In this paper we present a semi-automatic method for labeling image sequences based on probabilistic classification. The method utilizes an initial color segmentation and models the distribution of the features around each resulting region as a mixture of multivari-ate Gaussians. The label field is estimated using a deterministic local search algorithm so that the joint probability of the labels and the observations is maximized. A statistical representation of each "color-region" is utilized and the steps of the algorithm are expressed in terms of the "color-region" statistics. Experimental results are presented.
展开▼
