Rotated cascade R-CNN: A shape robust detector with coordinate regression

作者：

Highlights：

• We present an RCR-CNN, which can gradually regress the object in a two-stage manner and significantly improves the performance of our system.

• Our proposed method won first place in the ICPR 2018 Contest for Robust Reading for Multi-Type Web Images with a score of 0:796 and was our best single model in the ICPR 2018 Contest on Object Detection in Aerial Images (ODAI) with a 69:2% mean average precision (mAP), where we won first place. In addition, we also achieved the best results on the curved text detection dataset CTW1500, demonstrating the effectiveness and flexibility of our method.

摘要

•We present a novel LocSLPR method that can handle quadrangular/curved objects and well address the ambiguity problem of vertex order compared with direct regression. LocSLPR requires fewer parameters and achieves better results than segmentation-based methods.•We present an RCR-CNN, which can gradually regress the object in a two-stage manner and significantly improves the performance of our system.•Our proposed method won first place in the ICPR 2018 Contest for Robust Reading for Multi-Type Web Images with a score of 0:796 and was our best single model in the ICPR 2018 Contest on Object Detection in Aerial Images (ODAI) with a 69:2% mean average precision (mAP), where we won first place. In addition, we also achieved the best results on the curved text detection dataset CTW1500, demonstrating the effectiveness and flexibility of our method.

论文关键词：Object detection,Text detection,Aerial images,Curved text,Rotated cascade R-CNN

论文评审过程：Received 27 December 2018, Revised 30 June 2019, Accepted 10 July 2019, Available online 16 July 2019, Version of Record 19 July 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.106964