This paper addresses performance evaluation in the presence of imprecise ground-truth. Indeed, the most common assumption when performing benchmarking measures is that the reference data is awless. In previous work, we have shown that this assumption cannot be taken for granted, and that, in the case of perceptual interpretation problems it is most certainly always wrong but for the most trivial cases. We are presenting a statistical test that will allow measuring the con-dence one can have in the results of a benchmarking test ranking multiple algorithms. More specically, we can express the probability of the ranking not being respected in the presence of a given level of errors in the ground truth data.