Machine Learning Approaches to 3D Model Classification

A desirable 3D model classification system should be equipped with qualities such as highly correct classification accuracy, good enough classification speed, robustness to model noises and etc. Our objectives of this thesis are to 1)analyze different shape descriptors to concisely capture the geometric information of a shape in a finite feature descriptor 2) and train multiple multi-class supervised machine learning classifiers and evaluate their effectiveness in classifying 3D shapes. At first, some of the most representative shape descriptors of different categories are discussed. Then, we also review some mathematical background of supervised machine learning algorithms that will be implemented. We conduct our experimentations on our own created 3D database, in which all models are downloaded from Google 3D warehouse. We compare classification performances by applying different machine learning algorithms combined with three shape feature extraction methodologies: Light-field descriptor(LFD)+Angular radial transform descriptors(ARTD),Light-field descriptors (LFD)+ 2D Zernike descriptors(2D ZD), 3D Zernike descriptors(3D ZD). Our experimental results show that above three presented shape descriptors are effective in classifying 3D models. Also, we also extend the classification of virtual models to the real world point cloud models and evaluate its performance. Conclusions and possible future work directions are presented at the end of this thesis.