The paper contributes to the viewpoint invariant recognition of planar patterns, especially labels and signs under affine deformations. By their nature, the information of such eye-catchers is not contained in the outline or frame — they often are affinely equivalent like parallelograms and ellipses — but in the intensity content within. Moment invariants are well suited for their recognition. They need a closed bounding contour, but this is comparatively easy to provide for the simple shapes considered. On the other hand, they characterize the intensity patterns without the need for error prone feature extraction. This paper uses moments as the basic features, but extends the literature in two respects: (1) deliberate mixes of different types of moments to keep the order of the moments (and hence also the sensitivity to noise) low and yet have a sufficiently large number to safeguard discriminant power; and (2) invariance with respect to photometric changes is incorporated in order to find the simplest moment invariants that can cope with changing lighting conditions which can hardly be avoided when changing viewpoint. The paper gives complete classifications of such affine / photometric moment invariants. Experiments are described that illustrate the use of some of them.
Van Gool, L., Moons, T., & Ungureanu, D. (1996). Computer Vision — ECCV ’96. (B. Buxton & R. Cipolla, Eds.), Computer Vision — ECCV ’96 (Vol. 1064, pp. 642–651). Berlin/Heidelberg: Springer-Verlag. Retrieved from http://www.springerlink.com/content/f312l5u5843p4058