Don't forget that NIR cameras (often used for iris recognition) can still see through those glasses. But i agree, stuff similar to the challenges in the AR face dataset [1] will usually stump most employed systems.

[1] http://www2.ece.ohio-state.edu/~aleix/ARdatabase.html

True although the commercial deployment of NIR face recognition isn't that extensive (I'm not saying it doesn't exist however). I know for our company's software that academics have been trying to use it with NIR images even though our software wasn't trained at all for it.

I do think that NIR will eventually become more important (especially for car applications) and really the main stumbling block is massive sets of training data for it.

Yea the NIR wavelength makes the skin lose a lot of discrimination ability. Don't know how familiar you are with the FOCS, SCFace, or PolyU datasets, but they're a good start for training (of course very few papers use them, they'd rather gain an extra 0.1% on LFW).

What about mirror shades?