A keras port of swook/GazeML for eye region landmarks detection.
The dlib face detector is replaced by MTCNN.
Here is the demo jupyter notebook, or try it on Colaboratory.
- 
The preprocessing is very different (including not flipping left eye) from the official implementation, thus the results produced are suboptimal.
 - 
Model training as well as gaze estimation has not been ported yet.
 
- python 3.6
 - keras 2.2.4
 - tensorflow 1.12.0
 
ELG model weights are converted from the official repo swook/GazeML. We learnt a lot from there.



