R2CNN in PyTorch 1.2

Pytorch Implementation of "R2CNN Rotational Region CNN for Orientation Robust Scene Text Detection" paper , it is based on facebook's maskrcnn-benchmark

Installation

Check INSTALL.md for installation instructions.

Perform training on ICDAR2015 dataset

1. Download icdar2015 dataset and pretrain model from maskrcnn-bencmark

cd ./tools
mkdir datasets
ln -s PATH_ICDAR2015 datasets/ICDAR2015
mkdir pretrain
cd pretrain
wget https://download.pytorch.org/models/maskrcnn/e2e_faster_rcnn_R_50_FPN_1x.pth

2. Convert annotations to COCO style

cd ./tools/ICDAR2015
python convert_icdar_to_coco.py

3. start training

cd ./tools
python train_net.py

Inference on ICDAR 2015 dataset

1. Download model or use your own model

2. single image inference

cd ./tools
python inference_engine.py

New feature compared with maskrcnn-benchmark

new data structure quad_bbox(x1, y1, x2, y2, x3, y3, x4, y4) is defined to replace bbox(x1, y1, x2, y2)
an extra branch in box_head which regress offsets of 4 points
post processor of rpn is adjusted to detect text objects

TODO

Citations

Please consider citing this project in your publications if it helps your research. The following is a BibTeX reference. The BibTeX entry requires the url LaTeX package.

@misc{r2cnn,
author = {Yingying Jiang, Xiangyu Zhu, Xiaobing Wang, Shuli Yang, Wei Li, Hua Wang, Pei Fu, Zhenbo Luo},
title = {R2CNN: Rotational Region CNN for Orientation Robust Scene Text Detection},
conference = {ICPR2018}
year = {2017},
}