Reinterpreting Physically-Motivated Modeling

Deformable models in the “physically-based” paradigm are almost always formulated in an adhoc fashion and are not related to physical reality. We reinterpret these techniques by putting them into a framework of robust statistics and using this framework to analyze the problems and ad-hoc solutions found in common physically-based formulations. These include incorrect prior shape models; bad relative weights of various energies; and the two-stage approach to minimization (adjusting global, then local shape parameters). We examine the statistical implications of common deformable object formulations. In our reformulation, the units are meaningful, training data plays a fundamental role, different kinds of information may be fused, and certainties can be reported for the segmentation results. We suggest robust statistics to combat interference from the necessarily large amount of unmodeled image information.