--- pretty_name: Manga109 Segmentation license: other language: - ja task_categories: - image-segmentation - object-detection size_categories: - 10K **Manga109 images are not included.** Obtain Manga109 separately and follow > its terms. Every `images[].file_name` is relative to the Manga109 `images/` > directory. ## What changed in v2.0.0 This is a breaking supervision update intended for standard RF-DETR-style instance-segmentation training: - Text masks on 449 pages use the manually painted Zenodo Manga109 text-mask dataset as the highest-priority pixel source. - The remaining pages use [`mayocream/koharu-text-sam-ts-l`](https://huggingface.co/mayocream/koharu-text-sam-ts-l) to refine text/COO pixels inside authoritative human geometry. - Good existing text masks are unioned with clipped teacher ink. Filled box/polygon fallbacks are replaced when the teacher has sufficient support. - PP-DocLayoutV3 is used **only for bounding-box proposals**. For the 3,372 accepted train-only pseudo instances, the mask is always TextSeg ink clipped to the proposal; the stored box is tightened to the resulting mask. - 504 pages with materially incomplete positive labels were removed so their unlabeled text cannot become false-negative COCO background. This includes 102 `000.jpg` cover pages. - All 454,606 published annotations have `iscrowd: 0`. No custom dense head, ignore-region encoding, or synthetic negative-mask class is required. The previous release remains available at the immutable `v1.1.0` tag. ## Dataset summary The split remains book-disjoint. Filtering removes pages, not books. | Split | Books | Pages | Text | COO | Bubbles | Panels | All annotations | |---|---:|---:|---:|---:|---:|---:|---:| | Train | 87 | 8,128 | 129,608 | 45,165 | 102,088 | 81,638 | 358,499 | | Validation | 11 | 1,001 | 15,877 | 7,395 | 13,784 | 11,157 | 48,213 | | Test | 11 | 969 | 16,826 | 6,388 | 13,835 | 10,845 | 47,894 | | **Total** | **109** | **10,098** | **162,311** | **58,948** | **129,707** | **103,640** | **454,606** | The annotations contain 355,817 geometric containment relations. The three `review/*.jsonl` files contain sanitized per-page diagnostics for all 10,602 candidate pages, including the 504 excluded pages; they are not training annotations. ## Package layout ```text manga109-segmentation/ ├── annotations/ │ ├── train.coco.json │ ├── validation.coco.json │ └── test.coco.json ├── review/ │ ├── train.jsonl │ ├── validation.jsonl │ └── test.jsonl ├── build.json ├── checksums.sha256 └── package_manifest.json ``` Use the relative image paths with a separately obtained Manga109 release: ```python import json from pathlib import Path dataset_root = Path("manga109-segmentation") image_root = Path("Manga109_released_2026_05_21/images") with (dataset_root / "annotations/train.coco.json").open(encoding="utf-8") as f: coco = json.load(f) image_path = image_root / coco["images"][0]["file_name"] ``` Masks use compressed COCO RLE. `bbox` is COCO `[x, y, width, height]`, `area` is the mask-pixel count, and every annotation uses `iscrowd: 0`. ## Categories and relations | ID | Category | |---:|---| | 1 | `text` | | 2 | `onomatopoeia` | | 3 | `bubble` | | 4 | `panel` | The top-level `relations` array records `contained_by_bubble` and `contained_by_panel` geometry. Image and bubble reading-order fields are heuristic hints, not human reading-order ground truth. ## Mask provenance `attributes.quality_tier` gives the direct training provenance: - `gold_mangaseg`: retained MangaSegmentation bubble/panel mask. - `gold_zenodo_refined`: human-geometry instance refined with manually painted Zenodo text-mask pixels. - `silver_textseg_refined`: human-geometry instance refined with TextSeg. - `silver_pp_bbox_textseg_mask`: PP-DocLayoutV3 proposal whose pixels come from TextSeg; used only in train. Counts by split are recorded in `build.json`. Detailed page-level agreement, teacher coverage, proposal boxes, exclusion reasons, and visual-review flags are in `review/*.jsonl` without local filesystem paths. ## Page filtering and negative supervision A page is excluded when it has no target typography but at least 512 teacher foreground pixels, or when it has at least 10,000 teacher pixels and final recall below 0.20. The two tests may overlap. Seven retained pages have no typography instances and a near-empty teacher mask (at most 382 pixels). They provide safe implicit background supervision. There is deliberately no `negative` segmentation category: ordinary COCO background is the negative signal, while incomplete pages are omitted. ## Limitations - Most text/COO instance identities and envelopes are human-authored, but many final pixel masks are model-assisted. - PP-DocLayoutV3 can introduce class/proposal errors in the 3,372 train-only pseudo instances, even though TextSeg supplies their pixels. - Validation and test include teacher-refined pixels, so use independent human or Zenodo evaluation when measuring absolute mask quality. - Bubble/panel masks inherit MangaSegmentation ambiguity and source errors. - Relations and reading order are geometric heuristics. - Manga109 images and their usage rights are not distributed here. ## License and sources `license: other` is intentional because the package combines derived annotations from multiple sources. Users must follow every upstream license and attribution requirement. In particular: - [Manga109 and Manga109-v2026](https://manga109.github.io/manga109-project-website/en/) - [MangaSegmentation](https://huggingface.co/datasets/MS92/MangaSegmentation) - [COO: Comic Onomatopoeia Dataset](https://github.com/ku21fan/COO-Comic-Onomatopoeia) - [Zenodo Manga text-mask dataset](https://doi.org/10.5281/zenodo.4511796) (CC BY 4.0) - [`mayocream/koharu-text-sam-ts-l`](https://huggingface.co/mayocream/koharu-text-sam-ts-l) - [`PaddlePaddle/PP-DocLayoutV3_safetensors`](https://huggingface.co/PaddlePaddle/PP-DocLayoutV3_safetensors) The package does not grant access to or a license for Manga109 images. Never upload the Manga109 image files with this repository. Exact source revisions, policy thresholds, split counts, and quality-tier counts are in `build.json`; file integrity is recorded in `checksums.sha256`.