Feature-based estimation of radial basis mappings for non-rigid registration

We study the challenging problem of registering images of a non-rigid surface by estimating a Radial Basis Mapping from feature matches. We cast the problem as a Maximum Likelihood Estimation coupled with nested model selection. We propose an algorithm based on dynamically inserting centres and refining the transformation parameters under the control of a selection model criterion. We validate the algorithm using extensive simulations and by building on recent feature extraction and matching techniques, we report convincing results on real data.