Shape recovery of mail pieces using deformable models

Franc Solina and Ruzena Bajcsy Department of Computer and Information Science University of Pennsylvania, Philadelphia PA ABSTRACT A system for classification of mail pieces using range images is proposed. Position, orientation, size and shape parameters of volumetric models of single mail pieces are recovered using least squares minimization of a fitting function. The models are super- quadrics with global deformations. The recovered parameters serve for classification. I. INTRODUCTION Automatic sorting of all mail pieces is a difficult problem because mail pieces differ widely in size and shape. In general, one has to know the location, orientation, size and shape of mail pieces to be able to initiate the right handling procedure. Computer vision as a method for locating and describing of objects without direct physical contact seems to be the right approach to do that in a fast and reliable manner. Computer vision has been successfully applied in many industrial applications. The methods used in the majority of these industrial vision systems, however, cannot be applied to the problem of mail piece classification. Most so called "model-based" object recognition vision systems use a set of rigid and precise models for all objects that are expected to be found in the scene [3]. Based on detected local features in the scene, these models are projected onto the image, to find