/eigd

Official project website of the paper "A Unified Taxonomy and Multimodal Dataset for Events in Invasion Games" accepted at MMSports'21

Primary LanguageJupyter Notebook

A Unified Taxonomy and Multimodal Dataset for Events in Invasion Games

Conference arXiv

Official repository for the paper:

Biermann, H., Theiner, J., Bassek, M., Raabe, D., Memmert, D., & Ewerth, R. (2021, October). A Unified Taxonomy and Multimodal Dataset for Events in Invasion Games. In Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports (pp. 1-10).

Contents

Events in Invasion Games Dataset - Handball (EIGD-H)

This dataset contains the broadcast video streams of handball matches along with synchronized official positional data and human event annotations for 125min raw data in summary.


Data Source & Characteristics

  • Handball matches from the Handball-Bundesliga (HBL) captured in saison 2019/20
  • Size: 5 matches x 5 sequences x 5min
  • Video:
    • unedited broadcast video stream (no cuts, no overlays)
    • HD resolution (1280x720px)@30fps
  • Positional data:
    • official captured by Kinexon
    • manually synchronized to video streams (offsets and sampling rate (originally captured at 20Hz))
  • Events:
    • frame-wise annotations based solely on the video content
    • annotations according to the proposed taxonomy
    • multiple annotations for two matches (10 sequences) from 3 experts
    • hierarchical event format: <root_event>.<sub_event>.<sub_sub_event>
    • statistics: [event_statistics.ipynb]

License

Position and video data are provided by Kinexon with authorization of the Handball-Bundesliga (HBL). As EIGD-H is licensed under CC BY-NC-SA 4.0 you must give appropriate credit when using this dataset by

  1. naming the Handball-Bundesliga (HBL)
  2. citing this publication

Download

You can download the annotations, position and video data manually at https://data.uni-hannover.de/dataset/eigd or automatically using download_eigd.sh:

Visualization Positional Data

See visualize_positional_data.ipynb

Events in Invasion Games Dataset - Soccer (EIGD-S)

Annotations and URLs to the videos are available at https://data.uni-hannover.de/dataset/eigd .

  • Videos are captured from the official FIFA youtube channel
  • Size: 5 matches x 5 sequences x 5min
  • Video:
    • edited broadcast video stream
    • HD resolution (1280x720px)@25fps
  • Events:
    • frame-wise annotations based solely on the video content
    • annotations according to the proposed taxonomy
    • multiple annotations for two matches (10 sequences) from 4 experts and one inexperienced annotator
    • hierarchical event format: <root_event>.<sub_event>.<sub_sub_event>

Human Performance Evaluation

To measure the aggreement of multiple annotators, i.e. the expected human performance, you can use these two notebooks (evaluate_eigd-h.ipynb and evaluate_eigd-s.ipynb) to reproduce the results of the paper. The formatted output is also accessbile here.

Annotation Guidelines and Event Definitions

See definitions.md and examples.md.

Citation

@inproceedings{BiermannTaxonomyMMSports21,
author = {Biermann, Henrik and Theiner, Jonas and Bassek, Manuel and Raabe, Dominik and Memmert, Daniel and Ewerth, Ralph},
title = {A Unified Taxonomy and Multimodal Dataset for Events in Invasion Games},
year = {2021},
isbn = {9781450386708},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3475722.3482792},
doi = {10.1145/3475722.3482792},
booktitle = {Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports},
pages = {1–10},
numpages = {10},
keywords = {event detection, human performance analysis, datasets, events in sports},
location = {Virtual Event, China},
series = {MMSports'21}
}