Training a general purpose deformable template

We propose a general purpose deformable template that can be used to recognize images of different classes of objects. We demonstrate a method by which prior information about a class of objects can be systematically incorporated into the template to form a model for the class. This is done by having two separate levels of parameterization: a set of prior variables whose values specify a prior model, and a set of configuration variables which can be varied to match a model to a specific image. When trained on a set of faces, the template successfully distinguished between face and non-face, as well as between different faces