Abstract
In this thesis, we propose a new 3D shape descriptor to extract representative shape features from 3D models for 3D model retrieval. Our approach computes information from the local shape geometrical structures and employs the bag-of-features representation. First of all, we build the descriptor by creating a cube grid according to the local shape information of one vertex and intersecting this cube with the surface of this 3D model. For the intersected cube, we apply the distance transform to reduce the within-class variation. The Distance-Transform-Cubes (DTCs) sets created by the vertices from many 3D models are then collected for building a DTCs’ dictionary by k-means clustering. As the bag-of-feature method, we finally represent a 3D model based on the DTCs’ dictionary by a 1D histogram shape feature called Distance Transform Cube Histogram (DTCH) to perform the retrieval task. In addition, we combine the DTCH with geometrical information, such as distance and angle computed from the 3D model to generate higher-dimensional DTCH-based shape descriptiors. Our experimental results show the superior performance by using the new descriptors for 3D model retrieval.