Hopenet is an accurate and easy to use head pose estimation network.
First faces are detected using Dockerface, a Faster R-CNN trained on faces and deployed in a Docker image ( If the axis is jittery it is because of the center of the face bounding box is jittery.
Then head pose is obtained using Hopenet (
If a head doesn’t display an output this is because the face wasn’t detected by Faster R-CNN.
The blue axis points forward, the red axis is horizontal and the green axis is vertical.
You can check out the paper here:
source