PDF version of this entire document

Apparent Limitations

Any data we may have of few individuals and a breadth of facial expressions from them may be useful for a comparable demonstration of results. Had we had no access to their (apparently proprietary) data, experimenting with it is a walk through smoke and mirrors. Open Access and Open Data are particularly important for this reason, at the very least as complementary, auxiliary material in one's paper. In many cases, code too should be made available for audit (at all level), in order for one to defend the results.

Considerable time/attention should be dedicated to ensuring that the training phase is done with data which is known to have yielded good results; we do not quite have that from Mian and based on a quick survey of the stock of FRGC images, there are rarely cases where one individual was imaged more than a handful of times. Without the ability to prove that good eigenvectors can be derived from the set, leaping towards ROC curves and systematic experiments would be very premature and clearly time-consuming.

The intention here is not to defend GMDS by scrutinising a counterpart's work; rather, it's about understanding what it is exactly that they show and how it was achieved. For example, were comparable databases tested on? And if so, how were the models trained? Is there a preparatory phase the outsider if not being informed of? Models can surely be refined by studying variations 'off-line', without delving into large sets with significant variation.

Figure: A before/after overview
Image face-before-and-after

Roy Schestowitz 2012-01-08