Pose estimation baseline for the 3D Human Pose Estimation benchmark of the HARPER dataset (IROS 2024)

This repository contains the code for the baseline of the 3D Human Pose Estimation benchmark on the HARPER dataset. The baseline is based on the HRNet architecture and uses the depth maps captured by the Spot to estimate the 3D pose of the partially-visible human body.

Quick start

The quick start guide will be available soon.

Installation

You can install the required packages following the steps here.
You can find the pretrained HRNet model here. To use it, modify the TEST.MODEL_FILE parameter in the config file (experiments/harper/hrnet/w32_256x256_adam_lr1e-3_harper.yaml) with the correct path.

Data preparation

Follow the steps in the HARPER official repository to download the dataset and prepare the data.
Modify the DATASET.ROOT parameter in the config file with the correct path.

Credits

This code is based on the HRNet architecture, forking this implementation.

Citation

If you use this code in your research, please cite the following paper (IROS 2024 citation coming soon):

@article{avogaro2024exploring,
    title={Exploring 3D Human Pose Estimation and Forecasting from the Robot's Perspective: The HARPER Dataset},
    author={Avogaro, Andrea and Toaiari, Andrea and Cunico, Federico and Xu, Xiangmin and Dafas, Haralambos and Vinciarelli, Alessandro and Li, Emma and Cristani, Marco},
    journal={arXiv e-prints},
    pages={arXiv--2403},
    year={2024}
}