Deep Head Pose is an open-source PyTorch implementation of Hopenet, a neural network that estimates the orientation of a person’s head from a single image. With about 1,700 stars on GitHub, it predicts yaw, pitch and roll angles and has become a widely reused reference implementation for head pose estimation in computer vision.
The model was trained on the 300W-LP dataset and tested on real-world data, with the underlying method described in a CVPR Workshop paper that the repository links to. Pretrained weights and straightforward scripts are included for running estimation on images or video, using a face detector to locate heads first. Because head pose is a building block for gaze estimation, driver attention monitoring, augmented reality and human-computer interaction, the project is frequently embedded inside larger vision pipelines rather than used on its own. Setup is light โ PyTorch, OpenCV and a few standard libraries are all it needs โ and the qualitative examples in the repo show robust results on in-the-wild photos and video clips.
Deep Head Pose is free and open source; the code and pretrained models can be downloaded from GitHub at no cost. There is no hosted service or paid version โ you are responsible only for the compute it runs on.
Computer vision engineers prototyping gaze tracking, driver monitoring or AR features, and researchers who need a proven, easy-to-run head pose estimation baseline. It is research code, so expect to wire it into your own pipeline rather than install a finished application.