FootballCamSynth: A Synthetic Football Dataset with Camera Parameters

Overview

FootballCamSynth is a synthetic dataset created using a customized version of the Google Research Football Simulator. It contains 40,000 images with camera parameter annotations, making it useful for sports field registration, camera calibration, and synthetic-to-real adaptation.

Dataset Features

40,000 high-quality images of simulated football scenes.
Randomized textures: Grass, pitch lines, and uniforms are sampled from a predefined set for diversity.
Dynamic camera settings:
- Position (x, y, z) sampled within realistic field boundaries.
- Orientation (pan, tilt, roll): The roll is fixed at 0, while pan and tilt adjust to keep the field in view.
- Field of View (FoV) varies across images.

Usage

This dataset is ideal for:
✔ Camera parameter estimation
✔ Football scene understanding
✔ Synthetic-to-real adaptation

Dataset Statistics

Below is a visualization of key dataset distributions, including camera positions, orientations, and FoV:

Annotations

camera_parameters.csv contains annotations for all images. Each image is annotated with:

{
    "image_name": "000000000.jpg",
    "aov": 1.1664075131018652,
    "c_x": 41.564503873575774,
    "c_y": 91.01672333543827,
    "c_z": -10.775780120664296,
    "pan": -1.0167846147515798,
    "roll": 0.0,
    "tilt": 1.4723548512164744,
    "h": 1080,
    "w": 1920
}

Visualization & Projection Matrix

To visualize the labeled data and compute the projection matrix, use the Jupyter notebook:

📂 src/vis_labels.ipynb

License

📜 CC-BY-NC-4.0 – Free to use for research and non-commercial purposes with attribution.