Disparity search range estimation based on dense stereo matching

This paper presents a scheme for estimating disparity search range based on hierarchical stereo matching. It is important to specify a proper range of search space, since it prevents the solution from being trapped in local minima, and saves a lot of time for estimating disparity maps. Conventionally, it has been estimated by finding a sparse set of the correspondences via feature matching techniques. Instead, we address this problem by considering how the dense correspondences, the ultimate goal of estimating search range, can be estimated without search range. First, we estimate the dense correspondences by adapting a simple local stereo matching technique with an arbitrary search range. The hierarchical scheme is leveraged for reducing computational costs and memory usages. Then, reliable checking techniques are performed for eliminating unreliable correspondences. Finally, the range of search space is estimated by observing the distribution of the reliable correspondences. For the quantitative evaluation, a new error metric, biased root mean squared error (B-RMSE), is proposed, which differentiates the estimated search range whether it is narrower or wider than true disparity range. The experimental results show that the proposed method gives more accurate search range compared to the conventional method.