cv2 → mediapipe ( library order should be like this)
Previously mentioned, the dlib and YOLO combination model faces a bottleneck problem since dlib runs only on CPU while YOLO runs on GPU, making it impractical for Jetson Nano.
Therefore, we opted to use MediaPipe exclusively to leverage GPU capabilities.
MediaPipe uses two lightweight AI architectures: BlazeFace for face detection and MobileNet for feature detection.
MediaPipe provides a face landmarks detection model that can identify 468 distinct features.
total face
0 ~ 467: Complete 3D face mesh (based on basic model)
👀(Eyes)
Right Eye: Multiple points between 33 ~ 133
Outline: 33, 133, 160, 159, 158, 157, 173 etc.
Iris: 468 ~ 472 (only when refine_landmarks=True option is used)
Left Eye: 263 ~ 362
Outline: 263, 362, 387, 386, 385, 384, 398 etc.
Iris: 473 ~ 477 (additional model)
👄 Lips
Full lips:
Outer contour: 61, 146, 91, 181, 84, 17, ...
Inner contour: 78, 191, 80, 81, 82, 13, ...
👃 Nose
Nose center (tip): 1
Nostrils: 97, 2, 326
👂 Ears and Jawline
Chin point: 152
Right ear reference point: 234
Left ear reference point: 454