Authors: Pei Xu, Hongzhen Wang, Shan Huang, Shen Huang, Qi Ju
Description: This method is based on Mask RCNN. We use resnet152 as backbone and don't use any ensemble methods. We train and test the model in multi scales. We synthesized curved data to pretrain the model. MLT2017 and a small part of LSVT data are used in training.