View-based techniques for 3D model retrieval

Abstract

With the increase in the number of available 3D models, the ability to accurately and efficiently search for 3D models is crucial in many applications such as Computer-Aided Design (CAD), on-line 3D model shopping and 3D game, movie and animation production. As a result, 3D model retrieval has become an important research area. In recent years, several typical algorithms that extract different types of 3D model features have been proposed. However, 3D model feature supporting multi-modal queries such as 3D models and 2D sketches is an important research direction which has little related work. In addition, 3D normalization is an important process in 3D model retrieval to extract rotation-dependent features and currently there still exists much room in terms of alignment accuracy and consistency. In this thesis work, we propose several algorithms to contribute solutions for the above issues. Motivated by the mechanism of human perception and multi-view vision, that is 3D shape information of a 3D object can be obtained based on multiple views, together with the retrieval performance comparison of previous retrieval work as well as the verifications of our proposed algorithms, we adopt a view-based approach which extracts features based on the rendered views of a 3D model. The first part of our work is dealing with 3D pose normalization. A novel Minimum Projection Area-based (MPA) alignment method is proposed for pose normalization based on the idea of successively finding two perpendicular principal axes with minimum projection area. Experimental results demonstrate that MPA has a good performance in finding accurate axes; can robustly find a consistent pose for similar models and outperforms PCA, CPCA, and NPCA in terms of 3D model retrieval performance. Next, we propose a view-based 3D model feature named view context to support both Query-by-Model and Query-by-Sketch retrieval. The view context of a particular view i This thesis could not be presented in this form without the help and support from so many people who are gratefully acknowledged here. Firstly, I would like to take this opportunity to appreciate Dr. Henry Johan, my supervisor, for his so much precious time and continuous effort dedicated for training me of doing good research. His invaluable suggestions, honest criticism and kind encouragement are constant sources of inspiration throughout the research. His personal enthusiasm, highly professional academic standards and serious attitude to research work will guide and benefit me in the future. I am also grateful to Prof. Seah Hock Soon for providing me a good research environment, and Dr. He Ying, Dr. Qian Kemao and Dr. Zheng Jianmin for their invaluable lectures and advices. Deep thanks also go to Mr. Budianto Tandianus and Mr. Nicholas Mario Wardhana for their help and the useful discussions during our weekly research meetings with our supervisor. Here, I also want to acknowledge Mr. Wei Yuanmin and Mr. Iskandarsyah in relation to the collaboration in 3D model alignment research.