We have been developing techniques for automatically detecting key facial feature points on new, unseen faces.
Our algorithm was an iterative, EM-based algorithm that deforming a 2D/3D parametric model to fit a set of noisy 2D points.
The initial positions of the points are given by a simple gradient-based feature point detector.
The algorithm iteratively regularizes this initial observation by shrinking it in principle subspace, and identifies outliers by measuring matching errors.
Comparing with previous alignment methods, we showed the algorithm is more robust and the algorithm runs fully automatically while working together with a multi-view face detector.