In this paper, we introduce a very large Chinese text dataset in the wild. While optical character recognition (OCR) in document images is well studied and many commercial tools are available, the detection and recognition of text in natural images is still a challenging problem, especially for some more complicated character sets such as Chinese text. Lack of training data has always been a problem, especially for deep learning methods which require massive training data. In this paper, we provide details of a newly created dataset of Chinese text with about 1 million Chinese characters from 3850 unique ones annotated by experts in over 30000 street view images. This is a challenging dataset with good diversity containing planar text, raised text, text under poor illumination, distant text, partially occluded text, etc. Besides the dataset, we give baseline results using state-of-the-art methods for three tasks: character recognition (top-1 accuracy of 80.5%), character detection (AP of 70.9%), and text line detection (AED of 22.1). The dataset, source code, and trained models are publicly available.

1. 32,285 high resolution images

2. 1,018,402 character instances

3. 3,850 character categories

4. 6 kinds of attributes

evaluation Server

• The evaluation server is available on CodaLab.

• You should submit a .zip file, which contains one .jsonl file in the top-level directory. Submission formats and evaluation metrics for classification task and detection task are described in tutorial part-2 and part-3, respectively.

• Sample submissions can be downloaded from "public submissions" of corresponding competition on CodaLab. You may need to login to CodaLab before downloading.

• Detailed results are provided in the "view detailed results" link for each submission.

Contact

If you have any questions about the dataset or code, please contact Tai-Ling Yuan (yuantailing[at]gmail.com).

Bibtex:

@article{yuan2019ctw,
  author  = {Tai{-}Ling Yuan and Zhe Zhu and Kun Xu and Cheng{-}Jun Li and Tai{-}Jiang Mu and Shi{-}Min Hu},
  title   = {A Large Chinese Text Dataset in the Wild},
  journal = {Journal of Computer Science and Technology},
  volume  = {34},
  number  = {3},
  pages   = {509--521},
  year    = {2019},
}

Change Log

• 06/17/2019 (GMT+8): replace the paper with A Large Chinese Text Dataset in the Wild

• 07/04/2018 (GMT+8): dataset moved to OneDrive

• 03/17/2018 (GMT+8): evaluation server available

• 03/15/2018 (GMT+8): dataset released on WeiYun and Google Drive

• 02/28/2018 (GMT+8): website comes online

Terms of Use

• The public annotations and trained models belong to the CSCG Group and are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

• The images belong to Tencent ltd. and are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

• Most of the baseline code belongs to Tai-Ling Yuan and is licensed under the MIT License.