Updated on May 30, 2019 (Important announcement on Total-Text vs. ArT dataset)
Updated on April 02, 2019 (Updated table ranking with default vs. our proposed DetEval)
Updated on March 31, 2019 (Faster version DetEval.py, support Python3. Thank you princewang1994.)
Updated on March 14, 2019 (Updated table ranking with evaluation protocol info.)
Updated on November 26, 2018 (Table ranking is included for reference.)
Updated on August 24, 2018 (Newly added annotation tool folder.)
Updated on May 15, 2018 (Added groundtruth in '.txt' format.)
Updated on May 14, 2018 (Added feature - 'Do not care' candidates filtering is now available in the latest python scripts.)
Updated on April 03, 2018 (Added pixel level groundtruth)
Updated on November 04, 2017 (Added text level groundtruth)
Released on October 27, 2017
TOTAL-TEXT is a word-level based English curve text dataset. If you are interested in text-line based dataset with both English and Chinese instances, we highly recommend you to refer SCUT-CTW1500. In addition, a Robust Reading Challenge on Arbitrary-Shaped Text (ArT), which is extended from Total-Text and SCUT-CTW1500, was held at ICDAR2019 to stimulate more innovative ideas on the arbitrary-shaped text reading task. Congratulations to all winners and challengers.
Total-Text and SCUT-CTW1500 are now part of the training set of the largest curved text dataset - ArT (Arbitrary-Shaped Text dataset). In order to retain the validity of future benchmarking on Total-Text datasets, the test-set images of Total-Text should be removed (with the corresponding ID provided here (https://github.com/cs-chan/Total-Text-Dataset/blob/master/Total_Text_ID_vs_ArT_ID.list)) from the ArT dataset shall one intend to leverage the extra training data from the ArT dataset. We count on the trust of the research community to perform such removal operation to attain the fairness of the benchmarking.
- The results from recent papers on Total-Text dataset are listed below where P=Precision, R=Recall & F=F-score.
- If your result is missing or incorrect, please do not hesisate to contact us.
- *Pascal VOC IoU metric; **Polygon Regression
| Method | Reported on paper |
DetEval (tp=0.4, tr=0.8) (Default) |
DetEval (tp=0.6, tr=0.7) (New Proposal) |
Published at | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| P | R | F | P | R | F | P | R | F | ||
| TextCohesion [paper] | 88.1 | 81.4 | 84.6 | - | - | - | - | - | - | arXiv:1904 |
| CRAFT [paper] | 87.6 | 79.9 | 83.6 | - | - | - | - | - | - | CVPR2019 |
| LOMO MS [paper] | 87.6 | 79.3 | 83.3 | - | - | - | - | - | - | CVPR2019 |
| ICG [paper] | 82.1 | 80.9 | 81.5 | - | - | - | - | - | - | PR2019 |
| FTSN [paper] | *84.7 | *78.0 | *81.3 | - | - | - | - | - | - | ICPR2018 |
| PSENet-1s [paper] | 84.02 | 77.96 | 80.87 | - | - | - | - | - | - | CVPR2019 |
| 1TextField [paper] | 81.2 | 79.9 | 80.6 | 76.1 | 75.1 | 75.6 | 83.0 | 82.0 | 82.5 | TIP2019 |
| CSE [paper] | 81.4 (**80.9) |
79.7 (**80.3) |
80.2 (**80.6) |
- | - | - | - | - | - | CVPR2019 |
| MSR [paper] | 85.2 | 73.0 | 78.6 | 82.7 | 68.3 | 74.9 | 81.4 | 72.5 | 76.7 | arXiv:1901 |
| ATTR [paper] | 80.9 | 76.2 | 78.5 | - | - | - | - | - | - | CVPR2019 |
| TextSnake [paper] | 82.7 | 74.5 | 78.4 | - | - | - | - | - | - | ECCV2018 |
| 1CTD [paper] | 74.0 | 71.0 | 73.0 | 60.7 | 58.8 | 59.8 | 76.5 | 73.8 | 75.2 | PR2019 |
| TextNet [paper] | 68.2 | 59.5 | 63.5 | - | - | - | - | - | - | ACCV2018 |
| 2Mask TextSpotter [paper] | 69.0 | 55.0 | 61.3 | 68.9 | 62.5 | 65.5 | 82.5 | 75.2 | 78.6 | ECCV2018 |
| CENet [paper] | 59.9 | 54.4 | 57.0 | - | - | - | - | - | - | ACCV2018 |
| Textboxes [paper] | 62.1 | 45.5 | 52.5 | - | - | - | - | - | - | AAAI2017 |
| EAST [paper] | 50.0 | 36.2 | 42.0 | - | - | - | - | - | - | CVPR2017 |
| Baseline [paper] | 33.0 | 40.0 | 36.0 | - | - | - | - | - | - | ICDAR2017 |
| SegLink [paper] | 30.3 | 23.8 | 26.7 | - | - | - | - | - | - | CVPR2017 |
Note:
1For the results of TextField and CTD, the improved versions of their original paper were used, and this explains why the performance is better.
2For Mask-TextSpotter, the relatively poor performance reported in their paper was due to a bug in the input reading module (which was fixed recently). The authors were informed about this issue.
End-to-end Recognition
(None refers to recognition without any lexicon; Full lexicon contains all words in test set.)
| Method | None (%) | Full (%) | Published at |
|---|---|---|---|
| TextNet [paper] | 54.0 | - | ACCV2018 |
| Mask TextSpotter [paper] | 52.9 | 71.8 | ECCV2018 |
| Textboxes [paper] | 36.3 | 48.9 | AAAI2017 |
In order to facilitate a new text detection research, we introduce Total-Text dataset (ICDAR-17 paper) (presentation slides), which is more comprehensive than the existing text datasets. The Total-Text consists of 1555 images with more than 3 different text orientations: Horizontal, Multi-Oriented, and Curved, one of a kind.
If you find this dataset useful for your research, please cite
@article{CK2019,
author = {Chee Kheng Ch’ng and
Chee Seng Chan and
Chenglin Liu},
title = {Total-Text: Towards Orientation Robustness in Scene Text Detection},
journal = {IJDAR},
year = {2019},
}
@inproceedings{CK2017,
author = {Chee Kheng Ch’ng and
Chee Seng Chan},
title = {Total-Text: A Comprehensive Dataset for Scene Text Detection and Recognition},
booktitle = {14th IAPR International Conference on Document Analysis and Recognition {ICDAR}},
pages = {935--942},
year = {2017},
doi = {10.1109/ICDAR.2017.157},
}
Suggestions and opinions of this dataset (both positive and negative) are greatly welcome. Please contact the authors by sending email to
chngcheekheng at gmail.com or cs.chan at um.edu.my.
The project is open source under BSD-3 license (see the LICENSE file). Codes can be used freely only for academic purpose.
©2017-2019 Center of Image and Signal Processing, Faculty of Computer Science and Information Technology, University of Malaya.