PDF version of this document

A Framework for Evaluation
of Appearance Models - DRAFT

概要:

Models of appearance are powerful tools for capturing data variability and they are capable of synthesising data. Such models have been shown to posses rich 'knowledge' of what data within a set comprises and the way such data can be decomposed and hence simplified. A framework was developed which is able to evaluate appearance model. It is able to tell apart models of varying quality, thereby promoting better algorithms for construction of appearance models. The method also allows the validation of models. By measuring 'distances' between images, it quantifies the proximity between a model and its data. To measure distance between images, a shuffle transform is used, which is robust. Two separate measures reflect on the quality of an entire model, given a large matrix of distances. Specificity measures how well data generated by the model fits data from which the model was constructed, whereas generalisation compares data from which the model was constructed and data generated by the model. The methods were shown to work well when applied to face data and MR brain data. In both cases, progressively perturbed models were correctly analysed by our measures. The framework was used to compare models of the brains, which were built automatically. Models which are known to be superior were merited by the framework.