Hierarchical contour point matching in image pairs

The authors describe a technique which combines region- and point-matching methods in a way that captures their strengths. Regions are extracted from two images and matched on the basis of features which are invariant to two-dimensional rotation, translation, and scale transformations. Matching at the gross region level is relatively quick and accurate compared to pixel-level matching, since there are many fewer regions in an image than there are pixels, and thus fewer candidate matches. The region displacement provides an estimate of the displacement of points within the region; this estimate is used to guide a correlation function matching all of the region´s contour points. Displacements of many points are thus obtained in a more efficient manner than in an unguided correlation search