Automatic shape model acquisition based on a generalization of convex/concave structure

This paper proposes a novel method for acquiring a shape model from shape samples of the same class. In conventional approaches, a priori knowledge and unique parts segmentation are needed because they perform generalization of predefined symbolic concepts. Moreover, they cannot form visual shape models because the model is represented as parts and parts relationships in the form of graph models such as semantic networks and multilevel graphs. The approach herein can directly acquire a visual model from a few samples without a priori knowledge because it is based on the generalization of multiscale convex/concave structure of a class of shapes. First, an optimal scale convex/concave segment structure common to shape samples using a multiscale segment matching method is extracted. Then the extracted segments are integrated to create generalized shapes. Finally, an importance measure based on the optimal scale is assigned to each convex/concave segment of the generalized shapes. That is, the proposed shape models consist of conjunctive descriptions as for the convex/concave segment structures, disjunctive descriptions as for the convex/concave segment forms, and the importance measure assigned to each segment of the shape model. A shape recognition method using the acquired models also is presented. Experimental results show the usefulness of the proposed method.