Feature points for a polygonal representation of silhouettes

The pixels of a silhouette located where an observer focuses his attention are useful feature points, since they can be regarded as vertices of a polygonal representation of the silhouette. The proposed method allows one to find feature points relative to details of different size, without explicitly exploiting curvature information. Moreover, no input parameters are required. The initial set of candidate vertices of the polygon is extracted from the chain-coded silhouette. Then, the set is iteratively modified by evaluating the perceptual relevance of the triangles defined by any triple of consecutive vertices. Geometrical features deduced from the set of ali the triangles are used to this purpose. The process terminates as soon as the organization of the current set of candidate vertices is regarded as perceptually significant.