PDF version of this document
There is no question about the possibility of improving previous results. Nonetheless, what needs to be pursued are ways of using obvious evidential methods to find models that are essentially optimal in some sense and are independent of the data under consideration. Algorithm that are ad-hoc and work better under certain conditions are of little interest as they will not generalise or offer any substantial progress in the long run.
Currently, the extensively used warps are not as flexible as one could hope for. While the theoretic principles work adequately, they sometimes prove impractical for use on contemporary machines. Better ways of applying such warps are being investigated so that fewer warps of lower complexity end up bringing two data samples to convergence rather quickly.
Texture patches have not seen significant enhancements in some previous work and further exploration seems worthwhile. Some of the principles from the work on ASM's could possibly be incorporated to produce more integral and consistent appearance models. The new models should exhibit good correlating between shape and intensity.
Some of the suggestions above are rather hard to convert into concrete implementations. Any such developments could lead to a real breakthrough. Experience and understanding of the research domain could suggest that a process of gradual trial-and-error would be fruitful. The next section presents some of the expected difficulties and Section 6 concludes and predicts the work plan that this document will entail.