Contrastive Learning for Sports Video: Unsupervised Player Classification

This repository provides code for our paper Contrastive Learning for Sports Video: Unsupervised Player Classification (Maria Koshkina, Hemanth Pidaparthy, James H. Elder).

Publication

If you use this code or the dataset please cite:

@inproceedings{koshkina2021contrastive,
  title={Contrastive Learning for Sports Video: Unsupervised Player Classification},
  author={Koshkina, Maria and Pidaparthy, Hemanth and Elder, James H},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4528--4536},
  year={2021}
}

Setup

Clone this repository.

git clone https://github.com/mkoshkina/teamId
cd teamId

Code has been written and test on Python 3.7. Install dependencies by running:

pip3 install -r requirements.txt

Download and unzip data into data directory. Our player classification code runs on player images segmented out from original video frames. Dataset will be released soon. To be notified when it becomes available please contact Maria Koshkina.

Download and unzip pre-trained models into trained_models directory.

Usage

To run evaluation of player clustering based on embedding features run:

python test_with_players_only.py with method=net

Alternatively, specify method=hist (for histogram), method=bag (for bag of colours), method=ae (for autoencoder). Refer to the paper for detailed method description.

Code Organization

• models.py - models for referee classifier, embedding network, autoencoder
• embedding_network.py - training code for embedding network
• referee_classifier.py - training and test for referee classifier
• autoencoder.py - training for autoencoder
• dataloader.py, utils.py - helper methods and constants; dataset split into train and test is defined in utils.py
• test_with_players_only.py - test player clustering using diffrerent features

Full Workflow

• Train referee classifier on the ground truth labels:

  python referee_classifer.py

• For convenience, we ran the referee classifier on all segmented images to save a list of predicted players_only

  python referee_classifier.py --save

• Using players_only images to train embedding network (and autoencoder for comparison):

  python embedding_network.py python autoencoder.py

• Run experiments using embedding network, histogram, bag of colors, or autoencoder features:

  python test_with_players_only.py with method=<method_name> , where method_name is one of net, hist, bag, ae

License

This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported License.