GNNets
Now in experimental release, suggestions welcome.
Paper
Youjiang Xu*, Jiaqi Duan*, Zhanghui Kuang§, Xiaoyu Yue, Hongbin Sun, Yue Guan, Wei Zhang. "Geometry Normalization Networks for Accurate Scene Text Detection" . [Paper]. In ICCV 2019.
@InProceedings{Xu_2019_Geometry,
author = {Xu, Youjiang and Duan, Jiaqi and Kuang, Zhanghui and Yue, Xiaoyu and Sun, Hongbin and Guan, Yue and Zhang, Wei},
title = {Geometry Normalization Networks for Accurate Scene Text Detection},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {Oct},
year = {2019}
}- Note that * means authors contributed equally, § means the corresponding author.
Framework
- The framework of the proposed Geometry Normalization Networks. The feature maps extracted by the backbone are fed into the Geometry Normalization Module (GNM) with multi-branches, each of which is composed of one Scale Normalization Unit (SNU)
$\mathcal{F}^s$ and Orientation Normalization Unit (ONU)$\mathcal{F}^o$ . There are two different scale normalization units ($\mathcal{S}, \mathcal{S}_\frac{1}{2} $ ) and four orientation normalization units ($\mathcal{O}, \mathcal{O}_r, \mathcal{O}f, \mathcal{O}{r+f}$). With different combinations of SNU and ONU, GNM generates different geometry normalized feature maps, which are fed into one shared text detection header.
The Rotated ICDAR 2015
- The Rotated ICDAR 2015 Benchmark could be download form here
Todo List
- update model performance
- add rotated ICDAR 2015 benchmark dataset