ROC curve evaluation of edge detector performance

We present a method of evaluating edge detector performance based on empirical receiver-operating characteristic (ROC) curves. The approach is based on using real images, for which the ground truth has been manually specified. The ROC curve for a given edge detector summarizes its range of tradeoffs between true positive and false positive edge pixels, as determined by comparing the detected edge pixels to the specified ground truth. Aggregate ROC curves are constructed for groups of images. To give a fairer evaluation relative to typical application use, a form of "train and test" methodology has been developed.