I agree that faces are easy and contains lots of internal features, so that wavelet features can represent the characteristics of the face well. Still, they have pulled the performance of this "easy" problem up to a level that it actually meets the normal people's standards, not to just vision people's who are pretty generous.

Regarding Schneiderman's paper, I like the general attitude of the author. There isn't much mathematically solid foundation or any fancy model from the machine learning community, but the author has a good idea of what is going on in each step and what the advantages and limitations are. He doesn't over-generalize or make overstatement of his methods.

The limitations are, of course, that they are looking at a tiny subwindow each step and doesn't take any global information into account.