Legacy media timing defect confirmed: the examined clip 0LtLS9wROrk_E_000147_000152 contains 149 frames, while its padded spotting annotation requires 266. Matching annotation IDs do not establish usable media alignment. A full rebuild from the original OpenDataLab videos is in progress. The complete source archive is at https://opendatalab.com/OpenDataLab/FineGym.

This repository currently contains 422 of the 5,374 clips in the T-DEED / E2E-Spot FineGym spotting benchmark. It is not a complete paper-reproduction release. The existing subset has 214 training, 66 validation and 142 test clips; the benchmark requires 3,327, 944 and 1,103 respectively.

Available samples' event labels and reconstructed frame indices match T-DEED commit d953b4de8a2ad941c565e1683c5b07255cbe644e. All referenced videos are present in the repository; full media decoding has not yet been audited. See the coverage report for missing sample IDs.

The repository was renamed from OpenSportsLab/soccernetpro-localization-gymnastics, retaining its history, main and raw branches, and manual access gate.

FineGym annotations retain the original authors' CC BY-NC 4.0 terms: https://sdolivia.github.io/FineGym/. The library's software license does not replace the source dataset's terms.

OSL Gymnastics Localization (Action Spotting)

This repository provides a Gymnastics action spotting / localization dataset in an OpenSportsLab / SoccerNet-style format.

The dataset is organized by split (train/, valid/, test/) with video clips and corresponding localization annotations in JSON.

📌 Task

Task type: action_spotting (a.k.a. temporal action localization / event spotting)
Sport domain: Artistic Gymnastics (VT, UB, BB, FX)
Annotation granularity: clip-relative timestamps in milliseconds (position_ms)
Label format: single-label events (one label per event)

This dataset supports fine-grained temporal localization of gymnastics elements such as:

Vault phases (VT_0, VT_1, VT_2, VT_3)
Uneven Bars elements (UB_circles_start, UB_dismounts_end, etc.)
Balance Beam skills (BB_turns_start, BB_flight_salto_end, etc.)
Floor Exercise elements (FX_back_salto_start, FX_leap_jump_hop_end, etc.)

📁 Main branch structure

Current structure on main:

main/
├── annotations-localization-train.json
├── annotations-localization-valid.json
├── annotations-localization-test.json
├── train/
│   ├── <clip>.mp4
│   └── ...
├── valid/
│   ├── <clip>.mp4
│   └── ...
└── test/
    ├── <clip>.mp4
    └── ...

The three folders train/, valid/, test/ contain short gymnastics video clips (.mp4).
The three JSON files contain the localization labels for the corresponding split.

Clip filenames follow the pattern:

<YouTubeID>_E_<start>_<end>.mp4

Example:

0LtLS9wROrk_E_000147_000152.mp4

🧾 Annotation format

Each annotation file follows a SoccerNet-like schema.

Top-level keys:

version: format version (e.g., "2.0")
task: "action_spotting"
dataset_name: dataset identifier
labels: list of valid event classes under a given head_name
data: list of items (each item corresponds to one clip)

`data[]` item fields

Each item contains:

id: stable item identifier
inputs: list containing a video descriptor
events: list of labeled events in that clip
metadata: extra information such as fps, width, height, segment boundaries, etc.

Example (simplified):

{
  "id": "Gymnastics_0LtLS9wROrk_E_000147_000152",
  "inputs": [
    {
      "type": "video",
      "path": "test/0LtLS9wROrk_E_000147_000152.mp4",
      "fps": 29.97
    }
  ],
  "events": [
    {
      "head": "gymnastics_action",
      "label": "VT_0",
      "position_ms": "1201",
      "comment": "round-off, flic-flac on, stretched salto backward with 1 turn off"
    }
  ]
}

⏱️ Timestamp ↔ video position relationship (IMPORTANT)

For each event:

position_ms is clip-relative time in milliseconds.
It is computed from the clip-relative frame index using:

position_ms = round(frame / fps * 1000)

So:

position_ms = 0 corresponds to the first frame of the clip.
position_ms = 4240 means the event happens around 4.240 seconds after the clip start.

If you need the approximate frame index back:

frame ≈ round(position_ms / 1000 * fps)

🏷️ Labels

Labels are stored under:

labels.<head_name>.labels

where <head_name> is typically:

gymnastics_action

Example classes:

BB_turns_start
BB_turns_end
FX_back_salto_start
FX_back_salto_end
UB_circles_start
UB_circles_end
UB_transition_flight_start
UB_dismounts_end
VT_0
VT_1
VT_2
VT_3

Each event is single-label, meaning one action per timestamp.

🧩 Clip naming convention

Clips are derived from longer competition videos.

Pattern:

<YouTubeID>_E_<start>_<end>

Where:

<YouTubeID>: original competition video ID
_E: event version marker
<start> / <end>: start and end time in seconds within the original long video

Example:

0LtLS9wROrk_E_000147_000152

This means:

Original video: 0LtLS9wROrk
Clip covers seconds 147 to 152
The clip file is stored as:
```
test/0LtLS9wROrk_E_000147_000152.mp4
```

🧰 Notes

Paths in inputs[].path are relative paths pointing to the split folder:
- train/<clip>.mp4
- valid/<clip>.mp4
- test/<clip>.mp4
position_ms values are clip-relative, not global video timestamps.
Clips may contain padding before or after annotated actions.
The repository includes .gitattributes for Git/LFS handling of large video files.

✅ Quick sanity check

Pick one entry in annotations-localization-test.json:

Open the clip video located at:
```
test/<clip>.mp4
```
Convert position_ms to seconds:
```
seconds = position_ms / 1000
```
Jump to that time in the video.
You should observe the corresponding gymnastics element near that timestamp.

📚 Data Source & Attribution

The gymnastics clips and raw annotations in this dataset are derived from the tennis data released in the official repository of the paper:

Spotting Temporally Precise, Fine-Grained Events in Video (ECCV 2022)
James Hong, Haotian Zhang, Michaël Gharbi, Matthew Fisher, Kayvon Fatahalian

Source repository (gymnastics data):
hhttps://github.com/jhong93/spot/tree/main/data/finegym

If you use this dataset, please cite the original paper:

@inproceedings{precisespotting_eccv22,
    author={Hong, James and Zhang, Haotian and Gharbi, Micha\"{e}l and Fisher, Matthew and Fatahalian, Kayvon},
    title={Spotting Temporally Precise, Fine-Grained Events in Video},
    booktitle={ECCV},
    year={2022}
}

Downloads last month: 14

Total file size:

13 GB

You need to agree to share your Hugging Face information to access this dataset

Benchmark coverage — partial release