Total-Text (Total-Text)

Research Tasks

Scene Text Recognition in Natural Scene Images

2018-07-26 (v. 1)

Contact author

Chee Kheng Ch'ng

University of Malaya


This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 Unported License.


Given an image patch with text information, a scene text recognition algorithm is expected to infer the text content in it. Being able to retrieve text information (which exists abundantly in our every day scene), has numerous real-life applications: text-aided autonomous driving system, mobile language translation apps, multimedia retrieval system, etc. Arbitrary-oriented text recognition is much more challenging than horizontal text recognition[1-2], yet it is part of the attributes in text in the wild. Being able to read text regardless of its orientation brings the robustness of scene text understanding algorithm to a new level.

[1] Lyu, Pengyuan, et al. "Mask TextSpotter: An End-to-End Trainable Neural Network for Spotting Text with Arbitrary Shapes." arXiv preprint arXiv:1807.02242 (2018).

[2] Shi, Baoguang, et al. "Robust scene text recognition with automatic rectification." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.



The evaluation protocol for this task follows the proposal of [3]. A word is considered correctly recognized if at least 80% and all letters are recognized correctly (case-sensitive). The expected output is a string of characters, in word level.

[3]S. M. Lucas, A. Panaretos, L. Sosa, A. Tang, S. Wong, and R. Young. ICDAR 2003 robust reading competitions. ICDAR, 2003.


No comments on this dataset yet.

Add your comment

In order to comment on a dataset you need to be logged on
Register Now!


In order to rate this dataset you need to be logged on
Register Now!