Efficient evidential indexing of three-dimensional models using prototypical parts

This thesis is concerned with efficient recognition of three-dimensional (3-D) objects using parametric part descriptions. The parametric shape models used are superquadrics, as recovered from depth data. The primary contribution of our research lies in a principled solution to the difficult problems of object part classification and model indexing. The novelty of our approach is in the use of a formal statistical approach for superquadric part classification and a formal evidential framework for model indexing. In addition, the method used for model indexing is amenable to the use of massive parallelism using a connectionist implementation of evidential semantic networks. A major concern in practical vision systems 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 prototypical part classes which we term protoparts. Each superquadric part recovered from the input is paired with the best matching protopart using precompiled class statistics. The features used by the classifier are the statistically most significant subset of parameters, computed using principal component analysis. The selected protoparts are used to index into the model database and the retrieved models are ranked by combining the evidence from the protoparts. We have implemented a working vision system based on the principles described above. Our work builds on a technique for fitting dense range data from simple 3-D scenes with superquadric models and uses a scheme for segmenting range data from complex 3-D objects into their constituent parts in terms of surface and volumetric primitives. Our contributions include a bootstrapping technique for dealing with small training databases, a protopart classifier and an evidential model indexing algorithm. To date, our vision system has been tested on a small database of bottles. Results, of recognition experiments with familiar, occluded and novel objects indicate the promise of our approach of using protoparts as model indexing keys