Detection and correction of disparity estimation errors via supervised learning

We propose a supervised learning method for detecting disparity estimation errors in a disparity map. A classifier is trained using features of low computational complexity. The proposed method can in principle be used to improve the performance of any disparity estimation algorithm. The results presented in this paper are therefore of general interest for those working on disparity estimation. In addition, our method solves the problem of needing a large variation of input stereo video with ground truth disparity. In our approach, we visually inspect a disparity map and manually annotate blocks that appear to be errors and blocks that appear to be correct. We then train a classifier to do this work automatically. Recursive predictions are used to correct errors. Our manual annotation approach has the advantage that `ground truth´ data is generated via low-cost annotation of arbitrary stereo video.