Learning Function Based Classification from 3D Imagery
Michael Pechuk, Octavian Soldea, and Ehud Rivlin
Computer Vision and Image Understanding, Volume 110, Issue 2, May, 2008, pp. 173-191
We propose a novel scheme for using supervised learning for function-based classiﬁcation of objects in 3D images. During the learning process, a generic multi-level hierarchical description of object classes is constructed. The object classes are described in terms of functional components. The multi-level hierarchy is designed and constructed using a large set of signature-based reasoning and grading mechanisms. This set employs likelihood functions that are built as radial-based functions from the histograms of the object instances. During classiﬁcation, a probabilistic matching measure is used to search through a ﬁnite graph to ﬁnd the best assignment of geometric parts to the functional structures of each class. An object is assigned to the class that provides the highest matching value. Reuse of functional primitives in different classes enables easy expansion to new categories. We tested the proposed scheme on a database of about 1000 different 3D objects. The proposed scheme achieved high classiﬁcation accuracy while using small training sets.