ditional geometry processing methods usually produce unsatisfactory results. The keyto process low-quality geometric data is to analyze and recover its underlying structureinformation. Our key observation here is that, although the low-quality geometric dataitself cannot provide enough information, the object it represents usually exhibits partic-ular kind of structure. Therefore, our key idea is to introduce suitable prior knowledgeand propose efective structure analysis methods based on it. Furthermore, the structureanalysis results can be applied to applications like reconstruction, manipulation.In this dissertation, we investigate the problem of structure analysis of low-qualitygeometric data and propose several efective solutions for some typical low-quality geo-metric data. Our main contributions are as follows:We propose a novel rectilinear mixture model as the prior structure model and anadaptive structure analysis method for point cloud of urban facades, which can au-tomatically generate a hierarchical representation. Compared with previous work,our method is more flexible and can uniformly handle various urban facades.We propose a part-assembly approach to analyze the structure of point cloud repre-senting man-made objects. Our method is able to compose suitable parts on the flywith respect to the input scan data. Unlike previous methods, it does not heavilyrely on suitable examplar models and only uses a small-scale shape repository.We propose a harmonic-field based approach to analyze the volume structure oflow-quality meshes. By adopting prior knowledge from physics, our method ismore robust compared with previous work and can handle input like triangle soup.