Quantitative evaluation of performance through bootstrapping: edge detection

A new quantitative performance evaluation technique for computer vision systems is proposed. In real situations the complexity of input data and/or computational procedure can make the traditional error propagation methods infeasible. Using bootstrapping, a numerical technique for deriving statistical characteristics from a single sample, the authors perturb the nuisance properties of the input image to obtain distributions for the output variables. The performance thus is evaluated for the given input and system and not under simplifying assumptions. The task of edge detection is used as example.