Parallel Evidence-Based Indexing of Complex Three-Dimensional Models Using Prototypical Parts and Relations (Dissertation Proposal)

This proposal is concerned with three-dimensional object recognition from range data using superquadric primitives. Superquadrics are a family of parametric shape models which represent objects at the part level and can account for a wide variety of natural and man-made forms. An integrated framework for segmenting dense range data of complex 3-D objects into their constituent parts in terms of bi-quadric surface patches and superquadric shape primitives is described in [29]. We propose a vision architecture that scales well as the size of its model database grows. Following the recovery of superquadric primitives from the input depth map, we split the computation into two concurrent processing streams. One is concerned with the classification of individual parts using viewpoint-invariant shape information while the other classifies pairwise part relationships using their relative size, orientation and type of joint. The major contribution of this proposal lies in a principled solution to the very difficult problems of superquadric part classification and model indexing. The problem is how to retrieve the best matched models without exploring all possible object matches. Our approach is to cluster together similar model parts to create a reasonable number of prototypical part classes (protoparts). Each superquadric part recovered from the input is paired with the best matching protopart using precomputed class statistics. A parallel, theoretically-well grounded evidential recognition algorithm quickly selects models consistent with the classified parts. Classified part relations (protorelations) are used to further reduce the number of consistent models and remaining ambiguities are resolved using sequential top-down search.