One curious oddity is the indifference of the algorithm to skin tone or albedo. Since the ``mean'' face is composed mostly of Caucasians, dark-skinned individuals lose their distinctive skin colour. Whether this necessarily has a negative impact on recognition algorithms is uncertain. It is evident, though, that the faces are still recognizable to a human observer after illumination normalization.
Overall, we note the regularity with which the faces appear in the normalized image gallery. The variance in appearance has been constrained to be a function of individual identity and expression alone, since lighting and pose have been filtered out of the image. This allows recognition and classification of identity to be performed on the basis of variance which we assume is dominated by the identity (not facial expression) of the individual in the image. This statement is true most of the time since people display a neutral expression in their daily activities (watching TV, walking, working, etc.)
On an SGI Indy workstation, the full normalization computation requires under 20 milliseconds for each 7000 pixel mug-shot shown above. The time required for the computation is almost proportional to the number of pixels in the output mug-shot being generated. For an 800 pixel mug-shot, the time required drops to below 2 milliseconds on the SGI Indy. Thus, the above normalization process is extremely efficient.