Authors: ShaoAn Lu

Description: Our network was based on the AdvancedEAST approach. We used a modified version of ResNet50 from [https://arxiv.org/abs/1901.06651] combined with feature fusion modules proposed in M2Det [https://arxiv.org/abs/1811.04533]. The model was trained from scratch using dataset from ICDAR2019 LSVT and ReCTS. Group normalization and weight normalization [https://arxiv.org/abs/1903.10520] are introduced for small batch size training. The multi-scale predictions generated by the aforementioned model are further refined via a YOLOv3 model, which was trained for character-level detection task, to catch missed small textlines.

Name: Shao-An Lu (呂紹安)
Organization: awesome techno