Robustness to noise of stereo matching

Abstract

We have measured the performance of several area-based stereo matching algorithms with noise added to synthetic images. Dense disparity maps were computed and compared with the ground truth using three metrics: the fraction of correctly computed disparities, the mean and the standard deviation of the disparity error distribution. For a noise-free image, S. Birchfield and C. Tomasi's pixel-to-pixel dynamic algorithm performed slightly better than a simple sum-of-absolute-differences algorithm (67% correct matches vs 65%) $considered to be within experimental error. A census algorithm performed worst at only 54%. The dynamic algorithm performed well until the S/N ratio reached 36 dB after which its performance started to drop. However, with correctly chosen parameters, it was superior to correlation and census algorithms until the images became very noisy (/spl sim/15 dB). The dynamic algorithm also ran faster than the fastest correlation algorithms using an optimum window radius of 4, and more than 10 times faster than the census algorithm.