Feature matching with Procrustes alignment and graph editing

We describe a novel approach for point-set registration which can deal with affine distortion and structural corruption. Procrustes analysis is used to pre-align the two point-sets. This eliminates the translation, scale and rotation deviations between both point-sets. The novelty of the approach is introducing structural constraints into Procrustes alignment to improve the registration. The constraints are established by means of Delaunay triangulation. Nodes are deleted as outliers if their matched pattern of correspondences are inconsistent. By iterating between procrustes alignment and graph editing we improve the performance of the point alignment process. The experimental results reveal that the new alignment method is more robust and accurate when structural corruption is a limiting factor.