News

• PSENet is included in MMOCR.
• We have upgraded PSENet from python2 to python3. You can find the old version here.
• We have implemented PSENet using Paddle. Visit it here.
• You can find code of PAN here.
• Another group also implemented PSENet using Paddle. You can visit it here. You can also have a try online with all the environment ready here.

Introduction

Official Pytorch implementations of PSENet [1].

[1] W. Wang, E. Xie, X. Li, W. Hou, T. Lu, G. Yu, and S. Shao. Shape robust text detection with progressive scale expansion network. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages 9336–9345, 2019.

Recommended environment

Python 3.6+
Pytorch 1.1.0
torchvision 0.3
mmcv 0.2.12
editdistance
Polygon3
pyclipper
opencv-python 3.4.2.17
Cython

Install

pip install -r requirement.txt
./compile.sh

Training

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py ${CONFIG_FILE}

For example:

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py config/psenet/psenet_r50_ic15_736.py

Test

python test.py ${CONFIG_FILE} ${CHECKPOINT_FILE}

For example:

python test.py config/psenet/psenet_r50_ic15_736.py checkpoints/psenet_r50_ic15_736/checkpoint.pth.tar

Speed

python test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --report_speed

For example:

python test.py config/psenet/psenet_r50_ic15_736.py checkpoints/psenet_r50_ic15_736/checkpoint.pth.tar --report_speed

Evaluation

Introduction

The evaluation scripts of ICDAR 2015 (IC15), Total-Text (TT) and CTW1500 (CTW) datasets.

ICDAR 2015

Text detection

./eval_ic15.sh

Total-Text

Text detection

./eval_tt.sh

CTW1500

Text detection

./eval_ctw.sh

Benchmark

Results

ICDAR 2015

Method	Backbone	Fine-tuning	Scale	Config	Precision (%)	Recall (%)	F-measure (%)	Model
PSENet	ResNet50	N	Shorter Side: 736	psenet_r50_ic15_736.py	83.6	74.0	78.5	Google Drive
PSENet	ResNet50	N	Shorter Side: 1024	psenet_r50_ic15_1024.py	84.4	76.3	80.2	Google Drive
PSENet (paper)	ResNet50	N	Longer Side: 2240	-	81.5	79.7	80.6	-
PSENet	ResNet50	Y	Shorter Side: 736	psenet_r50_ic15_736_finetune.py	85.3	76.8	80.9	Google Drive
PSENet	ResNet50	Y	Shorter Side: 1024	psenet_r50_ic15_1024_finetune.py	86.2	79.4	82.7	Google Drive
PSENet (paper)	ResNet50	Y	Longer Side: 2240	-	86.9	84.5	85.7	-

CTW1500

Method	Backbone	Fine-tuning	Config	Precision (%)	Recall (%)	F-measure (%)	Model
PSENet	ResNet50	N	psenet_r50_ctw.py	82.6	76.4	79.4	Google Drive
PSENet (paper)	ResNet50	N	-	80.6	75.6	78	-
PSENet	ResNet50	Y	psenet_r50_ctw_finetune.py	84.5	79.2	81.8	Google Drive
PSENet (paper)	ResNet50	Y	-	84.8	79.7	82.2	-

Total-Text

Method	Backbone	Fine-tuning	Config	Precision (%)	Recall (%)	F-measure (%)	Model
PSENet	ResNet50	N	psenet_r50_tt.py	87.3	77.9	82.3	Google Drive
PSENet (paper)	ResNet50	N	-	81.8	75.1	78.3	-
PSENet	ResNet50	Y	psenet_r50_tt_finetune.py	89.3	79.6	84.2	Google Drive
PSENet (paper)	ResNet50	Y	-	84.0	78.0	80.9	-

Citation

@inproceedings{wang2019shape,
  title={Shape robust text detection with progressive scale expansion network},
  author={Wang, Wenhai and Xie, Enze and Li, Xiang and Hou, Wenbo and Lu, Tong and Yu, Gang and Shao, Shuai},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={9336--9345},
  year={2019}
}

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.