Authors: Xingran Zhao, Jing Xian, Yadong Li, Hongbin Wang

Affiliation: AntGroup

Email: zhaoxingran.zxr@antgroup.com;xianjing.xj@antgroup;liyadong.lyd@antgroup.com;hongbin.whb@antgroup.com

Description: For word and line detection, we firstly crop patches from images for catching local mask results. Second, we also get global mask results by using full images as the input. Thirdly, we merge global and local results by using NMS postprocess procedure. For paragraph detection, we only use full images as input and get global mask results. All detectors are CBNetV2[1] with HTC[2]. For hierarchical text detection, we use IOS(intersection-of-sets) as metric to assign words into lines and use same strategy to assign lines into paragraphs.
[1]CBNetV2: A Composite Backbone Network Architecture for Object Detection.
[2]Hybrid Task Cascade for Instance Segmentation.