Coupled Region-Edge Shape Priors for Simultaneous Localization and Figure-ground Segmentation

We propose a new algorithm for simultaneous localization and figure-ground segmentation where coupled region-edge shape priors are involved with two different but complementary roles. We resort to a segmentation-based hypothesis-and-test paradigm to solve the problem, where the region prior is used to form a segmentation and the edge prior is used to evaluate the validity of the formed segmentation. Our fundamental assumption is that the opti-mal shape-constrained segmentation that maximizes the agreement with the edge prior occurs at the correctly hypothesized location. Essentially, the pro-posed algorithm addresses a mid-level vision issue that aims at producing a map image for part detection can be further used for high-level vision tasks. Our experiments demonstrated that this algorithm offers promising results in terms of both localization and segmentation. Key words: figure-ground segmentation, shape priors, segmentation, localization, watersheds, online learning, kernel-based color modelin