Perceiving Shapes through Region and Boundary Interaction

We present a computational algorithm for shape perception by studying the interaction between region and boundary cues. We formulate this problem in a graph partitioning framework, where region cues defined on a pixel graph and boundary cues defined on its dual edgel graph are coupled through edge-node incidence relationships. The consistency of simultaneous partitioning on such graphs can thus be guaranteed. We generalize normalized cuts criteria and algorithms to this model for globally optimal solutions. We demonstrate that by incorporating boundary smoothness, objects with heterogeneous region properties can stand out as one group and objects with weak contours can be segmented more readily without the suppression from objects of high contrast. This model can also encode higher-order shape information and preliminary results on shape selection are given.