Analysis of Means to Improve Cooperative Disparity Estimation

(Zitnick and Kanade, 2000) proposes a cooperative approach for disparity estimation from stereo imagery based on support and inhibition in three-dimensional (3D) disparity space. By several means we obtain a significant improvement over the results reported for (Zitnick and Kanade, 2000) in (Scharstein and Szeliski, 2002). The results are in the range of the best approaches in (Scharstein and Szeliski, 2002), while the numerical complexity of our approach compares favorably to these approaches. Our main contribution lies in analyzing the different means for improvement including their performance gain. The most important are the use of symmetric support, the combination of absolute differences and (normalized) cross-correlation weighted by the strength of the horizontal gradient, the use of auto-correlation to estimate the significance of a matching score, the preference for small disparities to obtain more meaningful results for occluded regions, and the enforcement of the alignment of disparity and image gradient. Results for additional images with the same set of parameters show that the means are applicable to a wider range of imagery.