On the confidence of stereo matching in a deep-learning era: a quantitative evaluation