Evaluating Edge Detection through Boundary Detection

Edge detection has been widely used in computer vision and image processing. However, the performance evaluation of the edgedetection results is still a challenging problem. A major dilemma in edge-detection evaluation is the difficulty to balance the objectivity and generality: a general-purpose edge-detection evaluation independent of specific applications is usually not well defined, while an evaluation on a specific application has weak generality. Aiming at addressing this dilemma, this paper presents new evaluation methodology and a framework in which edge detection is evaluated through boundary detection, that is, the likelihood of retrieving the full object boundaries from this edge-detection output. Such a likelihood, we believe, reflects the performance of edge detection in many applications since boundary detection is the direct and natural goal of edge detection. In this framework, we use the newly developed ratio-contour algorithm to group the detected edges into closed boundaries. We also collect a large data set (1030) of real images with unambiguous ground-truth boundaries for evaluation. Five edge detectors (Sobel, LoG, Canny, Rothwell, and Edison) are evaluated in this paper and we find that the current edge-detection performance still has scope for improvement by choosing appropriate detectors and detector parameters.