Source Datasets of REBU-Syn

• We collected labeled data from 16 publicly available real datasets to construct REBU-Syn. The details of these datasets are listed in the following table.

Data file name	Size	Link	License
OpenVINO	1.5M	https://storage.googleapis.com/openimages/web/index.html	Apache License 2.0
TextOCR	0.8M	https://textvqa.org/textocr/dataset/	CC BY 4.0
ICDAR2013	843	https://rrc.cvc.uab.es/?ch=2	Unknown
ICDAR2015	4,467	https://rrc.cvc.uab.es/?ch=4	CC BY 4.0
IIIT5K	2,000	https://cvit.iiit.ac.in/research/projects/cvit-projects/the-iiit-5k-word-dataset	MIT License
SVT	257	http://www.iapr-tc11.org/mediawiki/index.php/The_Street_View_Text_Dataset	Unknown
Total-Text	12,251	https://github.com/cs-chan/Total-Text-Dataset	BSD-3 license
CTW1500	3,170	https://github.com/Yuliang-Liu/Curve-Text-Detector	Unknown
Uber	127,850	https://s3-us-west-2.amazonaws.com/uber-common-public/ubertext/index.html	Unknown
RCTW17	10,245	https://rctw.vlrlab.net/dataset	Unknown
COCOv2.0	72,950	https://vision.cornell.edu/se3/coco-text-2/	CC BY 4.0
LSVT	8,164	https://rrc.cvc.uab.es/?ch=16	Unknown
MLT19	55,112	https://rrc.cvc.uab.es/?ch=15	CC BY-NC 4.0
ReCTS	26,040	https://rrc.cvc.uab.es/?ch=12	Unknown
ArT	31,966	https://rrc.cvc.uab.es/?ch=14	Unknown
Union14M_L_lmdb_format	3M	https://github.com/Mountchicken/Union14M/tree/main?tab=readme-ov-file#34-download	MIT License

• We collected labeled data from 4 publicly available synthetic datasets to construct REBU-Syn. The details of these datasets are listed in the following table.

Data file name	Size	Link	License
MJ	6M	https://www.robots.ox.ac.uk/~vgg/data/text/	Unknown
ST	9M	https://www.robots.ox.ac.uk/~vgg/data/scenetext/	Unknown
Curved SynthText	1.7M	https://github.com/Jyouhou/ICDAR2019-ArT-Recognition-Alchemy	Apache License 2.0
SynthAdd	1.2M	https://github.com/wangpengnorman/SAR-Strong-Baseline-for-Text-Recognition	Unknown

Datasets

Download the training dataset from the following links:

1. LMDB archives for MJ, ST, IIIT5k, SVT, SVTP, IC13, IC15, CUTE80, ArT, RCTW17, ReCTS, LSVT, MLT19, COCO-Text, and Uber-Text.
2. LMDB archives for TextOCR and OpenVINO.
3. LMDB archives for Union14M_L_lmdb_format.
4. CTW1500
5. Total-Text
6. SynthAdd
7. Curved SynthText

Then, organize the data as follows:

├── REBU-Syn
├── train
│   └── synth_and_real
│       ├── Curved_SynthText
│       │   ├── syntext1
│       │   └── syntext2
│       ├── SynthAdd
│       │   ├── data.mdb
│       │   └── lock.mdb
│       ├── Union14M_L_lmdb_format
│       │   ├── difficult
│       │   ├── hard
│       │   ├── hell
│       │   ├── medium
│       │   └── simple
│       ├── benchmark
│       │   ├── ICDAR2013
│       │   ├── ICDAR2015
│       │   ├── IIIT5K
│       │   └── SVT
│       ├── extra
│       │   ├── CTW1500
│       │   └── total_text
│       └── real_data
│       │   ├── ArT
│       │   ├── COCOv2.0
│       │   ├── LSVT
│       │   ├── MLT19
│       │   ├── OpenVINO
│       │   ├── RCTW17
│       │   ├── ReCTS
│       │   ├── TextOCR
│       │   └── Uber
│       └── mj_st
│           ├── data.mdb
│           └── lock.mdb
└── val
│   ├── CUTE80
│   ├── IC13_1015
│   ├── IC15_1811
│   ├── IIIT5k
│   ├── SVT
│   └── SVTP
├── test
│   ├── CUTE80
│   ├── IC13_1015
│   ├── IC13_857
│   ├── IC15_1811
│   ├── IIIT5k
│   ├── SVT
│   └── SVTP

Data Generation

We generated MJST+(60M) using TextRecognitionDataGenerator and SynthText. For specific generation methods, please refer to GenData.md

Acknowledgement

We sincerely thank all the constructors of the 20 datasets used in REBU-Syn.

• PARSeq: the dataset we built upon. Thanks for their wonderful work!
• Union14M: organizes a challenging STR training data. Don't forget to check this great open-source work if you don't know it before!