Stochastic Systems Group  

Junmo Kim  SSG, MIT
Image segmentation, the process of decomposing an image into meaningful regions, becomes difficult when the image is of low quality or we have missing data due to occlusion. In that case, some prior knowledge about the object to be segmented will be especially useful. The problem we are interested in is to extract such prior information from available example shapes. In particular, we want the prior information in terms of shape prior distribution such that for an arbitrary target shape we can evaluate the probability that the target shape belongs to the same class as example shapes.
In this talk, we will present a nonparametric shape prior model and its application to image segmentation. In this approach, we assume that the example shapes are drawn from an unknown shape distribution, and we estimate the underlying shape distribution using a Parzen density estimator. Such density estimates are expressed in terms of distances between target shape and example shapes. How to define and compute a distance (metric) between shapes is a big topic in itself, and we consider a few tractable metrics. Finally, we will discuss how to evolve an active contour given a shape prior and show some shapebased segmentation results.
Problems with this site should be emailed to jonesb@mit.edu