This is a pytorch implementation for paper ADNet (TMM2022). ADNet is a shrunk polygon-based scene text detector, which uses an instance-wise dilation factor to obtain more complete and tight results. This repository is built on DBNet.
- Release code
- Document for Installation
- Trained models
- Document for testing and training
- Evaluation
- Python==3.7
- Pytorch==1.2
- CUDA==9.2
git clone https://github.com/qqqyd/ADNet.git
cd ADNet/
conda create --name ADNet -y
conda activate ADNet
conda install ipython
pip install torch==1.2.0+cu92 torchvision==0.4.0+cu92 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirement.txt
cd assets/ops/dcn/
python setup.py build_ext --inplacePrepare the datasets and put them in datasets/.
Download the trained models in Google Drive and put them in models/.
Evaluate the models using following commands:
CUDA_VISIBLE_DEVICES=0 python eval.py experiments/seg_detector/totaltext.yaml --resume models/adnet_total --polygon --box_thresh 0.7
CUDA_VISIBLE_DEVICES=0 python eval.py experiments/seg_detector/td500.yaml --resume models/adnet_td500 --polygon --box_thresh 0.7
CUDA_VISIBLE_DEVICES=0 python eval.py experiments/seg_detector/ctw1500.yaml --resume models/adnet_ctw1500 --polygon --box_thresh 0.8
CUDA_VISIBLE_DEVICES=0 python eval.py experiments/seg_detector/ic15.yaml --resume models/adnet_ic15 --polygon --box_thresh 0.8CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py <path-to-yaml-file> --name <task-name> --resume <pretrained model on SynthText (optional)> --num_gpus 4If you find our method useful for your reserach, please cite
@ARTICLE{qu2022adnet,
author={Qu, Yadong and Xie, Hongtao and Fang, Shancheng and Wang, Yuxin and Zhang, Yongdong},
journal={IEEE Transactions on Multimedia},
title={ADNet: Rethinking the Shrunk Polygon-Based Approach in Scene Text Detection},
year={2023},
volume={25},
number={},
pages={6983-6996},
doi={10.1109/TMM.2022.3216729}}