crossLi / stela

Scene Text Detection with Learned Anchor

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

STELA: A Real-Time Scene Text Detector with Learned Anchor

STELA is a simple and intuitive method for multi-oriented text detection based on RetinaNet. The key idea is utilizing the learned anchor which is obtained through a regression operation to replace the original into the final predictions. In our experiments, it achieves an F-measure 0.887 on ICDAR 2013, 0.833 on ICDAR 2015 and 0.715 on ICDAR 2017 MLT. For more details, please refer to our paper.


This code is modified from RetinaNet and maskrcnn-benchmark. It has been tested on Ubuntu 16.04 with CUDA 9.0 and PyTorch 1.1. If you have any issue, please leave a message.

  1. Clone this repository

    git clone
  2. Build the Cython modules

    cd $STELA_ROOT/utils
  3. Prepare your own data directory. For our implement, it should follow the format of PASCAL VOC.

  4. Train with YOUR dataset

    cd $STELA_ROOT
  5. Test with YOUR models

    cd $STELA_ROOT


At last, if you find the paper and code useful in your research, please consider citing:

	Title = {STELA: A Real-Time Scene Text Detector with Learned Anchor},
	Author = {Linjie Deng and Yanxiang Gong and Xinchen Lu and Yi Lin and Zheng Ma and Mei Xie},
	Journal = {arXiv preprint arXiv:1909.07549},
	Year = {2019}


Scene Text Detection with Learned Anchor


Language:Python 99.9%Language:Makefile 0.1%Language:Shell 0.1%