A Closed-Form Solution for Image Sequence Segmentation with Dynamical Shape Priors

In this paper, we address the problem of image sequence segmentation with dynamical shape priors. While existing formulations are typically based on hard decisions, we propose a, formalism which allows to reconsider all segmentations of past images. Firstly, we prove that the marginalization over all (exponentially many) reinterpretations of past measurements can be carried out in closed form. Secondly, we prove that computing the optimal segmentation at time t given all images up to t and a dynamical shape prior amounts to the optimization of a. convex energy and can therefore optimized globally. Experimental results confirm that for large amounts of noise, the proposed reconsideration of past measurements improves the performance of the tracking method.