Hard coded values in video_test.py

[shounakshastri]

Hello,

I was going through the code and I couldn't make it out what these values are in the video_test.py file.

def face_align_landmarks_sk(img, landmarks, image_size=(112, 112), method="similar"):
    tform = transform.AffineTransform() if method == "affine" else transform.SimilarityTransform()
    src = np.array([[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.729904, 92.2041]], dtype=np.float32)

and

nimgs = (np.array(nimgs) - 127.5) * 0.0078125

How are these numbers derived? Thank you.

[leondgarse]

• src = np.array([[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.729904, 92.2041]], dtype=np.float32) are points where typical face left_eye, right_eye, nose, mouse_left, mouse_right located in a 112 x 112 image. Detected face key points are aligned to these points.

  src = np.array([[38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.729904, 92.2041]], dtype=np.float32)
  aa = np.ones([112, 112, 3])
  plt.imshow(aa)
  plt.scatter(src[:, 0], src[:, 1])

  

• nimgs = (np.array(nimgs) - 127.5) * 0.0078125 is rescaling image in [0, 255] to [-1, 1], where 0.0078125 == 1 / 128. Several ways can be applied for this aim:

  img = (img - 128) / 128
  # Or
  img = img / 128 - 1

  Not a big difference.

[shounakshastri]

Ah ok! Thank you!