Stella X. Yu : Papers / Google Scholar

Boundaries as Contours of Optimal Appearance and Area of Support
Christina Pavlopoulou and Stella X. Yu
Energy Minimization Methods in Computer Vision and Pattern Recognition, Bonn, Germany, 24-27 Aug 2009
Paper | Slides

Bayesian boundary models often assume that the evidence for each contour is derived from the entire image. Consequently, the normalization term in the Bayes rule is the same for every contour and becomes irrelevant when seeking the optimal. However, in practice these models only use the vicinity of a contour, making the normalization term contour-specific. We propose a formulation that acknowledges the normalization term and includes it in the optimization. We show that it can be interpreted as a confidence measure promoting contours which are far better than other nearby candidate contours. We validate our approach in an interactive boundary delineation setting and demonstrate that complex boundaries can be extracted with significantly smaller amount of user input than when traditional Bayesian models are employed.

Bayesian formulation, bias problem, normalization