Now I've made a version that samples an image under a number of tilts and compares the results.
It turns out this is a very handy approach.
http://workinprogress.ludatic.nl/ludatic/facedetection/
I used the deface project as a basis, this is without the tilting offcourse, which as you can see, adds a lot of accuracy.
http://code.google.com/p/deface/
