Robust ReadingCompetition
Challenges

Authors: Tencent-DPPR Team (Chunchao Guo, Weichen Zhang, Yi Li, Hui Song, Ming Liu, Hongfa Wang, Lei Xiao) & USTB-PRIR (Chun Yang, Zejun Li, Jianwei Wu, Jiebo Hou, Chang Liu, Longhuang Wu, Xu-Cheng Yin)

Description: Tencent-DPPR (Data Platform Precision Recommendation) Team. They detect text regions using improved Rotation Region Proposal Networks. After that they extract features from text lines and employ multiple LSTM-based models to generate different results for each image. Finally, they select the one with the maximum probability among all candidate results. Please refer to the paper entiled "AdaDNNs: Adaptive Ensemble of Deep Neural Networks for Scene Text Recognition", arXiv, 2017.

method: Foo & Bar2017-06-30

Authors: Zheqi He, Yongtao Wang, Xiang Bai

Description: The method used quadrangle regression network for text detection, and then use homography to transform quadrangle regions to rectangles and finally CRNN (https://github.com/bgshih/crnn) for recognition.

method: WPS2017-06-30

Authors: Hin Lee

Description: A word spotting system that combines a text proposal extractor and a attention-based text recognizer.

Ranking Table

Description Paper Source Code
DateMethodAverage Precision
2017-07-01Tencent-DPPR Team & USTB-PRIR43.58%
2017-06-30Foo & Bar27.01%
2017-06-30WPS18.82%
2017-10-06SSD + CRNN (sravya)0.86%
2017-06-30CNN-LSTM based text recognition0.73%

Ranking Graphic