PDF version of this entire document
An assumption which is demonstrated to be of practical use is that by warping images into a neutral (special expression-free) frame of reference all faces are to be reliably told apart. However, as suggested by research in other areas, detection is then biased towards the selection of this one reference, which innately favours a pairwise approach and not a groupwise one. Moreover, there is presumed availability of neutral datasets, which is hard to assure, and there tends to be a difference also between distinct neutral scans of the same subject, depending on the imaged pose for example. These issues, while still separate from the main investigation domain (and thus just a secondary matter), are worth taking into account. There is a vast body of literature, e.g. regarding brain MRI analysis, about reference selection and its ramifications.
