Geometrical invariance can be acheived by assuming that the regions to be matched are locally planar, and by describing them in a manner which is invariant under homographies. Many authors use feature descriptors which are invariant under special cases of this group e.g. similarities or affinities. For example, Schmid and Mohr [10] use rotationally symmetric Gaussian derivatives to characterise image regions. Lowe’s SIFT features [7] use a characteristic scale and orientation at interest points to form similarity invariant descriptors. Baumberg [2] uses the second moment matrix to form affine invariant features.