On the representation and methodology for wide and short range head pose estimation

We provide Python code in order to replicate the head pose experiments in our paper https://doi.org/10.1016/j.patcog.2024.110263

If you use this code for your own research, you must reference our journal paper:

On the representation and methodology for wide and short range head pose estimation
Alejandro Cobo, Roberto Valle, José M. Buenaposada, Luis Baumela.
Pattern Recognition, PR 2024.
https://doi.org/10.1016/j.patcog.2024.110263

Requisites

• images_framework https://github.com/pcr-upm/images_framework
• pytorch (v1.13.0)
• sciPy

Installation

This repository must be located inside the following directory:

images_framework
    └── alignment
        └── opal23_headpose

Usage

usage: opal23_headpose_test.py [-h] [--input-data INPUT_DATA] [--show-viewer] [--save-image]

• Use the --input-data option to set an image, directory, camera or video file as input.

• Use the --show-viewer option to show results visually.

• Use the --save-image option to save the processed images.

  usage: Alignment --database DATABASE

• Use the --database option to select the database model.

  usage: Opal23Headpose [--gpu GPU]

• Use the --gpu option to set the GPU identifier (negative value indicates CPU mode).

usage: Opal23Headpose [--rotation-mode {euler,quaternion,6d,6d_opal}]

• Use the --rotation-mode option to specify the internal pose parameterization of the network.

  > python images_framework/alignment/opal23_headpose/test/opal23_headpose_test.py --input-data images_framework/alignment/opal23_headpose/test/example.tif --database 300wlp --gpu 0 --rotation-mode euler --save-image

Notebooks

The directory notebooks contains examples regarding some of the contributions of our paper:

• Loading annotations: this notebook contains a simple implementation of a PyTorch Dataset class that uses our annotations for CMU Panoptic Dataset and shows sample images.

• Prediction alignment: this notebook shows an example on how to align model predictions to reduce systematic errors in cross-dataset evaluations.

• Opal loss: this notebook shows an usage example of our Opal Loss function for Head Pose Estimation.