method: hiertext_submit_0401_curve_199_v22023-04-01

Authors: Zhong Humen, Tang Jun, Yang zhibo, Song xiaoge

Affiliation: Alibaba DAMO OCR Team


Description: Our method is a single end-to-end model designed for hierarchical text detection. Our model utilizes the pipeline of DETR-like methods and design a hierarchical decoder so that the model can detect more text instances with less queries for reducing computational cost.
The model uses ImageNet pretrained Swin-S as backbone and is trained only on HierText training set. Single-scale inference is used during testing. No external data and synthetic data is used.