r-net's Introduction

R-Net

Description

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).

Result

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.

Updates

2020/5/18 We have updated the code.

Requirement

Python3
PyTorch-0.4.1
torchvision-0.2.1
shapely-1.6.4.post2
lanms-1.0.2
opencv-python(4.1.0.25)

Installation

Data

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

Pre-trained Model and Our trained Model.

Download pre-trained VGG16 (must rename as vgg16.pth) and our traind model (passward:oowz) Put both models into output/

Test

CUDA_VISIBLE_DEVICES=0 python test_rnet.py

Train

CUDA_VISIBLE_DEVICES=0,1 python train_rnet.py

Evaluation

We use online tool to evaluate our results.

Speed

Run

CUDA_VISIBLE_DEVICES=0 python speed_eval.py.

We add the model prediction time and the NMS time as the inference speed.

Citation

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},}

Feedback

Suggestions and discussions are greatly welcome. Please contact the authors by sending email to [email protected]

Recommend Projects

dreadlord1984 / r-net Goto Github PK