Obtaining A Part-based 3-D Object Representation from Range Data

We examine two methods of obtaining a part-based 3-D object representation from range images. A small number (12) of shape types based on geolis are used as 3-D part primitives. The first method fits superquadric models to the depth data and then distinguishes between the shape classes by using statistical classifiers. Features for the classification are derived from the estimated superquadric parameters. Fairly good results are obtained using synthetic images. A second proposed method takes a differential-geometric approach. Surface segmentation of the range image is performed based on curvatures, and the part primitive is recognized based on the surface types and their adjacencies. Surfaces types and adjacencies of the part primitives also determine a small number of characteristic views to be stored in a multi-view representation of the primitives. This is to be used in cases where the surface types and adjacencies in the image yield more than one possibility for the part primitive.