Framework to building a new ground image database: For segmentation and evaluation

International audienceEdge detection technique needs to be assessed before use it in a computer vision task. As dissimilarity evaluations depend strongly of a ground truth edge map, an inaccurate datum in terms of localization could advantage inaccurate precise edge detectors or/and favor inappropriate a dissimilarity evaluation measure. we demonstrate how to label these ground truth data in a semi-automatic way. The most common method for ground-truth definition in natural images remains manual labeling by humans Berkeley Segmentation Dataset proposed by Martin and Fowlkes in 2001. These data sets are not optimal in the context of the definition of low-level segmentation. Errors may be created by human labels (oversights or supplements); indeed, an inaccurate ground truth contour map in terms of localization penalizes precise edge detectors and/or advantages the rough algorithms. Then, an incomplete ground truth penalizes an algorithm detecting true boundaries and efficient edge detection algorithms obtain between 30% and 40% of errors. In fact, this new label processes in return to hand made ground truth. Indeed, in the first time, the contours are detected involving the convolution of the image with [-1 0 1 ] vertical and horizontal masks followed by a computation of a gradient magnitude and a suppression of local non-maxima in the gradient direction. Concerning color images, [-1 0 1 ] vertical and horizontal masks are applied to each channel of the image followed by a structure tensor. In a second time, undesirable edges are deleted while missing points are added both by hand. Using the [-1 0 1 ] mask enables to capture the majority of edge points and corners without deforming small objects, contrary to edge detectors involving Gaussian filters. By comparison with a real image where contours points are not precisely la- belled, experiments illustrate that the new ground truth database allows to evaluate the performance of edge detectors by filtering