AthleticsPose: Authentic Sports Motion Dataset on Athletic Field and Evaluation of Monocular 3D Pose Estimation Ability (MMSports at ACMMM 2025)

This repository contains the implementation for our paper "AthleticsPose: Authentic Sports Motion Dataset on Athletic Field and Evaluation of Monocular 3D Pose Estimation Ability". We introduce a comprehensive dataset and evaluation framework for 3D pose estimation in athletics scenarios.

• 2025-08-10: Added simple usage of codebase.
• 2025-08-04: Added dataset download instructions and license files
• 2025-07-18: Initial repository setup and basic file upload

To get started, clone this repository:

git clone https://github.com/your-username/AthleticsPose.git
cd AthleticsPose

Use the provided Makefile to set up the environment and download the dataset and pretrained checkpoints. This requires make and uv to be installed on your system. We also assume CUDA 11.8 is available for GPU acceleration.

If you don't have make, install it via your package manager:

• Ubuntu/Debian: sudo apt-get install make
• macOS: brew install make

If you don't have uv, install it with:

curl -LsSf https://astral.sh/uv/install.sh | sh

or on macOS with Homebrew:

brew install uv

Then run the initial setup:

make setup

This will:

1. Create and install the Python virtual environment with all dependencies
2. Download and extract the dataset and pretrained checkpoints

Environment only (no dataset/checkpoint download):

make venv

Dataset and checkpoints only (no environment setup):

make download

You can download only the AthleticsPose dataset using the following command:

curl -L -o data.zip \
  "https://github.com/SZucchini/AthleticsPose/releases/latest/download/data.zip"

or

wget -O data.zip \
  "https://github.com/SZucchini/AthleticsPose/releases/latest/download/data.zip"

For the model trained on AthleticsPose dataset, run:

uv run python scripts/evaluate.py evaluation=default

For the model trained on Human3.6M dataset, run:

uv run python scripts/evaluate.py evaluation=h36m_pretrained

For the model trained on AthletePose3D dataset, run:

uv run python scripts/evaluate.py evaluation=ap3d_pretrained model=small

To train a model from scratch, use the following command:

uv run python scripts/train.py exp_name=<your_experiment_name> \
  data.input_2d_type=det \  # or gt
  data.det_model_type=ft \  # or pretrained
  wandb.project=<your_wandb_project_name>  # You can use WandB for logging

For example, to run the predictions and analyses similar to those in the paper, use:

uv run python scripts/predict.py prediction=from_2d_markers prediction.input.marker_type=det_ft
uv run python scripts/analyze_predictions.py data/AthleticsPoseDataset/predictions

• Add dataset files and download instructions
• Complete README.md with detailed setup and usage instructions
• Implement inference functionality for arbitrary video inputs
• Provide pretrained model weights
• Add documentation for dataset structure
• Add documentation for running methods
• Add license for codes

If you find this work useful, please consider citing our paper:

@misc{suzuki2025athleticsposeauthenticsportsmotion,
      title={AthleticsPose: Authentic Sports Motion Dataset on Athletic Field and Evaluation of Monocular 3D Pose Estimation Ability},
      author={Tomohiro Suzuki and Ryota Tanaka and Calvin Yeung and Keisuke Fujii},
      year={2025},
      eprint={2507.12905},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.12905},
}

The AthleticsPose Dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) License. See the following files for details:

• LICENSE-DATASET - License for the dataset
• LICENSES/CC-BY-NC-SA-4.0.txt - Full license text

Note: This license applies specifically to the dataset. The code in this repository may be licensed differently.

For questions or issues, please open an issue on GitHub or contact the authors.