Relational Constraints for Point Distribution Models

Abstract

In this paper we present a new method for aligning point distribution models to noisy and unlabelled image data. The aim is to construct an enhanced version of the point distribution model of Cootes and Taylor in which the point-position information is augmented with a neighbourhood graph which represents the relational arrangement of the landmark points. We show how this augmented point distribution model can be matched to unlabelled point-sets which are subject to both additional clutter and point drop-out. The statistical framework adopted for this study interleaves the processes of finding point correspondences and estimating the alignment parameters of the point distribution model. The utility measure underpinning the work is the cross entropy between two probability distributions which respectively model alignment errors and correspondence errors. In the case of the point alignment process, we assume that the registration errors follow a Gaussian distribution. The correspondence errors are modelled using probability distribution which has been used for symbolic graph-matching. Experimental results are presented using medical image sequences.