PDF version of this document
Attempts were made at the identification of the point of convergence for the registration algorithm. A clear flaw with the cost that had been defined was discovered. It turned out that the way in which the algorithm presently evaluates models neglects to account for small artifacts. These artifacts must be encapsulated in this model. More crucially, these small artifacts which are left-out residuals need to form part of the model cost (description length term). In their absence, the objective function was able to drift away, thereby hiding vital structures in images. Not only was the result of registration poorer due to an improper model cost, but also it was impossible to contend that one unique solution can ever be reached.
It is now realised that the objective function must account for the model errors in some way or another. This is why the residuals for each image, as reconstructed from the model, need to be calculated. All inverse warps needs to be calculated first to do so - a step that is not trivial (on-going and future work is scheduled to be done on this).
Before this incorporation of residuals goes on, it was suggested that the shapes problem is looked at again. It is believed that description length should have a term accounting for model discrepancy so that the optimisation can be made stable. Technicality concerning MDL is expected to be discussed with Carole Twining in the near future. Eventually, a term must be precisely defined to account for the residuals and, having solved the problem for the simpler and well-founded case of shapes, the less trivial case of image registration can be resumed. Several discussion have raised disagreements regarding the way in which residuals are defined in the context of appearance models, images and transformations.
To summarise, of current interest is the way in which model residuals can nullify erosion of data. They can be used to compose a proper description length for models and images. By resolving such a problem, better registration performance will be yielded.