Automatic shape model acquisition using multiscale segment matching

A novel method for acquiring a shape model from shape samples of the same class is proposed. A critical point is that the method requires no prior knowledge of the class. Multiscale representations are first obtained using curvature scale space filtering to gain inflection point correspondence between consecutive smoothed shapes. The multiscale samples are then matched to extract the convex/concave structure common to the class. The matching is invariant under translation, rotation, and size change. Finally, generalized samples composing a model are generated by smoothly connecting the matched convex and concave segments. Experimental results show that the resulting model is useful for shape recognition