ShiniShiho/koharu-layout-rfdetr-seg-2xl-1152-onnx
Image Segmentation • Updated • 1
Datasets:
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
split: string
image: string
teacher_source: string
width: int64
height: int64
existing_typography_instances: int64
pp_pseudo_instances: int64
pp_boxes: list<item: struct<proposal_bbox: list<item: int64>, tight_mask_bbox: list<item: int64>, score: doubl (... 23 chars omitted)
child 0, item: struct<proposal_bbox: list<item: int64>, tight_mask_bbox: list<item: int64>, score: double, mask_pix (... 11 chars omitted)
child 0, proposal_bbox: list<item: int64>
child 0, item: int64
child 1, tight_mask_bbox: list<item: int64>
child 0, item: int64
child 2, score: double
child 3, mask_pixels: int64
refinement_policies: struct<replace_with_zenodo: int64, retain_existing_teacher_too_sparse: int64, union_existing_with_te (... 51 chars omitted)
child 0, replace_with_zenodo: int64
child 1, retain_existing_teacher_too_sparse: int64
child 2, union_existing_with_textseg: int64
child 3, replace_fallback_with_textseg: int64
old_vs_new: struct<reference_pixels: int64, prediction_pixels: int64, intersection_pixels: int64, union_pixels: (... 54 chars omitted)
child 0, reference_pixels: int64
child 1, prediction_pixels: int64
child 2, intersection_pixels: int64
child 3, union_pixels: int64
child 4, iou: double
child 5, precision: double
child 6, recall: double
teacher_vs_old: struct<reference_pixels: int64, prediction_pixels: int64, intersection_pixels: int64, union_pixels: (... 54 chars omitted)
child 0, reference_pixels: int64
...
d: timestamp[s]
child 4, annotation_policy: string
child 5, reading_order_policy: string
licenses: list<item: struct<id: int64, name: string, url: string>>
child 0, item: struct<id: int64, name: string, url: string>
child 0, id: int64
child 1, name: string
child 2, url: string
relations: list<item: struct<type: string, from: int64, to: int64, score: double>>
child 0, item: struct<type: string, from: int64, to: int64, score: double>
child 0, type: string
child 1, from: int64
child 2, to: int64
child 3, score: double
categories: list<item: struct<id: int64, name: string, supercategory: string>>
child 0, item: struct<id: int64, name: string, supercategory: string>
child 0, id: int64
child 1, name: string
child 2, supercategory: string
images: list<item: struct<id: int64, width: int64, height: int64, file_name: string, license: int64, book: s (... 146 chars omitted)
child 0, item: struct<id: int64, width: int64, height: int64, file_name: string, license: int64, book: string, page (... 134 chars omitted)
child 0, id: int64
child 1, width: int64
child 2, height: int64
child 3, file_name: string
child 4, license: int64
child 5, book: string
child 6, page_index: int64
child 7, source_split: string
child 8, dataset_split: string
child 9, text_reading_order_hint: list<item: int64>
child 0, item: int64
child 10, reading_order_quality: string
to
{'info': {'description': Value('string'), 'version': Value('string'), 'year': Value('int64'), 'date_created': Value('timestamp[s]'), 'annotation_policy': Value('string'), 'reading_order_policy': Value('string')}, 'licenses': List({'id': Value('int64'), 'name': Value('string'), 'url': Value('string')}), 'categories': List({'id': Value('int64'), 'name': Value('string'), 'supercategory': Value('string')}), 'images': List({'id': Value('int64'), 'width': Value('int64'), 'height': Value('int64'), 'file_name': Value('string'), 'license': Value('int64'), 'book': Value('string'), 'page_index': Value('int64'), 'source_split': Value('string'), 'dataset_split': Value('string'), 'text_reading_order_hint': List(Value('int64')), 'reading_order_quality': Value('string')}), 'annotations': List(Json(decode=True)), 'relations': List({'type': Value('string'), 'from': Value('int64'), 'to': Value('int64'), 'score': Value('float64')})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
split: string
image: string
teacher_source: string
width: int64
height: int64
existing_typography_instances: int64
pp_pseudo_instances: int64
pp_boxes: list<item: struct<proposal_bbox: list<item: int64>, tight_mask_bbox: list<item: int64>, score: doubl (... 23 chars omitted)
child 0, item: struct<proposal_bbox: list<item: int64>, tight_mask_bbox: list<item: int64>, score: double, mask_pix (... 11 chars omitted)
child 0, proposal_bbox: list<item: int64>
child 0, item: int64
child 1, tight_mask_bbox: list<item: int64>
child 0, item: int64
child 2, score: double
child 3, mask_pixels: int64
refinement_policies: struct<replace_with_zenodo: int64, retain_existing_teacher_too_sparse: int64, union_existing_with_te (... 51 chars omitted)
child 0, replace_with_zenodo: int64
child 1, retain_existing_teacher_too_sparse: int64
child 2, union_existing_with_textseg: int64
child 3, replace_fallback_with_textseg: int64
old_vs_new: struct<reference_pixels: int64, prediction_pixels: int64, intersection_pixels: int64, union_pixels: (... 54 chars omitted)
child 0, reference_pixels: int64
child 1, prediction_pixels: int64
child 2, intersection_pixels: int64
child 3, union_pixels: int64
child 4, iou: double
child 5, precision: double
child 6, recall: double
teacher_vs_old: struct<reference_pixels: int64, prediction_pixels: int64, intersection_pixels: int64, union_pixels: (... 54 chars omitted)
child 0, reference_pixels: int64
...
d: timestamp[s]
child 4, annotation_policy: string
child 5, reading_order_policy: string
licenses: list<item: struct<id: int64, name: string, url: string>>
child 0, item: struct<id: int64, name: string, url: string>
child 0, id: int64
child 1, name: string
child 2, url: string
relations: list<item: struct<type: string, from: int64, to: int64, score: double>>
child 0, item: struct<type: string, from: int64, to: int64, score: double>
child 0, type: string
child 1, from: int64
child 2, to: int64
child 3, score: double
categories: list<item: struct<id: int64, name: string, supercategory: string>>
child 0, item: struct<id: int64, name: string, supercategory: string>
child 0, id: int64
child 1, name: string
child 2, supercategory: string
images: list<item: struct<id: int64, width: int64, height: int64, file_name: string, license: int64, book: s (... 146 chars omitted)
child 0, item: struct<id: int64, width: int64, height: int64, file_name: string, license: int64, book: string, page (... 134 chars omitted)
child 0, id: int64
child 1, width: int64
child 2, height: int64
child 3, file_name: string
child 4, license: int64
child 5, book: string
child 6, page_index: int64
child 7, source_split: string
child 8, dataset_split: string
child 9, text_reading_order_hint: list<item: int64>
child 0, item: int64
child 10, reading_order_quality: string
to
{'info': {'description': Value('string'), 'version': Value('string'), 'year': Value('int64'), 'date_created': Value('timestamp[s]'), 'annotation_policy': Value('string'), 'reading_order_policy': Value('string')}, 'licenses': List({'id': Value('int64'), 'name': Value('string'), 'url': Value('string')}), 'categories': List({'id': Value('int64'), 'name': Value('string'), 'supercategory': Value('string')}), 'images': List({'id': Value('int64'), 'width': Value('int64'), 'height': Value('int64'), 'file_name': Value('string'), 'license': Value('int64'), 'book': Value('string'), 'page_index': Value('int64'), 'source_split': Value('string'), 'dataset_split': Value('string'), 'text_reading_order_hint': List(Value('int64')), 'reading_order_quality': Value('string')}), 'annotations': List(Json(decode=True)), 'relations': List({'type': Value('string'), 'from': Value('int64'), 'to': Value('int64'), 'score': Value('float64')})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Manga109 Segmentation
Manga109 Segmentation is an annotation-only dataset for manga layout and
instance segmentation. The current release is v3.0.0. It contains COCO RLE
masks for text, onomatopoeia, bubble, and panel, plus containment
relations, Japanese transcriptions, and the human polygons behind every
onomatopoeia.
Manga109 images are not included. Obtain Manga109 separately and follow its terms. Every
images[].file_nameis relative to the Manga109images/directory.
What changed in v3.0.0
v3.0.0 refines v2.0.0. Pages, splits, annotation ids, masks, bubbles, panels, and relations carry over, except where listed here:
- Human boxes. A text box is its Manga109 v2026 box (158,939 instances), and an onomatopoeia box spans its COO polygons (58,938). v2.0.0 tightened both to the mask, so a stroke the mask missed also shrank the box. A mask reaching outside its human box or polygons is clipped to it (1,060 instances, 2,715 pixels).
- No filled rectangles. A text or onomatopoeia mask covering at least 95%
of its own box is a filled fallback, not ink. 187 are replaced with the
MangaSegmentation or Zenodo ink for the same instance. The 61 with no ink
behind them keep their box and class, with an empty mask and
has_mask: false. - No page numbers. 1,930 of v2.0.0's 3,372 PP-DocLayoutV3 pseudo-instances were page numbers, which Manga109 leaves unlabelled on every other page. They are dropped. The 1,442 kept are editorial notes, titles, and credits.
- No orphan onomatopoeia. 10 onomatopoeia fragments (1–68 pixels) with no COO polygon behind them are dropped.
- New attributes. Onomatopoeia gain
human_polygon,rotated_box, andtranscription; text gainsmanga109_idandtranscription. See Boxes and Masks.
v2.0.0 remains available at the immutable v2.0.0 tag.
Dataset summary
The split is book-disjoint and unchanged from v2.0.0.
| Split | Books | Pages | Text | Onomatopoeia | Bubbles | Panels | All annotations |
|---|---|---|---|---|---|---|---|
| Train | 87 | 8,128 | 127,678 | 45,158 | 102,088 | 81,638 | 356,562 |
| Validation | 11 | 1,001 | 15,877 | 7,392 | 13,784 | 11,157 | 48,210 |
| Test | 11 | 969 | 16,826 | 6,388 | 13,835 | 10,845 | 47,894 |
| Total | 109 | 10,098 | 160,381 | 58,938 | 129,707 | 103,640 | 452,666 |
The annotations contain 355,807 containment relations.
Package layout
manga109-segmentation/
├── annotations/
│ ├── train.coco.json
│ ├── validation.coco.json
│ └── test.coco.json
├── review/
│ ├── train.jsonl
│ ├── train.v3_changes.jsonl
│ ├── validation.jsonl
│ ├── validation.v3_changes.jsonl
│ ├── test.jsonl
│ └── test.v3_changes.jsonl
├── build.json
├── checksums.sha256
└── package_manifest.json
review/{split}.jsonl are v2.0.0's per-page diagnostics, unchanged.
review/{split}.v3_changes.jsonl record each instance v3.0.0 dropped or whose
rectangle mask it replaced or removed. Neither is a training annotation.
Use the relative image paths with a separately obtained Manga109 release:
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"]
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.
Boxes
bbox is COCO [x, y, width, height]. attributes.box_source says where it
comes from:
box_source |
Category | Box |
|---|---|---|
manga109_v2026 |
text | The Manga109 v2026 box, clipped to the page; manga109_id names it |
coo |
onomatopoeia | The extent of its COO polygons, stored in human_polygon |
pp_doclayoutv3 |
text | v2.0.0's box, tightened to its TextSeg mask; train only |
| absent | bubble, panel | v2.0.0's box, from MangaSegmentation |
Human boxes surround the text with the annotator's margin, so a text or
onomatopoeia box is usually larger than its mask. An onomatopoeia linked
across several COO polygons has one polygon per piece. Its rotated_box is
the minimum-area rectangle around all of them, as
{"center", "size", "angle", "convention"} from OpenCV's minAreaRect.
Masks
Masks use compressed COCO RLE, area is the mask's pixel count, and every
annotation uses iscrowd: 0. A text or onomatopoeia mask always lies inside
its human box or polygons.
attributes.has_mask is false on 61 instances: 44 text and 17 onomatopoeia,
of which 57 are in train and 4 in test. Their mask is empty and their area
is the box area. Skip their mask loss in training, or drop them, since an
ordinary COCO loader would teach the model that these instances have no
pixels. Drop them from the ground truth when evaluating masks.
attributes.quality_tier gives v2.0.0's mask provenance:
| Tier | Text | Onomatopoeia | Bubble | Panel |
|---|---|---|---|---|
gold_mangaseg: MangaSegmentation mask |
129,707 | 103,640 | ||
gold_zenodo_refined: manually painted Zenodo pixels |
6,095 | 2,239 | ||
silver_textseg_refined: refined by TextSeg |
152,844 | 56,699 | ||
silver_pp_bbox_textseg_mask: PP-DocLayoutV3 box, TextSeg pixels; train only |
1,442 |
Where v3.0.0 replaced a rectangle, attributes.mask_replaced_by names the new
pixels' source: mangaseg (186) or zenodo (1).
Page filtering
v3.0.0 keeps v2.0.0's pages. v2.0.0 removed 504 pages whose unlabelled text
would have become false-negative background, including 102 000.jpg covers.
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. There is deliberately no negative category: ordinary COCO
background is the negative signal.
Limitations
- Most text and onomatopoeia identities and boxes are human-authored, but many final pixel masks are model-assisted.
- Manga109 labels neither editorial notes nor page numbers, and COO misses some small handwritten onomatopoeia. Outside the train-only pseudo-instances these stay background.
- PP-DocLayoutV3 can introduce class and box errors in the 1,442 train-only pseudo-instances.
- Validation and test include teacher-refined pixels, so use independent human or Zenodo evaluation when measuring absolute mask quality.
- Bubble and panel masks inherit MangaSegmentation's 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
- MangaSegmentation
- COO: Comic Onomatopoeia Dataset
- Zenodo Manga text-mask dataset (CC BY 4.0)
mayocream/koharu-text-sam-ts-lPaddlePaddle/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 checksums
and revisions, policy thresholds, and per-split change counts are in
build.json; file integrity is recorded in checksums.sha256 and
package_manifest.json.
- Downloads last month
- 262
