Kotomiya07/kuzushiji-dataset-yolo

Kuzushiji Dataset (YOLO format)

This dataset contains page images from historical Japanese documents (Kotenseki) with character-level bounding box annotations for Kuzushiji (cursive Japanese) recognition.

Dataset Structure

{
    "image": Image(),                    # Page image
    "image_id": str,                     # Image ID (e.g., 100241706_00004_2)
    "book_id": str,                      # Book ID (e.g., 100241706)
    "width": int,                        # Image width in pixels
    "height": int,                       # Image height in pixels
    "objects": {
        "bbox": List[List[float]],       # Bounding boxes ([x_center, y_center, width, height] (normalized 0-1))
        "category": List[str],           # Unicode strings (e.g., U+3042)
        "category_id": List[int],        # Category IDs
        "char": List[str],               # Actual characters (e.g., あ)
    },
    "columns": {
        "bbox": List[List[float]],       # Column boxes
        "column_id": List[str],          # Column IDs (e.g., COL0001)
        "char_ids": List[List[str]],     # Member Char IDs
        "segment_id": List[str],         # Parent Segment IDs if available
    },
    "segments": {
        "bbox": List[List[float]],       # Segment boxes
        "segment_id": List[str],         # Segment IDs (e.g., SEG0001)
        "column_ids": List[List[str]],   # Member Column IDs
    }
}

Bounding Box Formats

Format	Description	Normalized
COCO	[x_min, y_min, width, height]	No (pixels)
YOLO	[x_center, y_center, width, height]	Yes (0-1)

Usage

from datasets import load_dataset
import json

dataset = load_dataset("Kotomiya07/kuzushiji-dataset-yolo")

# Access first example
example = dataset["train"][0]
print(f"Image ID: {example['image_id']}")
print(f"Number of characters: {len(example['objects']['bbox'])}")

# Load label mappings
from huggingface_hub import hf_hub_download

label2id_path = hf_hub_download(
    repo_id="Kotomiya07/kuzushiji-dataset-yolo",
    filename="label2id.json",
    repo_type="dataset"
)
with open(label2id_path) as f:
    label2id = json.load(f)

print(f"Number of categories: {len(label2id)}")

Label Mappings

This dataset includes the following mapping files:

• label2id.json: Unicode string (e.g., "U+3042") to category ID mapping
• id2label.json: Category ID to Unicode string mapping

License

This dataset is licensed under CC BY-SA 4.0.

Data Source

This dataset is derived from the 日本古典籍くずし字データセット (Japanese Historical Character Dataset).

• Provider: National Institute of Japanese Literature and other institutions (国文学研究資料館ほか所蔵)
• Processing: ROIS-DS Center for Open Data in the Humanities (CODH)
• DOI: 10.20676/00000340
• Website: https://codh.rois.ac.jp/char-shape/

Citation

If you use this dataset, please cite:

『日本古典籍くずし字データセット』（国文研ほか所蔵／CODH加工）doi:10.20676/00000340

English:

"Japanese Historical Character Dataset" (Owned by NIJL and others, Processed by CODH) doi:10.20676/00000340

Acknowledgments

Data provided by: ROIS-DS Center for Open Data in the Humanities (人文学オープンデータ共同利用センター)