infactory-ai/ball-tracking

Soccer Ball Tracking Dataset

A curated dataset for detecting and tracking soccer balls in broadcast footage, specifically designed for tiny object detection challenges. This dataset supports the development of models robust to motion blur, long-shot scales, and occlusions.

Dataset Description

This public sample consists of 1,450 frames extracted from 10 video clips of professional soccer broadcasts. The data is split into two categories based on ball visibility:

• Visible: Frames where the ball is clearly visible and annotated with a bounding box.
• Not Visible: Frames where the ball is occluded, out of frame, or otherwise not visible (negative samples).

This is a representative subset of a larger internal dataset, selected to cover diverse match conditions (SD/HD, different teams, mined vs. segmented clips).

Statistics

Category	Samples	Description
Visible	884	Frames with at least one ball bounding box
Not Visible	566	Frames with no visible ball (hard negatives)
Total	1,450	Total frames from 10 clips

Source Data

• Domain: Professional Soccer Broadcasts
• Resolution: Varied
• Annotation Style: YOLO format (normalized xywh)
• Labeling Method: Active Learning Loop (Model Mining -> Pseudo-labeling -> Manual Verification)
• Anonymization: Source video names have been replaced with UUIDs.

Dataset Structure

infactory-ai/ball-tracking/
├── README.md
├── metadata.csv
├── dataset_info.json
└── data/
    ├── visible/
    │   ├── {uuid}_{frame}.jpg
    │   └── {uuid}_{frame}.txt   # YOLO label
    └── not_visible/
        └── {uuid}_{frame}.jpg

Metadata Fields (metadata.csv)

Field	Type	Description
file_path	string	Relative path to the image file
video_source	string	UUID of the source video clip
frame_index	int	Frame number in the original clip
visibility	string	visible or not_visible
bboxes_count	int	Number of bounding boxes in the frame

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("infactory-ai/ball-tracking", data_dir="data")

# Filter for visible frames
visible_frames = dataset.filter(lambda x: x["visibility"] == "visible")

Parsing Labels

Labels are in standard YOLO format: <class_id> <x_center> <y_center> <width> <height>

• class_id: 0 (ball)
• Coordinates are normalized to [0, 1].

Team

Name	Role
Valentino Constantinou	Head of Infrastructure
Dr. Mehdi Iranmanesh	Applied AI Engineer
John Kanalakis	Chief Technology Officer

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

You are free to:

• Share -- copy and redistribute the material in any medium or format
• Adapt -- remix, transform, and build upon the material

Under the following terms:

• Attribution -- You must give appropriate credit to Infactory, provide a link to the license, and indicate if changes were made.
• Non-Commercial -- You may not use the material for commercial purposes without a separate commercial license from Infactory.

Commercial licensing: For commercial use, contact hello@infactory.ai.

Citation

@dataset{ball_tracking_2026,
  title={Soccer Ball Tracking Dataset},
  author={Constantinou, Valentino and Iranmanesh, Mehdi and Kanalakis, John},
  year={2026},
  publisher={Infactory},
  url={https://huggingface.co/datasets/infactory-ai/ball-tracking}
}