PDF version of this document
This idea is concerned with sub-division of the problem and localisation of computation which otherwise becomes very demanding. Since the model constructed is conventionally build from the entire set under consideration, there is a clear correlation between the complexity of the entire problem and that size of the set. This correlation is not linearly proportional to the size of the set either. To get decent results (either in landmark selection or image correspondence), a very long optimisation is required for increasingly larger sets of data. Although the principles are genuine and sophisticated at first sight, they suffer from this unappealing relational complexity.
The next section explains this principle in the context of images. It was decided to try to apply the same concepts to shapes after it was arguably successful when images were under consideration.