Introduction

This repository is the official implementation of PAN++. Compared to pan_pp.pytorch, this repository is specific to PAN++, and more stable.

Installation

First, clone the repository locally:

git clone https://github.com/whai362/pan_pp_stable.git

Then, install PyTorch 1.1.0+, torchvision 0.3.0+, and other requirements:

conda install pytorch torchvision -c pytorch
pip install -r requirement.txt

Finally, compile codes of post-processing:

# build pa and other post-processing algorithms
sh ./compile.sh

Dataset

Please refer to dataset/README.md for dataset preparation.

Training & Testing

ICDAR2015: please refer to IC15_RESULTS.md for training and testing.

RCTW-17: please refer to RCTW17_RESULTS.md for training and testing.

Total-Text: please refer to TT_RESULTS.md for training and testing.

CTW1500: please refer to CTW_RESULTS.md for training and testing.

MSRA-TD500: please refer to MSRA_RESULTS.md for training and testing.

Evaluate the performance

cd eval/
./eval_{DATASET}.sh

Evaluate the speed

python test.py XXX --report_speed true

Visualization

python test.py XXX --vis true

Citation

Please cite the related works in your publications if it helps your research:

PAN++

@article{wang2021pan++,
  title={PAN++: Towards Efficient and Accurate End-to-End Spotting of Arbitrarily-Shaped Text},
  author={Wang, Wenhai and Xie, Enze and Li, Xiang and Liu, Xuebo and Liang, Ding and Zhibo, Yang and Lu, Tong and Shen, Chunhua},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2021},
  publisher={IEEE}
}

License

This project is developed and maintained by IMAGINE Lab@National Key Laboratory for Novel Software Technology, Nanjing University.

This project is released under the Apache 2.0 license.