Range Image Interpretation of Mail Pieces with Superquadrics

Although mail pieces can be classified by shape into parallelopipeds and cylinders, they do not conform exactly to these perfect geometrical shapes due to rounded edges, distorted comers, and bulging sides. Segmentation and classification of mail pieces hence cannot rely on a limited set of specific models. Variations and deformations of shape can be conveniently expressed when using superquadrics. We show how to recover superquadric models for mail pieces and segment the range image at the same time.