Object Recognition: Modelling and the Interface to a Control Strategy for Matching

Abstract

Amodelling system for object recognition and pose estimation is presented in this work, based on approximating the aspect/appearance graph of arbitrary rigid objects for a spherical viewing surface using simulated image data. The approximation is achieved by adaptively subdividing the viewing sphere starting with an icosahedral tessellation and iteratively decreasing the patch size until the desired resolution is reached. The adaptive subdivision is controlled by both the required resolution and object detail. The decision whether a patch should be divided is based on a similarity measure, which is obtained from applying graph matching to attributed relational graphs generated from image features.

Patches surrounded by similar views are grouped together and reference classes for the aspects are established. The reference classes are indexed by contour types encountered in the views within the group, where the contour types are computed via unsupervised clustering performed on the complete set of contours for all views of a given object.

Classification of an unknown pose is done efficiently via simple or weighted bipartite matching of the contours extracted from the unknown pose to the equivalence classes. The best suggestions are selected by a scoring scheme applied to the match results.

Themodelling system is demonstrated by experimental results for a number of objects at varying levels of resolution. Pose estimation results from both synthetic and real images are also presented.