DANIEL FREEDMAN

Many problems in computer vision can be posed in terms of energy minimization, where the relevant energy function models the interactions of many pixels. Finding the global or near-global minimum of such functions tends to be difficult, precisely due to these interactions of large ð> 3Þ numbers of pixels. In this paper, we derive a set of sufficient conditions under which energies which are functions of discrete binary variables may be minimized using graph cut techniques. We apply these conditions to the problem of incorporating shape priors in segmentation. Experimental results demonstrate the validity of this approach.