Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract


Segmentation with MCMC Curve Sampling


Ayres Fan
SSG, LIDS, MIT


Segmentation is the process of dividing an image into coherent regions. We introduce a segmentation method based on Markov Chain Monte Carlo (MCMC) sampling techniques. The advantages of sampling methods over traditional optimization-based methods include better robustness to local minima, natural handling of multi-modal distributions, and access to higher-order statistics. With MCMC sampling methods, we wish to sample from a distribution p. To do so, we iteratively sample from a proposal distribution q and accept or reject those samples according to a decision rule. If the decision rule satisfies detailed balance, then the iterations will converge to samples from p. We detail some sampling methods based on both continuous and discrete curve evolution formulations. We show how our methods satisfy detailed balance and demonstrate some initial results on the prostate.


Problems with this site should be emailed to jonesb@mit.edu