Multiple appearance models

This paper investigates a concept for modelling complex data based on sub-models. The task of building and choosing optimal models is addressed in a generic information theoretic fashion. We propose an algorithm based on minimum description length to find an optimal sub-division of the data into sub-parts, each adequate for linear modelling. This results in an overall more compact model configuration called a model clique and in better generalization behavior. The algorithm is applied to active appearance models, active shape models and eigenimages and is evaluated on 4 different data sets. Experiments indicate that model cliques exhibit better generalization behavior than single models and mimic intuitive sub-division of data.