On representation and invariant recognition of complex objects based on patches and parts

Abstract

This paper presents an outline of our view of 3D object modeling and recognition. The problem of interest is the recognition of which free-form object from a large possible set of articulated deformable objects in arbitrary positions is present in sensed data when clutter is also present and the objects are partially occluded. The modeling approach is to represent complex objects by patches or parts that individually capture significant local information and to geometrically relate them in order to provide the complete structure of an object. A representation that we use generally consists of implicit polynomial curves and surfaces for 2D data and 3D data, respectively. Recognition is based on geometric invariants-functions of patches or parts that capture the shape but are invariant to the geometric transformations. Since an object to be recognized may be a member of a class, or since the invariants for an object may take values in a class, recognition of class membership is necessary, and for this we use Bayesian recognizers. The technology of self and mutual invariants and Bayesian recognizers for implicit polynomial patches or parts is touched on in the paper. Another topic briefly discussed is representation by generalized cylinders within the framework of algebraic curves and invariants.