Skip to content
A tensorflow implementation of EAST text detector
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
demo_images demo server Aug 6, 2017
templates also show loadavg Aug 7, 2017
.gitignore demo server Aug 6, 2017
LICENSE add init file and fix path problem Aug 1, 2017 remove keras dependency Aug 3, 2017 update deploy script Aug 7, 2017 minor fix Feb 17, 2019 Update Aug 27, 2017 fix batch_size variable Aug 3, 2017

EAST: An Efficient and Accurate Scene Text Detector


This is a tensorflow re-implementation of EAST: An Efficient and Accurate Scene Text Detector. The features are summarized blow:

  • Online demo
  • Only RBOX part is implemented.
  • A fast Locality-Aware NMS in C++ provided by the paper's author.
  • The pre-trained model provided achieves 80.83 F1-score on ICDAR 2015 Incidental Scene Text Detection Challenge using only training images from ICDAR 2015 and 2013. see here for the detailed results.
  • Differences from original paper
    • Use ResNet-50 rather than PVANET
    • Use dice loss (optimize IoU of segmentation) rather than balanced cross entropy
    • Use linear learning rate decay rather than staged learning rate decay
  • Speed on 720p (resolution of 1280x720) images:
    • Now
      • Graphic card: GTX 1080 Ti
      • Network fprop: ~50 ms
      • NMS (C++): ~6ms
      • Overall: ~16 fps
    • Then
      • Graphic card: K40
      • Network fprop: ~150 ms
      • NMS (python): ~300ms
      • Overall: ~2 fps

Thanks for the author's (@zxytim) help! Please cite his paper if you find this useful.


  1. Installation
  2. Download
  3. Demo
  4. Test
  5. Train
  6. Examples


  1. Any version of tensorflow version > 1.0 should be ok.


  1. Models trained on ICDAR 2013 (training set) + ICDAR 2015 (training set): BaiduYun link GoogleDrive
  2. Resnet V1 50 provided by tensorflow slim: slim resnet v1 50


If you want to train the model, you should provide the dataset path, in the dataset path, a separate gt text file should be provided for each image and run

python --gpu_list=0 --input_size=512 --batch_size_per_gpu=14 --checkpoint_path=/tmp/east_icdar2015_resnet_v1_50_rbox/ \
--text_scale=512 --training_data_path=/data/ocr/icdar2015/ --geometry=RBOX --learning_rate=0.0001 --num_readers=24 \

If you have more than one gpu, you can pass gpu ids to gpu_list(like --gpu_list=0,1,2,3)

Note: you should change the gt text file of icdar2015's filename to img_*.txt instead of gt_img_*.txt(or you can change the code in, and some extra characters should be removed from the file. See the examples in training_samples/


If you've downloaded the pre-trained model, you can setup a demo server by

python3 --checkpoint-path /tmp/east_icdar2015_resnet_v1_50_rbox/

Then open http://localhost:8769 for the web demo. Notice that the URL will change after you submitted an image. Something like ?r=49647854-7ac2-11e7-8bb7-80000210fe80 appends and that makes the URL persistent. As long as you are not deleting data in static/results, you can share your results to your friends using the same URL.

URL for example below: web-demo



python --test_data_path=/tmp --gpu_list=0 --checkpoint_path=/tmp/east_icdar2015_resnet_v1_50_rbox/ \

a text file will be then written to the output path.


Here are some test examples on icdar2015, enjoy the beautiful text boxes! image_1 image_2 image_3 image_4 image_5


Please let me know if you encounter any issues(my email boostczc@gmail dot com).

You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.