Learning image alignment without local minima for face detection and tracking