A Performance Measure for Boundary Detection Algorithms

摘要

In this paper we discuss the issues involved when trying to compare the performance of algorithms which seek to find boundaries and boundary models between regions of differing mean gray-level value. The usefulness of such measures is not confined to comparing different approaches, but provides an important step to building self-optimizing vision systems that automatically adjust algorithm parameters at each level of the system to improve performance. We discuss the issues that such a performance measure should address, and then present a theoretical framework for the performance measure we proposed. We demonstrate the power of the measure by characterizing the performance of a three-stage medium level vision system for detecting straight line boundaries. Both synthetic imagery, for which ground truth is known, and real imagery are used to test the ability of the criterion to characterize the output of the system.