A novel similarity measure under Riemannian metric for stereo matching

Stereo matching has been one of the most active areas in computer vision for decades. Many methods, ranging from similarity measures to local or global matching cost optimization algorithms, have been proposed. In this paper, we propose a novel similarity measure under Riemannian metric. A generalized structure tensor is applied to describe a point and the similarity is measured by the distance between the associated tensors. Since the structure tensor lies in a Riemannian manifold, the distance between structure tensors is the geodesic distance on Riemannian manifold. We show that our similarity measure provides an efficient way to fuse different features and it is independent of illumination change and window scaling. Experiments on standard dataset prove that our similarity measure outperforms many traditional measures such as SSD, SAD and normalized cross-correlation (NCC).