Robust ReadingCompetition
Challenges

method: CLOVA-AI / PAPAGO2018-07-03

Authors: Youngmin Baek, Bado Lee, Hwalsuk Lee

Description: Character-level text detection based on weakly-supervised learning. Multi-scale experiment result. CLOVA-AI team, Naver Corp.  (Paper in preparation)

method: ATL Cangjie OCR2018-03-12

Authors: Yang Fan, Liu Yang, Guo Shan, Alibaba Turing Lab

Description: An end to end text recognition framework (both text detection and recognition) was used. In detection part, we used modified SSD and improved NMS to detect both text confidence and its quadrilateral location. In recognition part, we used CNN+CTC. Finally, we refined the detection results using both the information of detection and recognition.

method: FOTS_v22018-01-22

Authors: Xuebo Liu, Ding Liang, Junjie Yan

Description: FOTS_v2 improves the performance on long text, such as Chinese.

Ranking Table

Description Paper Source Code
DateMethodAverage PrecisionPrecisionRecallHmean
2018-07-03CLOVA-AI / PAPAGO54.64%82.37%66.64%73.68%
2018-03-12ATL Cangjie OCR64.30%78.88%68.84%73.52%
2018-01-22FOTS_v259.93%83.06%65.61%73.31%
2017-11-09EAST++54.94%80.42%66.61%72.86%
2018-05-18PSENet_NJU_ImagineLab (single-scale)52.51%77.01%68.40%72.45%
2018-01-22FOTS56.95%81.86%62.30%70.75%
2017-06-28SCUT_DLVClab150.34%80.28%54.54%64.96%
2017-06-30Sensetime OCR61.24%56.93%69.43%62.56%
2017-06-29SARI_FDU_RRPN_v150.33%71.17%55.50%62.37%
2017-06-28SARI_FDU_RRPN_v048.76%67.07%55.37%60.66%
2017-06-30TH-DL30.88%67.75%34.78%45.97%
2017-06-30linkage-ER-Flow15.47%44.48%25.59%32.49%

Ranking Graphic

Ranking Graphic