Stereo confidence metrics using the costs of surrounding pixels

In this paper, we propose two new stereo confidence metrics which are used for estimating the reliability of a stereo matching result more accurately. Unlike the conventional metrics which exploit the cost information of the pixel being processed only, the proposed metrics utilize the cost information of the surrounding pixels as well as the pixel being processed. By checking whether the costs of the surrounding pixels at the target disparity are the local minimum or not, or calculating the ratio between costs of surrounding pixels, the proposed metrics can estimate the reliability of a stereo matching result more correctly. Experimental results show that the two proposed metrics improve the area under the curve of the error rate as a function of disparity map density by up to 0.012 and 0.025, compared to the existing methods, respectively.