Authors: Xiao Duan(Fudan University)

Description: Due to the texts have arbitrary shapes in the contest, so I choose the method based on segmentation. I use mask-rcnn as the benchmark model. Based on this, cascade bounding box detection (regressing bounding box for several times) and deformable convolution are added to this model. Moreover, I use soft_nms to filter the model's result. However, the result is not very satisfying, so I use some postprocessing operations, such as mask_nms to suppress the redundant mask and filtering the results that confidence below 0.5. By those steps, I get the final results.