This Space use the Apache 2.0 Licensed Mediapipe FaceLandmarker
One of json format is from MIT licensed face_recognition
I should clarify because it is confusing: I'm not using dlib's non-MIT licensed 68-point model at all.
Contour points are not point-to-point mapping,because when face rotated some point move inside.
Most of points are always divided equal for balancing when face rotated