PDF version of this entire document

Normalisation

Normalisation (with scale) of the face as a geometric structure would be a mistake, so it is worth specifying what we refer to as ``normalisation''. While rotation may be fine, scale is a trickier transformation because one must think about the fact that the distance from nose to chin could in fact be used to identify a person. So, the normalisation applied is in no way modifying anything in the image itself. In fact, nothing is being rescaled or even rotated. The image is being translated only, so as to bring into alignment the many noses, assuming no-one's nose is behind the chest (easy to check for these special cases), the chin, or the forehead (requiring the person to look up or down, although that too can be checked to avoid misclassification). The use of the term normalisation refers not to any real transformation but rather to the acquisition of additional data, which allows the algorithm to determine:

• How much of the face to crop for PCA/GMDS
• How low a surface beneath the eyes and the nose should be treated as rigid for the sake of reliable alignment
• Suffice to say, this may also come handy when applying ICPv2.

One could, in theory, use normalisation for decomposition (PCA) and then plug back the normalisation scale to compensate for an aforementioned and previously-applied scaling. That is not the method being adhered to, however, because scale can be treated just fine as long as the sample point are selected correctly (with reasonable correspondences marked up); PCA can overcome scale anomalies.

As the next sub-section explains, concerns remain.