I'm not involved with the app, but there's no need for a hidden marker in this case -- the page IS the marker. It is a planar surface with one of a few known black and white images on it. So: Detect features in each blank drawing, store them in a database included in the app, then use the same corner detector and descriptor at runtime and you have a very constrained feature matching problem.

If you're not familiar with feature detection methods, SIFT is a famous one, though there are much faster methods existing today: http://en.wikipedia.org/wiki/Scale-invariant_feature_transfo...