Date of Degree


Document Type


Degree Name



Computer Science


Ioannis Stamos

Subject Categories

Computer Sciences


parametrization, reconstruction, segmentation, simplification


The laser scanner is nowadays widely used to capture the geometry of art, animation maquettes, or large architectural, industrial, and land form models. It thus poses specific problems depending on the model scale. This thesis provides a solution for simplification of triangulated data and for surface reconstruction of large data sets, where feature edges provide an obvious segmentation structure. It also explores a new method for model segmentation, with the goal of applying multiresolution techniques to data sets characterized by curvy areas and the lack of clear demarcation features. The preliminary stage of surface segmentation, which takes as input single or multiple scan data files, generates surface patches which are processed independently. The surface components are mapped onto a two-dimensional domain with boundary constraints, using a novel parametrization weight coefficient. This stage generates valid parameter domain points, which can be fed as arguments to parametric modeling functions or surface approximation schemes. On this domain, our approach explores two types of remeshing. First, we generate points in a regular grid pattern, achieving multiresolution through a flexible grid step, which nevertheless is designed to produce a globally uniform resampling aspect. In this case, for reconstruction, we attempt to solve the open problem of border reconciliation across adjacent domains by retriangulating the border gap between the grid and the fixed irregular border. Alternatively, we straighten the domain borders in the parameter domain and coarsely triangulate the resulting simplified polygons, resampling the base domain triangles in a 1-4 subdivision pattern, achieving multiresolution from the number of subdivision steps. For mesh reconstruction, we use a linear interpolation method based on the original mesh triangles as control points on local planes, using a saved triangle correspondence between the original mesh and the parametric domain. We also use a region-wide approximation method, applied to the parameter grid points, which first generates data-trained control points, and then uses them to obtain the reconstruction values at the resamples. In the grid resampling scheme, due to the border constraints, the reassembly of the segmented, sequentially processed data sets is seamless. In the subdivision scheme, we align adjacent border fragments in the parameter space, and use a region-to-fragment map to achieve the same border reconstruction across two neighboring components. We successfully process data sets up to 1,000,000 points in one pass of our program, and are capable of assembling larger scenes from sequential runs. Our program consists of a single run, without intermediate storage. Where we process large input data files, we fragment the input using a nested application of our segmentation algorithm to reduce the size of the input scenes, and our pipeline reassembles the reconstruction output from multiple data files into a unique view.