README.md

---
license: cc-by-nc-4.0
task_categories:
- object-detection
language:
- ja
tags:
- manga
- speech-bubble
- comics
- yolo
- computer-vision
- object-detection
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: images/train/*.jpg
  - split: validation
    path: images/val/*.jpg
  - split: test
    path: images/test/*.jpg
---

# Manga Speech Bubble Detection Dataset

Dataset for detecting speech bubble locations in manga pages.  
**2 671 images** — 1 class: `location-of-bubbles`.

## Dataset Structure

| Split | Images | Labels |
|-------|-------:|-------:|
| train | 2 056  | 2 056  |
| val   |   404  |   404  |
| test  |   211  |   211  |

```
images/
  train/   # 2056 × .jpg
  val/     # 404  × .jpg
  test/    # 211  × .jpg
labels/    # YOLO .txt, mirrors images/
dataset.yaml
manga_bubbles_detect.py   # HF loading script
```

## Annotation Format

**YOLO** — each `.txt` contains one row per bounding box:

```
<class_id> <cx> <cy> <w> <h>   # normalized [0..1], class 0 = location-of-bubbles
```

When loaded via `load_dataset()` the script converts bboxes to **COCO format** (absolute pixels `[x_min, y_min, width, height]`) and also exposes the raw YOLO values in `bbox_yolo`.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("PSImera/manga_bubbles_detect")

sample = ds["train"][0]
print(sample["image_id"])          # e.g. "img_000001"
print(sample["width"], sample["height"])
for obj in sample["objects"]:
    print(obj["category"], obj["bbox"])   # [x_min, y_min, w, h] absolute px
```

## Sources

1. **Roboflow** — [manga-6puie](https://universe.roboflow.com/diplom-uhct7/manga-6puie) (pre-labeled)
2. **Google Drive / ikefir34/DLS_Manga_Translator** — [Drive](https://drive.google.com/drive/folders/198OVEXLxY9hyhC0bdALxtR66BtBHp_Oj) · [GitHub](https://github.com/ikefir34/DLS_Manga_Translator) (labeled manually in CVAT)

## License

[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — **non-commercial use only**.

The underlying manga images remain the intellectual property of their respective publishers.
This dataset is intended solely for academic research and non-commercial purposes.
Do not use for commercial applications without obtaining rights from the original copyright holders.