Conficonv: A Confidence Measure For Disparity Estimation Based On Deep Learning

In this paper, a confidence measure for disparity estimation is proposed to encode the degree of uncertainty of each point in disparity map. Based on Convolutional Neural Network (CNN), a network is set up which is named as Confidence Convolutional neural network (ConfiConv). Compared with four different Confidence Measures (CMs) in Middlebury 2014 dataset using two stereo vision algorithms, it is shown that the average Area Under Curve (AUC) values of ConfiConv are better than other measures. And ConfiConv is also tested in another dataset which contains multiple movie clips.