We are playing with the capabilities of the algorithm published by Gabriele Fanelli, for an upcoming interactive installation. The results are very promising as you can see in the following video...
No training or special setup needed, heads are detected instantly. This funny hack is the first rough test, applied to augmented reality.
We are working to improve our code, specially the speed and quality of the triangulation. The detection algorithm uses random regression forest to match the pose of the user and estimate the position and orientation of the head. The result is surprisingly robust, accurate and fast, even with partial occlusion of the head. However, with the provided training database, it is usual that you get false positives.
FYI, Kyle McDonald has the source code in his repo: https://github.com/kylemcdonald/ofxHeadPoseEstimator