Image segmentation based on prior probabilistic shape models

In the domain of shape modeling researchers have investigated various shape models in the abstract. In particular, Zhu and co-workers have proposed a maximum-entropy based model of object shape derived from data. In this paper we investigate the possibility of using such a shape model for image segmentation. To our knowledge, this is the first attempt at using such models to solve practical problems. One of the challenges of these models is their computational cost. We develop an efficient Monte Carlo Markov Chain (MCMC) method to train the model. A Bayesian framework is then used to perform segmentation based on the identified prior shape model and observed data. Based on limited evaluation we can demonstrate superior segmentation results relative to segmentations performed with no prior information and segmentations performed using a curve evolution approach with a generic boundary length penalizing prior.