Content-based 3D model retrieval based on volumetric extended gaussian image shape representation

Abstract

With the large collections of 3D models in the Internet, there is an increasing need to develop effective retrieval techniques for these models. Since traditional textual based retrieval methods lost their effectiveness when the models are not well annotated, many researchers have proposed a number of shape representations to describe the 3D models based on the 3D models’ content. One of the most challenging issues in content-based 3D model retrieval research is how to extract effcient feature vector from the original model with translation, scaling and rotation independence. In this thesis, I introduce a novel shape signature, called Volumetric Extended Gaussian Image (VEGI) for content-based 3D model retrieval. It directly captures the volumetric distribution of a 3D mesh model along the latitude-longitude direction without conventional pose normalization and guarantees translation and scaling invariance. Rotation invariance is accomplished by further calculating the spherical harmonic transform of this directional distribution, exploiting the norm of its spherical harmonic coeffcients as the shape descriptor. Due to the completeness and orthonormality properties of spherical harmonics, the VEGI descriptor also supports multi-resolution description of a 3D model, providing coarse to detailed reconstruction. Although spherical harmonic transform based methods can maintain the property of rotation independence, there is a pre-requisite for performing spherical harmonic transform: the surface sphere should be uniformly sampled. However, spherical harmonic transform based methods divide the spherical surface by sam¬pling the latitude and longitude direction uniformly which causes singularities in the two poles. In other words, the poles are over-sampled while the equator parts are under-sampled. Hence, I further proposed an improvement approach to fix this problem through incorporating a variant principal component analysis into the Volumetric Extended Gaussian Image (VEGI) to get Adaptive Volumetric Extended Gaussian Image (A-VEGI) shape representation. In order to reduce the search time, an indexing scheme based on Hierarchical Self Organizing Map (HSOM) will be established to improve retrieval effciency through hierarchical searching. We evaluate VEGI shape descriptor on a combined 3D head model database and public Princeton Shape Benchmark (PSB) database regarding computational aspects and shape retrieval performance. The experimental results show that our retrieval architecture not only has high descriptive and discriminative power, but also outperforms many existing methods which are dependent on the canonical alignment. Thanks to the excellent performance of Adaptive Volumetric Extended Gaussian Image (A-VEGI) shape representation, we developed a public online 3D model retrieval system for users to find the most similar 3D models to their queries.

Date of Award	15 Feb 2008
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Hau San WONG (Supervisor)