PDF version of this document
Few conclusions were listed already, but what follows is a concise and generic summary.
Registration using statistical models is viable. It has clear drawbacks because it is (1) slow; (2) able to drift away (and destroy data) and (3) complex. When data registration is performed with the methods proposed, models of deformation are produced and correspondences identified.
Model-based algorithms result in appearance models whose determinant is by orders of magnitude lower than that which is measured at the start. The execution time they impose is inferior to MI, much as was expected all along, but might be superior to that of MSD.
The MDL term is improperly defined at present since it ignores the model discrepancies. When MDL is approximated by the determinant of the covariance matrix of the model, problems arise and registration (or landmark identification) is exacerbated past the stage when convergence should hold. Instead of convergence, a logarithmically decreasing 'tail' is observed and it indicates the need for objective functions being revised.