This repository is a reference implementation for the inference part of "Optimizing Through Learned Errors for Accurate Sports Field Registration", WACV 2020. For more details, please refer to our WACV 2020 or [arXiv] paper. A video showing the results is available [here]
The released code is freely available for free non-commercial academic research use, and may be redistributed under these conditions. Any commercial use is prohibited.
Note: We decided to not release the training code. Sorry for any inconvenience.
The Optimization Based Image Registration is patent protected (pending applications US 62/850,910; US 16/049,546; EP17746676.0; CA 3,012,721)[here] and shall not be used for any commercial application. For information about licensing please contact If you are interested in a commercial license, contact Sportlogiq or SLiQ Labs.
- The trained weights (both initial guess net, and loss surface net) for soccer.
- The inference code for soccer.
- A Jupiter notebook to for simple user interaction.
- The code to generate a soccer field template(Processing language) and a h5 format test dataset used in the paper.
This implementation is based on Python3 and PyTorch.
You can install the environment by: conda env create -f environment.yml
Activate the env by: conda activate sportsfield
We provide the pretrained weights for soccer on Google drive. Download "out.zip", and extract all the content to ./out, such that the ./out folder contains pretrained_init_guess and pretrained_loss_surface .
Users can overlay the template to a soccer image or video using the notebook.
Users can simply run: python test_end2end.py loss_surface init_guess --load_weights_upstream "pretrained_init_guess" --load_weights_error_model "pretrained_loss_surface" --batch_size 32 to start the evaluation.
A reference evaluation result is provided for comparison:
----- Summary -----
original IOU part mean: 0.90211654
original IOU part median: 0.91872334
original IOU whole mean: 0.8406853
original IOU whole median: 0.857767
optimized IOU part mean: 0.9530167
optimized IOU part median: 0.9701195
optimized IOU whole mean: 0.9019278
optimized IOU whole median: 0.9253305
----- -----
spent 290.74491572380066 seconds for 186 images
1.5631447081924768 seconds per single image
----- End -----
If you use this code in your research, cite the paper:
@inproceedings{jiang2020optimizing,
author={Wei Jiang and Juan Camilo Gamboa Higuera and Baptiste Angles and Weiwei Sun and Mehrsan Javan and Kwang Moo Yi},
booktitle={2020 IEEE Winter Conference on Applications of Computer Vision (WACV)},
title={Optimizing Through Learned Errors for Accurate Sports Field Registration},
year={2020},
organization={IEEE}
}
The released code is freely available for free non-commercial academic research use, and may be redistributed under these conditions. Please, see the license for further details. If you are interested in a commercial license, contact Sportlogiq or SLiQ Labs for licensing information.
Note: The Optimization Based Image Registration is patent protected (pending applications US 62/850,910; US 16/049,546; EP17746676.0; CA 3,012,721) and shall not be used for any commercial application.