This work is used for pose estimation(yaw, pitch and roll) by Face landmarks(left eye, right eye, nose, left mouth, right mouth and chin). Roll:+90°：-90°/Pitch:+90°：-90°/Yaw:+90°：-90°, like the picture below:

The order of numbers is ROLL, PITCH, YAW：

• I fine-tune the MTCNN into the output of 6 landmark feature points, reference and make some adjustments in this article 'Head Pose Estimation using OpenCV and Dlib'.
• Because the MTCNN's eyes are the middle of the position rather than the corner of the eye, so we modify the world coordinate(model point) from original to (-150.0, -150.0, -125.0)# Left Mouth corner/(150.0, -150.0, -125.0)# Right mouth corner
• Modify the camera matrix's focal_length from original to img_size[1]/2 / np.tan(60/2 * np.pi / 180).

1. imgpts, jac = cv2.projectPoints(axis, rotation_vector, translation_vector, camera_matrix, dist_coeffs)
2. modelpts, jac2 = cv2.projectPoints(model_points, rotation_vector, translation_vector, camera_matrix, dist_coeffs)
3. rvec_matrix = cv2.Rodrigues(rotation_vector)[0]
4. proj_matrix = np.hstack((rvec_matrix, translation_vector))
5. eulerAngles = cv2.decomposeProjectionMatrix(proj_matrix)[6]
6. pitch, yaw, roll = [math.radians(_) for _ in eulerAngles]
7. pitch = math.degrees(math.asin(math.sin(pitch)))
8. roll = -math.degrees(math.asin(math.sin(roll)))
9. yaw = math.degrees(math.asin(math.sin(yaw)))

1. Head Pose Estimation using OpenCV and Dlib
2. MTCNN-tensorflow