A Bayesian Multiple-Hypothesis Approach to Edge Grouping and Contour Segmentation

Abstract

A contour segmentation algorithm is presented that takes an edge map and extracts continuous curves of arbitrary smoothness, correctly handling curve intersections and capable of extrapolating over significant measurement gaps. The algorithm incorporates noise models of the edge-detection process and limited scene statistics. It is based on an explicit contour model and employs a statistical distance measure to quantify the likelihood of each segmentation hypothesis. A Bayesian multiple-hypothesis tree organizes possible segementations, making it possible to postpone grouping decisions until a sufficient amount of information is available. We have demonstrated its performance on real and synthetic images.