Landmarks & ratios
Once the model returns facial landmarks, we use only these 8 points. Distances use normalized coordinates, then divide by the same face width.
H = d(10, 152)
W = d(234, 454)
F = d(54, 284)
J = d(172, 397)
d(a, b) = 鈭?(x鈧?鈭?x岬?虏 + (y鈧?鈭?y岬?虏)