PDF version of this document
When approaching the problem of image interpretation, one has several paradigms for solving it. One such paradigm is the modelling of objects and the use of models to learn something about an object. Along this simple approach, shape models [3] were developed and they assisted in analysing contours of objects in an image. The natural extension to shape models was one which encapsulated intensity information, as well as contours. The work of Edwards et al. [4] brought about models which were capable of synthesising photo-realistic images.
Following the success of this approach, several groups have built appearance models using different methods and obtained results of differing quality. Stegmann [5], for example, reproduced algorithms for appearance model construction and extended them to account for 4-D models that include the dimension of time. Rueckert et al. [14] have taken this approach and embedded it in registration algorithms so that statistical deformation models are built subsequent to registration.
Davies et al. [2] have adopted a method for the evaluation of shape models, i.e. the derivation of values for a given shape descriptor. This was done by a minimum description length [16] approach, which relates to the simplicity of a model. Ever since, the method allowed evaluation and comparison between shape models - possibly built by different algorithms - to be compared. Furthermore, it enabled the formation of an information-theoretic objective function. Such as objective function, when treated an an optimisation problem, allowed optimal shape models to be constructed.
While methods of evaluation are available for shape models and, therefore, quantitative comparison is possible, none is available for appearance models. Since appearance models are far more complex and far heavier than shapes (as they include texture information), a method has been thought for evaluating them and arguing about their validity.
This paper outlines a successful method for the evaluation of appearance models. The method is shown to be well-behaved and its applicability to faces and brains is illustrated. Furthermore, it is used to compare different methods of model construction, all of which do so without need for manual mark-up of the data. Finally, it is shown that the method is able to correctly distinguish between models that appear identical to the naked eye.