PDF version of this document

Perturbation Framework

In order to test our method, it is required that we take ground-truth registrations and transform them as to degrade them in a well-understood fashion. By doing so, a set comprising different registrations of varying quality can be obtained. By applying our method to each such set in turn, we should be able to demonstrate that our method discerns good registrations from worse ones. Ideally, our method would be able to provide a benchmark whose scale is monotonic and span a wide range of possible mis-registrations.

To perturb the registration we use clamped-plate splines which are composed of 25 knot-points, all of which are randomly distributed across the images. By increasing the magnitude of the warp, we can directly increase the level of mis-registration.