Efficiency of Simple Shape Descriptors

Characterizing objects by their shape is a critical part in many computer vision applications. Several theoretically interesting approaches exist which, however, remain computationally too expensive. The goal of this paper is to emphasize the value of primitive shape descriptors. A set of simple descriptors is used as a feature vector in order to group similar objects together. Each descriptor is discussed separately and discriminatory power of their combination is demonstrated on irregular sample objects. The advantages of using simple shape descriptors are obvious: faster calculation and more general applicability. Sometimes \general ap-plicability" implies poor performance in a speciic application. A neural network method (Self-Organizing Map) is used in adapting a system to an application-speciic shape space. The results of the experiments are good. It is shown that fairly simple shape descriptors can be exibly used in complex recognition tasks involving irregular objects.