APeiZou / DMHead

Dual model head pose estimation. Fusion of SOTA models. 360° 6D HeadPose detection.

DMHead

Dual model head pose estimation. Fusion of SOTA models. 360° 6D HeadPose detection.

1. Summary

2. [WIP] Atmosphere

Kazam_screencast_00091_.mp4

3. [WIP] Benchmark

• Trained on 300W-LP (Custom, Mask-wearing face image augmentation)
• Test on AFLW2000
  • June 22, 2022

    Yaw: 3.6129, Pitch: 5.5801, Roll: 3.8468, MAE: 4.3466
    

4. Model Structure

• INPUTS: Float32 [N,3,224,224]
• OUTPUTS: Float32 [N,3], [Yaw,Roll,Pitch]

5. Citation

@misc{https://doi.org/10.48550/arxiv.2005.10353,
    doi = {10.48550/ARXIV.2005.10353},
    url = {https://arxiv.org/abs/2005.10353},
    author = {Zhou, Yijun and Gregson, James},
    title = {WHENet: Real-time Fine-Grained Estimation for Wide Range Head Pose},
    publisher = {arXiv},
    year = {2020},
}

@misc{hempel20226d,
    title={6D Rotation Representation For Unconstrained Head Pose Estimation},
    author={Thorsten Hempel and Ahmed A. Abdelrahman and Ayoub Al-Hamadi},
    year={2022},
    eprint={2202.12555},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}