Skip to content

Latest commit

 

History

History

pan

README.md

Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network

Introduction

@inproceedings{wang2019efficient,
  title={Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network},
  author={Wang, Wenhai and Xie, Enze and Song, Xiaoge and Zang, Yuhang and Wang, Wenjia and Lu, Tong and Yu, Gang and Shen, Chunhua},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={8440--8449},
  year={2019}
}

Note that, the original PAN is based on Python 2.7 and Pytorch 0.4.1. When migrating it to Python 3.6 and Pytorch 1.1.0, we make the following two changes to the default settings.

  • Using Adam optimizer;
  • PolyLR is also used in the pre-training phase.

Results and Models

  • Total-Text
Method Backbone Finetune Precision (%) Recall (%) F-measure (%) Config Download
PAN ResNet18 N 87.9 79.6 83.5 config model
PAN (paper) ResNet18 N 88.0 79.4 83.5 - -
PAN ResNet18 Y 88.5 81.7 85.0 config model
PAN (paper) ResNet18 Y 89.3 81.0 85.0 - -
  • CTW1500
Method Backbone Finetune Precision (%) Recall (%) F-measure (%) Config Download
PAN ResNet18 N 85.1 79.1 82.0 config model
PAN (paper) ResNet18 N 84.6 77.7 81.0 - -
PAN ResNet18 Y 86.0 81.0 83.4 config model
PAN (paper) ResNet18 Y 86.4 81.2 83.7 - -
  • ICDAR 2015
Method Backbone Finetune Precision (%) Recall (%) F-measure (%) Config Download
PAN ResNet18 N 84.4 77.5 80.8 config model
PAN (paper) ResNet18 N 82.9 77.8 80.3 - -
PAN ResNet18 Y 86.6 79.7 83.0 config model
PAN (paper) ResNet18 Y 84.0 81.9 82.9 - -
  • MSRA-TD500
Method Backbone Finetune Precision (%) Recall (%) F-measure (%) Config Download
PAN ResNet18 N 82.0 79.4 80.7 config model
PAN (paper) ResNet18 N 80.7 77.3 78.9 - -
PAN ResNet18 Y 85.7 83.4 84.5 config model
PAN (paper) ResNet18 Y 84.4 83.8 84.1 - -
  • SynthText
Method Backbone Config Download
PAN ResNet18 config model