This example shows an implementation of Facial Landmark detection that's used in MediaPipe Face Mesh. The TFLite model was taken from MediaPipe and converted to blob so that it can run on OAK devices.
Model takes input image of size 192 x 192, and predicts 468 facial landmarks and a score. Note that this demo only supports once face. For multiface support, additional face detection stage must be first implemented.
We use input image 416 x 416 and resize it to 192 x 192 using ImageManip node.
The model also includes effect renderer from 2D PNG image.
-
Purchase a DepthAI (or OAK) model (see shop.luxonis.com).
-
Install requirements.
python3 -m pip install -r requirements.txt
python3 main.py [options]
Options:
-confor--confidence_thresh: Set score threshold. If score falls bellow the threshold, facial landmarks are not shown.