This is an implement of "R-Net: A Relationship Network for Efficient and Accurate Scene Text Detection". R-Net is a direct regression based method for multi-oriented scene text detection, which outperforms parallel methods by a large margin. The paper is avaliabel at Paperlink(TMM2020).
| Model | recall | precision | F-measure | TIoU-R | TIoU-P | TIoU-F |
|---|---|---|---|---|---|---|
| This implementation (IC15) | 84.5 | 88.1 | 86.3 | 60.1 | 67.9 | 63.7 |
The released model is pre-trained on SynthText (1 epoch) and then finetune on ICDAR2015 (batchsize=8 on 1 gpu). Reasults on other datasets (e.g. MSRA-TD500, ICDAR2013, MLT, etc.) can be easily obtained by following our training setting.
2020/5/18 We have updated the code.
- Python3
- PyTorch-0.4.1
- torchvision-0.2.1
- shapely-1.6.4.post2
- lanms-1.0.2
- opencv-python(4.1.0.25)
Download data images and annotions from ICDAR 2015 Challenge 4. Prepare data as:
./dataset/train/imgs
./dataset/train/txt
./dataset/test/imgs
./dataset/test/txt
Download pre-trained VGG16 (must rename as vgg16.pth) and our traind model (passward:oowz) Put both models into output/
CUDA_VISIBLE_DEVICES=0 python test_rnet.pyCUDA_VISIBLE_DEVICES=0,1 python train_rnet.pyWe use online tool to evaluate our results.
Run
CUDA_VISIBLE_DEVICES=0 python speed_eval.py.We add the model prediction time and the NMS time as the inference speed.
If you find our method useful for your reserach, please cite
@ARTICLE{rnet2020wang,
author={Y. {Wang} and H. {Xie} and Z. {Zha} and Y. {Tian} and Z. {Fu} and Y. {Zhang}},
journal={IEEE Transactions on Multimedia},
title={R-Net: A Relationship Network for Efficient and Accurate Scene Text Detection},
year={2020},
pages={1-1},}Suggestions and discussions are greatly welcome. Please contact the authors by sending email to [email protected]