This repo reproduce the result from mediapipe FaceMesh on Raspberry pi or other platform with Coral EdgeTPU. The model first performs face detection and then align the face. The aligned faces are fed into mesh generator for face landmark detection.
- add head pose estimation
- add kalman filter
The performance for 1 face on Edge TPU, make sure the Pi has enough power otherwise the performance drops a lot!
| platform | Desktop | Raspberry Pi 4 |
|---|---|---|
| TPU | 44FPS | 14.5FPS |
| No TPU | 35FPS | 7.7FPS |
Install the requirements by:
pip install -r requirements.txt
Then run the demo with:
python main.py
I borrowed the model from: PINTO_model_zoo