PDF version of this document

Contributions

Taking a more positive stance, it has been illustrated not only that group-wise registration based on appearance model is possible, but also that it surpasses registration methods that are based purely on a reference image, as judged by the corresponding appearance model.

The current algorithms are being interpreted rather than compiled and no multi-scale approach is yet in use. The extension of the algorithms to 2-D and 3-D would require a long time to run, but remains practical. Compiled implementations might be available within months as well as heuristic optimisations that will make run-time more competitive with that of pair-wise approaches. Furthermore, the results have been shown to be better in a global sense and are not dependent on just one individual image.

Contributions of the work can be subdivided into three aspects:

1. It provides a benchmark environment and results for many methods, including several new ones.
2. Unprecedented model-based group-wise registration is introduced.
3. Automatic construction of increasingly better appearance model becomes practicable. Correspondences are obtained using techniques borrowed from image registration.