PDF version of this entire document

PCA

Work similar to ours predates or concurs with what we achieved by the middle of 2011.

In 2006, Russ at al. [29] produced a paper on using PCA for face recognition in 3-D, using FRGCv1 and FRGCv2. They tackled the issue of face alignment which is required for adequate sampling of signal for PCA. To quote the paper, they ``achieve correspondence of facial points by registering a 3D face to a scaled generic 3D reference face and subsequently perform a surface normal search algorithm.''

A later paper, one from Mena-chalco at al. in Brazil [22], demonstrates early work that is carried out on a single subject and many acquisitions (much like the GIP dataset for expressions). The texture is incorporated too, building a model using PCA with a very small training set. This work is very different from what we do and their data sets are their own.

Using a variant of classic PCA, (2D)$^{2}$ PCA.A or 2DPCA, Gervei at al. [11] showed recognition rates of 83.3%. This deals with facial expressions too and is similar to what we already have implemented, including the pre-processing and the sampling phases. Even their charts are similar to ours (block diagrams), not just the recognition rates. They use the Gavab 3D face database which contains 540 3-D images from 60 individuals. For the sake of comparison, our June experiments use similar numbers to train the model, whereas UWA uses data of the orders of magnitude of thousands (mostly collected from 3 individuals at their lab).