Datasets:

---

MS92

/

MangaSegmentation

like 2

Thanks for releasing MangaSegmentation! I’ve confirmed that:
• The dataset follows standard COCO format (RLE masks with size / counts).
• A helper script, parse_rle_manga.py, is included and correctly decodes the masks.
• Every mask has non-zero area (checked via pycocotools), so the segmentations themselves are intact.

However, all annotations currently have category_id = 1, even though the JSON header defines six categories (frame, text, face, body, balloon, onomatopoeia). It looks as if the masks were exported correctly but the category mapping was lost.

Note: in Discussion #4 we fixed a minor path typo in parse_rle_manga.py (json/ → jsons/). That bug doesn’t affect the label issue, but the corrected script may help others reproduce the checks.

Questions
1. Is there a version of the JSON where category_id = 1-6 is preserved?
2. If so, will it be published, or is there a recommended way to obtain it for research purposes?

Thanks again—looking forward to using the full multi-class annotations!

Best