Perspective Matching Using the EM Algorithm

This paper describes a new approach to perspective matching which simultaneously exploits both rigidity-structure and point distribution information. The structural component of the model is represented by a Delaunay triangulation of the point-set. The point-distribution model is represented by a perspective deformation of the point-set. Model-matching is realised using a variant of the EM algorithm. This involves coupling the correspondence matching of the Delaunay triangulation to the recovery of the point deformation parameters. We use a Bayesian consistency measure to gauge the relational structure of the point correspondences. Maximum-likelihood point deformation parameters are estimated using a mixture-model defined over the point error-residuals. In effect, the Bayesian consistency measure is used to weight the contributions to a mean-squares error-criterion. The method is evaluated on matching 2D objects under varying pose.