FootAndBall Save

FootAndBall: Integrated player and ball detector

FootAndBall: Integrated player and ball detector VISAPP 2020, Valletta, Malta.

Created by Jacek Komorowski, Grzegorz Kurzejamski and Grzegorz Sarwas at Sport Algorithmics and Gaming

Abstract

The paper describes a deep neural network-based detector dedicated for ball and players detection in high resolution, long shot, video recordings of soccer matches. The detector, dubbed FootAndBall, has an efficient fully convolutional architecture and can operate on input video stream with an arbitrary resolution. It produces ball confidence map encoding the position of the detected ball, player confidence map and player bounding boxes tensor encoding players’ positions and bounding boxes. The network uses Feature Pyramid Network desing pattern, where lower level features with higher spatial resolution are combined with higher level features with bigger receptive field. This improves discriminability of small objects (the ball) as larger visual context around the object of interest is taken into account for the classification. Due to its specialized design, the network has two orders of magnitude less parameters than a generic deep neural network-based object detector, such as SSD or YOLO. This allows real-time processing of high resolution input video stream.

System overview

Citation

If you find our work useful, please consider citing:

@conference{footandball,
author={Jacek Komorowski and Grzegorz Kurzejamski and Grzegorz Sarwas},
title={Foot{A}nd{B}all: Integrated Player and Ball Detector},
booktitle={Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 5: VISAPP,},
year={2020},
pages={47-56},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0008916000470056},
isbn={978-989-758-402-2},
}

Environment and Dependencies

Code was tested using Python 3.8 with PyTorch 1.10.2 on Ubuntu 18.04 with CUDA 10.2. Dependencies:

• PyTorch (version 1.10 or above)
• opencv-python
• tqdm
• scipy
• PIL

Datasets

FootAndBall detector is trained using ISSIA-CNR Soccer dataset and SoccerPlayerDetection_bmvc17_v1 dataset (alternative link) .

Training

To train FootAndBall detector, download and decompress ISSIA-CNR Soccer and SoccerPlayerDetection (optionally) datasets, edit config.txt and set paths to downloaded datasets. Then, run:

python train_detector.py --config config.txt

Testing

The pre-trained model model_20201019_1416_final.pth is saved in models/ folder. The model was trained with ISSIA-CNR dataset (cameras 1,2,3,4) and SoccerPlayerDetection dataset (set 1). To run the trained model use the following command:

python run_detector --path datasets/issia/filmrole5.avi --weights models/model_20201019_1416_final.pth --out_video out_video.avi --device <cpu|cuda>

This will create a video with name given by out_video parameter with bounding boxes around the ball and players position. Detection confidence level (in 0..1 range) will be displayed above each bounding box. Exemplary videos with detection results on ISSIA-CNR Soccer dataset can be downloaded here: camera 5 camera 6 .

License

Our code is released under the MIT License (see LICENSE file for details).