jawildtext_cropped

Per-polygon scene-text-recognition crops derived from llm-jp/jawildtext.

Each row of the source dataset carries a polygons column with quadrilateral text regions and their transcriptions. For every polygon we perspective-warp the source image onto the rectified bounding rectangle, yielding a tight, horizontally-aligned crop suitable for training/evaluating Japanese scene-text recognition models.

Stats

• Samples: 108 403
• Shards: 22 (shard-{00000..00021}.tar)
• Format: webdataset
• Source subset: default

Layout

Each shard is a tar with two files per sample:

data/shard-00000.tar
    jawildtext-00000-00000000.jpg   # rectified RGB crop, JPEG q=95
    jawildtext-00000-00000000.txt   # UTF-8 transcription
    ...

Loading

import webdataset as wds

url = "https://huggingface.co/datasets/nagohachi/jawildtext_cropped/resolve/main/data/shard-{00000..00021}.tar"
ds = wds.WebDataset(url).decode("pil").to_tuple("jpg", "txt")
for image, text in ds:
    ...

Preprocessing

1. Stream llm-jp/jawildtext (default, train split).
2. For each polygon: reorder its 4 vertices to (TL, TR, BR, BL) via sum/diff sorting, compute the target rectangle size from edge lengths, then apply a perspective transform (PIL Image.Transform.PERSPECTIVE, BICUBIC).
3. Drop crops with empty text or with either side < 8 px.
4. Pack 5 000 samples per shard.

Skipped from the source: 2 583 empty-text polygons, 1 130 polygons below the 8 px size threshold. No warp failures.

License

Apache 2.0, matching the upstream llm-jp/jawildtext license.