Authors: Xiangcheng Du, Xingjiao Wu, Shufan Wu, Shuchen Kong, Hao Ye, Yingbin Zheng, Liang He

Description: We proposed a detection method based on segmentation. We use feature pyramid networks(fpn) as the backbone, we consider classification and location as the output. Specially we predict each pixel with three types: positive, negative and ambiguous. We use an offset vector to cluster the pixels belong to the same text instance.