STN-OCR: A single Neural Network for Text Detection and Text Recognition

This repository contains the code for the paper: STN-OCR: A single Neural Network for Text Detection and Text Recognition

Please note that we refined our approach and released new source code. You can find the code here

Please use the new code, if you want to experiment with FSNS like data and our approach. It should also be easy to redo the text recognition experiments with the new code, although we did not release any code for that.

Structure of the repository

The folder datasets contains code related to datasets used in the paper. datasets/svhn contains several scripts that can be used to create svhn based ground truth files as used in our experiments reported in section 4.2., please see the readme in this folder on how to use the scripts. datasets/fsns contains scripts that can be used to first download the fsns dataset, second extract the images from the downloaded files and third restructure the contained gt files.

The folder mxnet contains all code used for training our networks.


In order to use the code you will need the following software environment:

  1. Install python3 (the code might work with python2, too, but this is untested)
  2. it might be a good idea to use a virtualenv
  3. install all requirements with pip install -r requirements.txt
  4. clone and install warp-ctc from here
  5. go into the folder mxnet/metrics/ctc and run python build_ext --inplace
  6. clone the mxnet repository
  7. checkout the tag v0.9.3
  8. add the warpctc plugin to the project by enabling it in the file
  9. compile mxnet
  10. install the python bindings of mxnet
  11. You should be ready to go!


You can use this code to train models for three different tasks.

SVHN House Number Recognition

The file is the entry point for training a network using our purpose build svhn datasets. The file as such is ready to train a network capable of finding a single house number placed randomly on an image.

Example: centered_image

In order to do this, you need to follow these steps:

  1. Download the datasets
  2. Locate the folder generated/centered
  3. open train.csv and adapt the paths of all images to the path on your machine (do the same with valid.csv)
  4. make sure to prepare your environment as described in installation
  5. start the training by issuing the following command:

    python <path to train.csv> <path to valid.csv> --gpus <gpu id you want to use> --log-dir <where to save the logs> -b <batch size you want ot use> --lr 1e-5 --zoom 0.5 --char-map datasets/svhn/svhn_char_map.json

  6. Wait and enjoy.

If you want to do experiments on more challenging images you might need to update some parts of the code in The parts you might want to update are located around line 40 in this file. Here you can change the max. number of house numbers in the image (num_timesteps), the maximum number of characters per house number (labels_per_timestep), the number of rnn layers to use for predicting the localization num_rnn_layers and whether to use a blstm for predicting the localization or not use_blstm.

A quite more challenging dataset is contained in the folder medium_two_digits, or medium in the datasets folder. Example: 2_digits_more_challenge

If you want to follow our experiments with svhn numbers placed in a regular grid you'll need to do the following:

  1. Download the datasets
  2. Locate the folder generated/easy
  3. open train.csv and adapt the paths of all images to the path on your machine (do the same with valid.csv)
  4. set num_timesteps and labels_per_timestep to 4 in
  5. start the training using the following command: python <path to train.csv> <path to valid.csv> --gpus <gpu id you want to use> --log-dir <where to save the logs> -b <batch size you want ot use> --lr 1e-5
  6. If you are lucky it will work ;)

Text Recognition

Following our text recognition experiments might be a little difficult, because we can not offer the entire dataset used by us. But it is possible to perform the experiments based on the Synth-90k dataset provided by Jaderberg et al. here. After downloading and extracting this file you'll need to adapt the groundtruth file provided with this dataset to fit to the format used by our code. Our format is quite easy. You need to create a csv file with tabular separated values. The first column is the absolute path to the image and the rest of the line are the labels corresponding to this image.

To train the network you can use the script. You can start this script in a similar manner to the script.


In order to redo our experiments on the FSNS dataset you need to perform the following steps:

  1. Download the fsns dataset using the script located in datasets/fsns
  2. Extract the individual images using the script located in datasets/fsns/tfrecord_utils (you will need to install tensorflow for doing that)
  3. Use the script to transform the original fsns groundtruth, which is based on a single line to a groundtruth containing labels for each word individually. A possible usage of the script could look like this:

    python <path to original gt> datasets/fsns/fsns_char_map.json <path to gt that shall be generated>

  4. Because MXNet expects the blank label to be 0 for the training with CTC Loss, you have to use the script in datasets/fsns and swap the class for space and blank in the gt, by issuing:

    python <original gt> <swapped gt> 0 133

  5. After performing these steps you should be able to run the training by issuing:

    python <path to generated train gt> <path to generated validation gt> --char-map datases/fsns/fsns_char_map.json --blank-label 0

Observing the Training Progress

We've added a nice script that makes it possible to see how well the network performs at every step of the training. This progress is normally plotted to disk for each iteration and can later on be used to create animations of the train progress (you can use the and scripts located in mxnet/utils for this purpose). Besides this normal plotting to disk it is also possible to directly see this progress while the training is running. In order to see this you have to do the following:

  1. start the script in mxnet/utils
  2. start the training with the following additional command line params:

    --send-bboxes --ip <localhost, or remote ip if you are working on a remote machine> --port <the port the script is running on (default is 1337)

  3. enjoy!

This tool is especially helpful in determining whether the network is learning anything or not. We recommend that you always use this tool while training.


If you want to evaluate already trained models you can use the evaluation scripts provided in the mxnet folder. For evaluating a model you need to do the following:

  1. train or download a model
  2. choose the correct evaluation script an adapt it, if necessary (take care in case you are fiddling around with the amount of timesteps and number of RNN layers)
  3. Get the dataset you want to evaluate the model on and adapt the groundtruth file to fit the format expected by our software. The format expected by our software is defined as a csv (tab separated) file that looks like that: <absolute path to image> \t <numerical labels each label separated from the other by \t>
  4. run the chosen evaluation script like so

    python eval_<type> <path to model dir>/<prefix of model file> <number of epoch to test> <path to evaluation gt> <path to char map>

You can use for evaluating a model trained with CTC on the original svhn dataset, the script for evaluating a model trained for text recognition, and the for evaluating a model trained on the FSNS dataset.


This Code is licensed under the GPLv3 license. Please see further details in


If you are using this Code please cite the following publication:

  title={STN-OCR: A single Neural Network for Text Detection and Text Recognition},
  author={Bartz, Christian and Yang, Haojin and Meinel, Christoph},
  journal={arXiv preprint arXiv:1707.08831},

A short note on code quality

The code contains a huge amount of workarounds around MXNet, as we were not able to find any easier way to do what we wanted to do. If you know a better way, pease let us know, as we would like to have code that is better understandable, as now.