Shape Estimation and Object Classification in Images Using Geometric Priors

We propose a method for modeling and incorporating prior shape information for Bayesian shape estimation from images. In this approach, shapes are treated as elements of an infinite-dimensional, non-linear, quotient space. Prior probability models on shape classes are defined and computed intrinsically on the tangent bundle of this shape space. MAP shape estimation is posed as a problem of gradient-based energy minimization where this energy has contributions from the image data and the prior model.