model_run_and_postprocess.py

def post_process(self, outputs):
  loc, conf, landmarks = outputs
  height, width = 640, 640

  priorbox = PriorBox(image_size=(height, width))
  priors = priorbox.forward()
  variance = [0.1, 0.2]
  scale = np.array([width, height, width, height])
  scale1 = np.array(
      [width, height, width, height, width, height, width, height, width, height]
  )
  # bboxe
  boxes = decode(loc[0], priors, variance)
  boxes = boxes * scale
  # score
  scores = np.squeeze(conf, axis=0)[:, 1]
  # landmark
  landmarks = decode_landm(landmarks[0], priors, variance)
  landmarks = landmarks * scale1
  # ignore low scores
  inds = np.where(scores > self.conf_threshold)[0]
  boxes, landmarks, scores = boxes[inds], landmarks[inds], scores[inds]
  # sort
  order = scores.argsort()[::-1]
  boxes, landmarks, scores = boxes[order], landmarks[order], scores[order]
  # do NMS
  bounding_boxes = np.hstack((boxes, scores[:, np.newaxis])).astype(
      np.float32, copy=False
  )
  keep = nms_boxes(
      bounding_boxes[:, :4], bounding_boxes[:, 4], self.nms_threshold
  )
  bounding_boxes, landmarks = bounding_boxes[keep, :], landmarks[keep]

  return np.concatenate((bounding_boxes, landmarks), axis=1)



resizedImg = pre_process(mainImg)
start = int(round(time.time() * 1000))

outputs = ort_session.run(None, {"input": resizedImg})
bboxes = post_process(outputs)
_, all_landmarks_5 = get_face_landmarks_5(bboxes, eye_dist_threshold=5)

new_landmarks=find_landmark_orginal_image(mainImg,all_landmarks_5)

cropped_faces, affine_matrices = align_warp_face(mainImg, new_landmarks)

end = int(round(time.time() * 1000))
estimation_time = end - start
logger.info(f"'RetinaFace infer time: {estimation_time} ms")
save_faces(cropped_faces,"result_faces")



return cropped_faces, affine_matrices