FIFA Skeletal Tracking Starter Kit

This repository provides a naïve baseline for the FIFA Skeletal Tracking Challenge. It includes a simple, fully documented implementation to help participants get started with 3D pose estimation using bounding boxes, skeletal data, and camera parameters.

📌 Features

• Baseline Implementation: A simple approach for 3D skeletal tracking.
• Camera Pose Estimation: Computes camera transformations from bounding box correspondences.
• Field Markings Refinement: Improves camera rotation using detected Field Markings.
• Pose Projection & Optimization: Projects 3D skeletons onto 2D images and refines translation via optimization.

🚀 Getting Started

📦 Installation

Make sure you have the required dependencies installed:

pip install numpy torch opencv-python tqdm scipy

📂 Data Preparation

The script expects the following dataset structure:

data/
│── cameras/
│   ├── sequence1.npz
│   ├── sequence2.npz
│── images/
│   ├── sequence1/
│   │   ├── 00001.jpg
│   │   ├── 00002.jpg
│── skel_2d.npz
│── skel_3d.npz
│── boxes.npz
│── pitch_points.txt

• images/: Stores frame images for each sequence. Please Ensure that the filenames are sequentially numbered (e.g., "00000.jpg", "00001.jpg", etc.).
• cameras/: Contains .npz files with camera parameters for each sequence.
• boxes.npz: Stores bounding box data for each sequence.
• skel_2d.npz: Contains estimated 2D skeletal keypoints.
• skel_3d.npz: Contains estimated 3D skeletal keypoints.

You can find details about the cameras, bounding boxes, and images on the Kaggle page. For skel_2d.npz and skel_3d.npz, you can generate them automatically using the provided preprocess.py script. Alternatively, we have also uploaded preprocessed data here.

📺 Sample Visualization

To help you visualize the results, we provide a short sample sequence in media/sample.mp4.

🔧 Running the Baseline

To run the baseline model on the dataset, simply execute:

python baseline.py

By default, the script reads from the data/ directory and generates a .npz file (dummy-solution.npz) in the root folder:

You can then use the prepare-submission.py to create a submission file:

python prepare-submission.py -i dummy-solution.npz

📌 Notes

• This is a naïve baseline — you are encouraged to improve the accuracy by refining camera estimation, leveraging better keypoint tracking, or integrating deep learning approaches.

🤝 Contributing

If you find a bug or have suggestions for improvements, feel free to submit a pull request or open an issue.

Acknowledgement

We use 4DHuman in the preprocess.py for estimating both 2D and 3D skeletons from bounding boxes. We appreciate the contributions of the developers and the broader research community in advancing human pose estimation.

📜 License

This project is licensed under the MIT License.