Authors: Yi Yu, Haiyang Guo, Xiaobing Wang, Yingying Jiang

Description: We use Mask R-CNN for text detection in our submission, which extends Faster R-CNN by adding a branch for predicting an object mask in parallel with the existing branch for bounding box recognition. Mask R-CNN is simple to train and adds only a small overhead to Faster R-CNN. And Mask R-CNN is the state-of-the-art object detection framework now. Therefore, it is used here for text detection task. We use the Mask R-CNN in https://github.com/facebookresearch/Detectron and the backbone network is ResNext 101. Meanwhile, the object is text here and the number of classification classes is 2. Besides, polygon based NMS is used for post-processing to remove overlapped text regions.

Authors: Yaowu Wei, Shangchao Su, Tairu Qiu, Xunyan Wang, Shaokang Lin, Zili Yi, Lei Deng, Mulin Xu, Jianqi Ma, Bin Li, Xiangyang Xue

Description: We propose a text segmentation algorithm based on Cascaded Mask-RCNN with deformable convolution and DenseASPP. And furthermore, we do some changes.

Authors: Pei Xu, Shan Huang. (Tencent)

Description: This is an instance segmentation based method. We segment the text polygons from region proposals. In the training process, We first train on images synthesized by the method of VGG-synthtext, and then fine tune on the LSVT training images. In the post-processing, we adopt NMS and filter out those polygons with low scores.