Edge Detector Evaluation Using Empirical ROC Curves

摘要

We demonstrate a method for evaluating edge detector performance based on receiver operating characteristic (ROC) curves. Edge detector output is matched against ground truth to count true positive and false positive edge pixels. A detector's parameter settings are trained to give a best ROC curve on one image and then tested on separate images. We compute aggregate ROC curves based on 1 set of 50 object images and another set of 10 aerial images. We analyze the performance of 11 different edge detectors reported in the literature.