University of Konstanz
Graduiertenkolleg / PhD Program
Computer and Information Science

Graduation Talks


Perceptual distances and texture registration for 3D models


Ioan Cleju, University Konstanz
Konstanz, Germany

date & place

Wednesday, 04.07.2007, 17:45 h
Room C252


Quantifying the perceived dissimilarity between 3D models and their simplified copies is necessary for assessing the quality of simplification and compression algorithms, and for Level-Of-Detail (LOD) management. This thesis examines several strategies used for the evaluation of objective distances with respect to user studies. We show that ordinal analysis on the LOD sequences does not provide enough data to differentiate among several objective distances. Popular evaluation strategies, such as those based on ratings, provide however ordinal data. We propose a new experimental setup that allows parametric evaluation of objective distances with respect to the user study. The case study included six objective distances and we found that all considered image-based measures were better than the geometric-based.

The second topic covered in this thesis is texture registration for 3D models. The common 3D acquisition pipeline considers geometry acquisition (by 3D scanning) and texture acquisition (by photographing) as two independent steps. The texture registration solves the 2D-3D mapping problem by recovering the parameters of the photo cameras. Commonly, the patches of the model are visible in several images. We propose to use this additional information to add registration objective functions between images. We define two types of objective functions, between each image and the surface model and for each pairs of images that sample a common patch of the surface, all based on mutual information. The mutual information does not need preprocessing and feature extraction, and is robust to varying illumination and surface reflectance. We propose a joint optimization model for registering all images. In various experiments we showed that the extended optimization approach is more robust with respect to the initialization and leads to increased accuracy of registration.