Principal Manifolds and Bayesian Subspaces for Visual Recognition