CTW1500 / README.md
HB16888's picture
Upload README.md with huggingface_hub
58fb765 verified
|
Raw
History Blame Contribute Delete
1.5 kB
metadata
license: other
license_name: ctw1500-research
license_link: https://github.com/open-mmlab/mmocr/blob/main/dataset_zoo/ctw1500/metafile.yml
tags:
  - text-detection
  - ocr
  - scene-text
  - mmocr

CTW1500 (MMOCR format)

CTW1500 is a curve scene text detection benchmark with 1,000 training and 500 test images. The official download links (Box / CloudStor) are currently unreliable, so this repository re-hosts the dataset already prepared in MMOCR format for convenience.

Contents

ctw1500_mmocr.zip unpacks to:

ctw1500/
├── textdet_imgs/
│   ├── train/          # 1000 images
│   └── test/           # 500 images
├── textdet_train.json  # MMOCR textdet training annotations
└── textdet_test.json   # MMOCR textdet test annotations

Usage

hf download HB16888/CTW1500 --repo-type dataset --local-dir ./ctw1500_dl
unzip ctw1500_dl/ctw1500_mmocr.zip -d data/

Then point MMOCR configs at data/ctw1500 (the default data_root='data/ctw1500' in the textdet configs works out of the box).

The same archive is mirrored on ModelScope.

License

The dataset is provided for academic research purposes, following the original CTW1500 release terms. The preparation pipeline is the one shipped with MMOCR (tools/dataset_converters/prepare_dataset.py ctw1500 --task textdet).