PDF version of this entire document

Summary and Discussion

This chapter demonstrated a general approach and framework for building appearance models directly from NRR. A registration algorithm can be used to identify correspondences in a group of images. In turn, these correspondences are used to build models of shape and appearance. While there are challenges to address, such as the speed of the optimisation and accuracy of the results, the method proves to be satisfactory enough to produce a model which embodies modes of variation that were introduced artificially. At the core of this algorithm, there exists an approach to estimating model complexity - the determinant of an appearance model. This method is fast, but it does not have a reasonable explanation that justifies it; hence the interest in MDL.

Subsequent experiments were carried out which followed the path towards automated model construction. They built upon this work on 1- and 2-D and were used by colleagues who extended the ideas further and produced publications, e.g. []. The author of this thesis meanwhile proceeded to the problem of assessing NRR, which is the topic that the next few chapters cover in detail.