POOF: Efficient Goalie Pose Annotation using Optical Flow

Brennan Gebotys, Alexander Wong, David Clausi

2021

Abstract

Due to the wide range of applications for human pose estimation including sports analytics and more, research has optimized pose estimation models to achieve high accuracies when trained on large human pose datasets. However, applying these learned models to datasets that are from a different domain (which is usually the goal for many real-world applications) usually leads to a large decrease in accuracy which is not acceptable. To achieve acceptable results, a large number of annotations is still required which can be very expensive. In this research, we leverage the fact that many pose estimation datasets are derived from individual frames of a video and use this information to develop and implement an efficient pose annotation method. Our method uses the temporal motion between frames of a video to propagate ground truth keypoints across neighbouring frames to generate more annotations to provide efficient POse annotation using Optical Flow (POOF). We find POOF achieves the best performance when used in different domains than the pretrained domain. We show that in the case of a real-world hockey dataset, using POOF can achieve 75% accuracy (a +15% improvement, compared to using COCO-pretrained weights) with a very small number of ground truth annotations.

Paper Citation

in Harvard Style

Gebotys B., Wong A. and Clausi D. (2021). POOF: Efficient Goalie Pose Annotation using Optical Flow. In Proceedings of the 9th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS, ISBN 978-989-758-539-5, pages 116-122. DOI: 10.5220/0010657000003059

in Bibtex Style

@conference{icsports21,
author={Brennan Gebotys and Alexander Wong and David Clausi},
title={POOF: Efficient Goalie Pose Annotation using Optical Flow},
booktitle={Proceedings of the 9th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS,},
year={2021},
pages={116-122},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0010657000003059},
isbn={978-989-758-539-5},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 9th International Conference on Sport Sciences Research and Technology Support - Volume 1: icSPORTS,
TI - POOF: Efficient Goalie Pose Annotation using Optical Flow
SN - 978-989-758-539-5
AU - Gebotys B.
AU - Wong A.
AU - Clausi D.
PY - 2021
SP - 116
EP - 122
DO - 10.5220/0010657000003059