deep-head-pose

Why we included this project

Hopenet, the model at the heart of this repo, estimates head orientation directly from a detected face and has become one of the more widely referenced PyTorch implementations of the idea. Trained on the 300W-LP dataset, it returns yaw, pitch, and roll angles without relying on facial keypoints, which makes it easy to slot into driver-monitoring, gaze-tracking, and VR or AR calibration work. Pre-trained weights ship with the repo, along with ready-made scripts that run inference on video using either dlib or your own face detections, so you can try it on real footage before committing to an integration. It is a research project rather than a maintained library, and you will need a GPU, but teams already comfortable in Python and PyTorch should find the code clear enough to adapt. The paper and the linked example videos help you judge how the output holds up on real-world input.