jahid114593/foul-vs-dive-dataset

CAS-FD: Single-View Foul vs. Dive Dataset

A balanced 600-clip single-view broadcast video dataset for binary Foul vs. Dive classification in association football. Every clip is 5 seconds long at 1280×720 / 25 fps, drawn from a single live broadcast camera (no replays, no tactical overlays), with class-balanced train / validation / test splits.

To our knowledge this is the largest publicly available single-view Foul/Dive corpus, approximately 24× larger than the only previously reported single-view set (25 clips, Al-Theiabat & Aljarrah, 2018), which is not publicly distributed.

Dataset Summary

Property	Value
Total clips	600
Classes	2 (Foul, Dive)
Class balance	300 / 300
Resolution	1280 × 720 px
Frame rate	25 fps
Clip duration	5 seconds (≈125 frames)
Container / codec	MP4 / H.264
Source view	Single broadcast camera (live feed only)
Splits	Train 400 / Val 100 / Test 100
Per-split balance	50 / 50 enforced in every split

Splits

Split	Foul	Dive	Total
Train	200	200	400
Validation	50	50	100
Test	50	50	100
Total	300	300	600

Splits are fixed and shipped as directory partitions. Please don't re-shuffle if you want results comparable to others using this dataset.

Directory Structure

foul-vs-dive-dataset/
├── README.md
├── train/
│   ├── Foul/
│   │   ├── action_0_clip_0.mp4
│   │   ├── action_1_clip_0.mp4
│   │   └── ...
│   └── Dive/
│       └── ...
├── val/
│   ├── Foul/
│   └── Dive/
└── test/
    ├── Foul/
    └── Dive/

The class label of each clip is its parent directory name. Clip-level binary labels are the only annotation provided; no frame-level or pixel-level labels are included.

Loading the Dataset

Option A — datasets videofolder loader

from datasets import load_dataset

ds = load_dataset(
    "videofolder",
    data_dir="path/to/foul-vs-dive-dataset",
    drop_labels=False,
)

print(ds)
# DatasetDict({
#   train: Dataset({ features: ['video', 'label'], num_rows: 400 })
#   validation: Dataset({ features: ['video', 'label'], num_rows: 100 })
#   test: Dataset({ features: ['video', 'label'], num_rows: 100 })
# })

Option B — direct download with huggingface_hub

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="jahid114593/foul-vs-dive-dataset",
    repo_type="dataset",
)

Option C — clone with git

git clone https://huggingface.co/datasets/jahid114593/foul-vs-dive-dataset

Option D — read a single clip with Decord

from decord import VideoReader

vr = VideoReader("foul-vs-dive-dataset/train/Foul/action_0_clip_0.mp4")
print(len(vr), vr[0].shape)   # ~125 frames, (720, 1280, 3)

How the Dataset Was Built

Foul clips (n = 300)

Extracted from the publicly available SoccerNet-v2 action spotting dataset. Each annotated foul timestamp was used to extract a 5-second window of the live broadcast feed centred approximately on the event. Replay segments, tactical overlay views, and slow-motion inserts were excluded. Clips span multiple European professional leagues and varied camera angles, lighting conditions, pitch zones, and player configurations.

Dive clips (n = 300)

Curated manually using the Stop Diving initiative (https://www.stopdiving.org/) as a catalogue of verified incident timestamps. Only timestamps were used; no frames, video files, or processed data from Stop Diving were redistributed. The corresponding full-broadcast footage was obtained separately and trimmed to 5-second windows around each referenced timestamp. Every candidate clip was manually reviewed to verify that the fall-onset moment was contained within the 5-second window; clips outside this criterion or of insufficient visual quality were discarded.

Annotation philosophy

• Single binary label per clip. No frame-level supervision required.
• Marginal cases excluded. Genuinely ambiguous incidents (e.g. real contact deliberately exaggerated) were excluded during curation, so the dataset trains a clean binary classifier rather than a marginal-case adjudicator. This is a deliberate scope choice and a known limitation.
• Visual-domain consistency. Both classes are restricted to professional broadcast footage under comparable production conditions.

Intended Use

This dataset is suitable for:

• Research on fine-grained contact-event action recognition in sports video
• Benchmarking temporal sampling strategies on a small, event-centric corpus
• A single-view counterpart to multi-view benchmarks such as SoccerNet-MVFouls
• Educational use in video understanding courses

Out of scope

• Real-time deployment in live officiating without further validation
• Multi-class severity prediction (the labels are strictly binary)
• Cross-domain transfer to amateur footage (broadcast professional only)
• Multi-view tasks (use SoccerNet-MVFouls instead)

Limitations

• Modest scale. 400 training clips is small by modern video-recognition standards; out-of-distribution generalisation is bounded.
• Clip-level splitting. Splits are stratified by class but not by source match. Clips from the same broadcast match may appear in different splits, which can inflate validation accuracy through stadium / kit-colour shortcuts. Match-level re-splitting is recommended for stricter benchmarks.
• In-group annotations. Clip labels were assigned by the dataset authors with an internal 96.7% consistency check; no external inter-annotator agreement study has been conducted.
• Binary simplification. Real refereeing decisions span a continuum (careless / reckless / excessive force; marginal / clear simulation). Multi-attribute labelling is left to future work.
• Single-view ceiling. Multi-view systems benefit from close-up replay angles that single-view inherently cannot access.

Ethical Considerations and Responsible Use

• Source. All clips are derived from publicly broadcast professional football matches. Player identities are not used as supervisory signals.
• No automated officiating claims. This dataset is for research only. It is not suitable for unsupervised deployment in real refereeing decisions. Any model outputs intended for consequential use must be reviewed by a qualified human official.
• Bias. All clips are from professional European leagues; models trained on this dataset may not generalise to amateur, women's, or non-European football without further adaptation.
• Non-commercial. Under CC BY-NC 4.0, commercial deployment (e.g. integration into a paid VAR product) is not permitted under this licence.

License

This dataset is released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International). You may use, share, and adapt the dataset for non-commercial research and educational purposes with appropriate attribution. Commercial use is not permitted under this licence.

Full text: https://creativecommons.org/licenses/by-nc/4.0/

The underlying broadcast footage may be subject to separate rights held by the original broadcasters and competition organisers. This release does not grant any rights to that underlying footage beyond what is permitted under fair-dealing / fair-use for academic research in your jurisdiction. Users are responsible for ensuring their use complies with applicable copyright law.

Citation

If you use this dataset, please cite:

@misc{islam2026foulvsdive,
  title        = {CAS-FD: Single-View Foul vs.\ Dive Dataset},
  author       = {Islam, MD. Jahidul and Alam, Mahfujul and Seyam, Nazmul Islam},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/jahid114593/foul-vs-dive-dataset}
}

Please also acknowledge the upstream sources:

• SoccerNet-v2 (Foul clip timestamps):

  @inproceedings{deliege2021soccernetv2,
    title     = {SoccerNet-v2: A Dataset and Benchmarks for Holistic Understanding of Broadcast Soccer Videos},
    author    = {Deli{\`e}ge, Adrien and others},
    booktitle = {CVPR Workshops},
    year      = {2021}
  }
  

• Stop Diving (Dive incident catalogue): https://www.stopdiving.org/

Contact

For dataset issues, corrections, or research correspondence, please open a discussion on the Hugging Face repository.