One of the main challenges of using image distance functions is
that function results do not always coincide with the perceived
similarity between two images. Therefore,it is not straightforward
to for mulate suitable compactness criteria based on this. We use
whole-image dissimilarity functions between two arbitrary ima-ges of an ordered image sequence S. In order to cope with lighting
changes and small perspective deformations, this function is
implemented using a region descriptor [34] and to further mini-mize crosstalk due to small camera movements, the functionis
explicitly made invariant to small affine image transforms.