Bootstrap Based Surface Reconstruction

Abstract

Surface reconstruction is one of the main research areas in computer graphics. The goal is to find the best surface representation of the boundary of a real object. The typical input of a surface reconstruction algorithm is a point cloud, possibly obtained by a laser 3D scanner. The raw data from the scanner is usually noisy and contains outliers. Apart from creating models of high visual quality, assuring that a model is as faithful as possible to the original object is also one of the main aims of surface reconstruction.

Most surface reconstruction algorithms proposed in the literature assess the reconstructed models either by visual inspection or, in cases where subjective manual input is not possible, by measuring the training error of the model. However, the training error underestimates systematically the test error and encourages overfitting.

In this thesis, we provide a method for quantitative assessment in surface reconstruction. We integrate a model averaging method from statistics called bootstrap and define it into our context. Bootstrapping is a resampling procedure that provides statistical parameter. In surface fitting, we obtained error estimate which detect error caused by noise or bad fitting. We also define bootstrap method in context of normal estimation. We obtain variance and error estimates which we use as a quality measure of normal estimates. As application, we provide smoothing algorithm for point clouds and normal smoothing that can handle feature area. We also developed feature detection algorithm.