提出了一种基于共同视域的自监督立体匹配算法,该算法根据视差的左右一致性来确定双目图像的共同可视区域,从而抑制被遮挡区域产生的噪声,为网络模型的学习提供了更加准确的反馈信号。研究结果表明:在没有任何标签数据的前提下,所提算法的预测误差降低了11%~42%,且与有监督立体匹配算法的性能相当。

A self-supervised stereo matching algorithm is proposed based on common view. In this algorithm, the common visible region of the binocular images is determined according to the left-right consistency of disparity and thus the noise generated in the occluded region is suppressed, which provides more accurate feedback signals for the network model learning. The research results show that the prediction error of the proposed algorithm can be reduced by 11%-42% without any label data, and the performance of the proposed algorithm is comparable to that of the supervised stereo matching algorithm.