A quantitative error measure for the evaluation of roof edge detectors
Baptiste Magnier, Alexandre Le, Albert Zogo · 2016
Crest line extraction remains a hard task in image processing. Indeed, these roof edges represent narrow edges on the image surface and whatever undesirable pixel close to or on the crest line may disturb the detection. This communication presents a new crest line detection overall evaluation. Comparing the ground truth contour image and the candidate crest line image, the proposed algorithm is based upon a new criterion that take into account the list of ground truth, the recall and their associated spacial nearness. Doubtlessly, an efficient evaluation penalizes a misplaced edge point proportionally to the distance to the true contour. This quantitative performance evaluation proves its efficiency on several crest line images of different types, bringing a favorable indicator for tow closest edge images or a poor in the presence of a degraded/distorted candidate edge image.