---
license: cc-by-nc-sa-4.0
task_categories:
  - object-detection
tags:
  - table-tennis
  - sports
  - ball-detection
  - yolo
  - yolov8
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: "images/game_*/*.jpg"
      - split: test
        path: "images/test_*/*.jpg"
---

# OpenTTGames ball-detection YOLOv8 subset

A YOLOv8-format training subset extracted from the
[OpenTTGames](https://lab.osai.ai/) dataset. Every source-video frame for
which OpenTTGames provides a ball-centre annotation is included as a
single JPEG + a single-row YOLO label (class `0` = ball).

> **Dataset viewer note.** The viewer above only renders the JPEG
> images. Per-frame YOLO bounding-box labels live alongside each image
> at `labels/<match_id>/<frame_index>.txt`, one label file per image,
> paired by filename. The viewer doesn't render the bbox overlays.
> Consume the dataset via Ultralytics / your training pipeline of choice
> to get image + label pairs.

## Source

Derivative work of [OpenTTGames](https://lab.osai.ai/), the table-tennis
perception dataset by OSAI. The original dataset is described in:

> Voeikov, R., Falaleev, N., & Baikulov, R. (2020). *TTNet: Real-time
> temporal and spatial video analysis of table tennis*. arXiv:2004.09927.
> https://arxiv.org/abs/2004.09927

Original archives at `https://lab.osai.ai/datasets/openttgames/data/` —
`<match>.mp4` + `<match>.zip` siblings, twelve matches total
(`game_1`–`game_5` train, `test_1`–`test_7` test).

## License

This subset is licensed under
[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/),
matching the source dataset's license. By the ShareAlike clause, any
further derivatives must also be licensed under CC BY-NC-SA 4.0 or a
compatible license. **NonCommercial use only.**

Attribution: OSAI, https://lab.osai.ai/ — see citation above. See the
`LICENSE` file in this repo for the full text reference.

## Extraction recipe

For every frame whose integer index appears as a key in the match's
`ball_markup.json` (i.e. every frame OpenTTGames has annotated with a
ball-centre point):

1. Seek the source mp4 to that frame index.
2. JPEG-encode the decoded frame (quality 95).
3. Emit a YOLO label with a single row: class `0` (ball), centre at the
   annotated `(x, y)` pixel coordinate, bounding box a fixed 32 px
   square (16 px half-size), normalised against the frame dimensions.

Frames absent from `ball_markup.json` (ball occluded, out of frame, or
not annotated) are not included.

## Layout

```
images/<match_id>/<frame_index>.jpg
labels/<match_id>/<frame_index>.txt
data.yaml
README.md
LICENSE
```

Where `<match_id>` is one of `game_1`..`game_5` (train) or
`test_1`..`test_7` (test).

## Class map

| id | name |
|----|------|
| 0  | ball |

## Intended use

Research and non-commercial single-class ball-detection model training.
No commercial use.