PDF version of this document
The evaluation of appearance models becomes practical through the use of large sets of data. Fitting the many possible cross-pairings of data and model instance leads to measures which are robust independently of data properties. Results have been shown for brain data as well as a challenging set of face data and graphs have always appeared encouraging. The evaluation method relies upon a distance measure between a pair of images and measures such as shuffle distance appear most suitable to handle the task.
The method able to compare similar models and distinguish between them successfully. This was shown to be the case when models were corrupted intentionally, but also in cases where models and their quality were poorly understood. In such circumstances, the method was able to provide answers and be used for benchmarking. It appears to suggest that registration in a group-wise manner results in better models of appearance. This opens the door to a framework which validates registration. The observation which motivates it is that correct registration identifies the correspondence perfectly, and therefore builds optimal models.