qa_headpose_regression

Head Pose Regression Using Quantization Aware MobilenetV2

State of the art head pose detection ( yaw, roll, pitch ) dated to @2020-2021

Estimate three angles using a face image, the angles correspond to yaw, roll, pitch. ( note that the face image should be cropped using some tool, lets say mtcnn or hog )

Consider that some of this is inspired from FSANet at the time FSANet, while imporving and speeding up inference, the model achieves the same inferece time as FSANet on GPU (3ms), using only tflite and 8 cpu cores.

Examples

( YAW PITCH ROLL order ):

Model Prediction: [ 47. -45. -39.]
Ground Truth: [ 37.76773363 -51.35657112 -34.43247194]

Model Prediction: [ 68. 12.  4. ]
Ground Truth: [ 71.62278432 13.5477656   9.57131228 ]

Results

BIWI
YAW: 3.8884045387677673
PITCH: 4.948589362227766
ROLL: 2.656535269831331

MAE: 3.8311763902756213
=========================
AFLW
YAW: 4.029069105922883
PITCH: 5.747231110398772
ROLL: 3.975032158809743

MAE: 4.583777458377132

Comparison with SOTA

@misc{head_pose_regression,
  title =        {Head Pose Regression Using Quantization Aware MobilenetV2},
  author =       {Abdolkarim Saeedi},
  publisher =    {GitHub},
  howpublished = {\url{[https://github.com/KiLJ4EdeN/qa_headpose_regression](https://github.com/KiLJ4EdeN/qa_headpose_regression)https://github.com/KiLJ4EdeN/qa_headpose_regression}},
  year =         {2024}
}