PDF version of this entire document
In order to better understand what the reference should be, a closer look at our dataset was seen as necessary; therein, several sets contain neutral and non-neutral image captures (of the same subject/s). In the case of Mian's group, information is available which was taken from 3 subjects only, containing roughly 1,000 images from each and involving a talking sequence a la Hack and Taylor[13]. This is valuable for a training phase, wherein PCA is used for intra-personal variability learning. Whether it generalises to other subjects or not is another matter altogether.
