SwinTextSpotter

This is the pytorch implementation of Paper: SwinTextSpotter: Scene Text Spotting via Better Synergy between Text Detection and Text Recognition (CVPR 2022). The paper is available at this link.

• We use the models pre-trained on ImageNet. The ImageNet pre-trained SwinTransformer backbone is obtained from SwinT_detectron2.

Installation

• Python=3.8
• PyTorch=1.8.0, torchvision=0.9.0, cudatoolkit=11.1
• OpenCV for visualization

Steps

1. Install the repository (we recommend to use Anaconda for installation.)

conda create -n SWINTS python=3.8 -y
conda activate SWINTS
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install opencv-python
pip install scipy
pip install shapely
pip install rapidfuzz
pip install timm
pip install Polygon3
git clone https://github.com/mxin262/SwinTextSpotter.git
cd SwinTextSpotter
python setup.py build develop

2. dataset path

datasets
|_ totaltext
|  |_ train_images
|  |_ annotations/train.json
|_ mlt2017
|  |_ train_images
|  |_ annotations/train.json
.......

Downloaded images

• ICDAR2017-MLT [image]
• Syntext-150k:
  • Part1: 94,723 [dataset]
  • Part2: 54,327 [dataset]
• ICDAR2015 [image]
• ICDAR2013 [image]
• Total-Text [image]
• ReCTs [images&label]
• LSVT [images&label](8.2G)
• ArT [images&label]
• SynChinese130k [images&label]

Downloaded label[Google Drive] [BaiduYun] PW: a8gj

3. Pretrain SWINTS (e.g., with Swin-Transformer backbone)

python projects/SWINTS/train_net.py --num-gpus 8 --config-file projects/SWINTS/configs/SWINTS-swin-pretrain.yaml

4. Fine-tune model on the mixed real dataset

python projects/SWINTS/train_net.py --num-gpus 8 --config-file projects/SWINTS/configs/SWINTS-swin-mixtrain.yaml

5. Fine-tune model

python projects/SWINTS/train_net.py --num-gpus 8 --config-file projects/SWINTS/configs/SWINTS-swin-finetune-totaltext.yaml

6. Evaluate SWINTS (e.g., with Swin-Transformer backbone)

python projects/SWINTS/train_net.py --config-file projects/SWINTS/configs/SWINTS-swin-finetune-totaltext.yaml --eval-only MODEL.WEIGHTS ./output/model_final.pth

7. Visualize the detection and recognition results (e.g., with ResNet50 backbone)

python demo/demo.py --config-file projects/SWINTS/configs/SWINTS-swin-finetune-totaltext.yaml --input input1.jpg --output ./output --confidence-threshold 0.4 --opts MODEL.WEIGHTS ./output/model_final.pth

Example results:

Acknowlegement

Part of the codes are built on top of Detectron2, ISTR, SwinT_detectron2, Focal-Transformer and MaskTextSpotterV3.

Citation

If our paper helps your research, please cite it in your publications:

@article{huang2022swints,
  title = {SwinTextSpotter: Scene Text Spotting via Better Synergy between Text Detection and Text Recognition},
  author = {Mingxin Huang and YuLiang liu and Zhenghao Peng and Chongyu Liu and Dahua Lin and Shenggao Zhu and Nicholas Yuan and Kai Ding and Lianwen Jin},
  journal={arXiv preprint arXiv:2203.10209},
  year = {2022}
}

Copyright

For commercial purpose usage, please contact Dr. Lianwen Jin: eelwjin@scut.edu.cn

Copyright 2019, Deep Learning and Vision Computing Lab, South China China University of Technology. http://www.dlvc-lab.net