Distinguished Regions for Wide-baseline Stereo

The problem of establishing correspondences between a pair of images taken from different viewpoints, i.e. the “wide-baseline stereo” problem, is studied in the paper. The choice of image elements that are put into correspondence in the wide-baseline matching problem is discussed. The concept of a distinguished region is introduced and formally defined and it is argued distinguished regions are very good candidates for matching. Two new types of distinguished regions, the Separated Elementary Cycles of the Edge Graph (SECs) and the Maximally Stable Extremal Regions (MSERs), are introduced. For both types, an efficient (near linear complexity) and practically fast detection algorithm is presented. Experimentally the stability of the proposed DRs is shown on disparate views of real-world scenes with significant change of scale, camera rotation and 3D translation of the viewpoint. A new robust similarity measure for establishing tentative correspondences is proposed. The robustness ensures that invariants from multiple measurement regions, some that are significantly larger (and hence discriminative) than the distinguished region, may be used to establish tentative correspondences. In experiments on indoor and outdoor image pairs, good estimates of epipolar geometry are obtained on challenging wide-baseline problems with the robustified matching algorithm operating on the output produced by the proposed detectors of distinguished regions. Locally fully affine distortions and significant occlusion were present in the tests.