A performance analysis of dense stereo correspondence algorithms and error reduction techniques

Dense stereo correspondence has been intensely studied and there exists a wide variety of proposed solutions in the literature. Different datasets have been constructed to test stereo algorithms, however, their ground truth formation and scene types vary. In this paper, state-of-the-art algorithms are compared using a number of datasets captured under varied conditions, with accuracy and density metrics forming the basis of a performance evaluation. Pre- and post-processing disparity map error reduction techniques are quantified.