SEE: Towards Semi-Supervised End-to-End Scene Text Recognition
Code for the AAAI 2018 publication "SEE: Towards Semi-Supervised End-to-End Scene Text Recognition". You can read a preprint on Arxiv
Installation
You can install the project directly on your PC or use a Docker container
Directly on your PC
- Make sure to use Python 3
- It is a good idea to create a virtual environment (example for creating a venv)
- Make sure you have the latest version of CUDA (>= 8.0) installed
- Install CUDNN (> 6.0)
- Install NCCL (> 2.0) installation guide
- Install all requirements with the following command:
pip install -r requirements.txt
- Check that chainer can use the GPU:
- start the python interpreter:
python
- import chainer:
import chainer
- check that cuda is available:
chainer.cuda.available
- check that cudnn is enabled:
chainer.cuda.cudnn_enabled
- the output of both commands should be
True
- start the python interpreter:
Using Docker
- Install
Docker
- Windows: Get it here
- Mac: Get it here
- Linux: User your favourite package manager i.e.
pacman -S docker
, or use this guide for Ubuntu.
- Install CUDA related things:
- CUDA (>= 8.0) installed
- CUDNN (> 6.0)
- nvidia-docker (Ubuntu, Arch Like OS))
- Get NCCL
- make sure to download the version for Ubuntu 16.04, that fits to your local CUDA configuration (i.e. you have installed CUDA 9.1 take the version for CUDA 9.1, if you have CUDA 8, take the version for CUDA 8)
- place it in the root folder of the project
- Build the Docker image
docker build -t see .
- If your host system uses CUDA with a version earlier than 9.1, specify the corresponding docker image to match the configuration of your machine (see this list for available options). For example, for CUDA 8 and CUDNN 6 use the following instead:
docker build -t see --build-arg FROM_IMAGE=nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04 .
- if you did not download a file called
nccl-repo-ubuntu1604-2.1.15-ga-cuda9.1_1-1_amd64.deb
, set the argumentNCCL_NAME
to the name of the file you downloaded. For example:
docker build -t see --build-arg NCCL_NAME=nccl-repo-ubuntu1604-2.1.15-ga-cuda9.0_1-1_amd64.deb .
- Check that everything is okay, by entering a shell in the container and do the following:
- run the container with:
nvidia-docker run -it see
- start the python interpreter:
python3
- import chainer:
import chainer
- check that cuda is available:
chainer.cuda.available
- check that cudnn is enabled:
chainer.cuda.cudnn_enabled
- the output of both commands should be
True
- run the container with:
- Hint: make sure to mount all data folders you need into the container with the
-v
option for running a container.
General Training Hints
If you like to train a network with more than 4 words per image, you will need to adjust or delete the loss_weights
(see this line). Otherwise, the code will throw errors at you. They are mainly meant for training FSNS models and should be discarded when training other models.
SVHN Experiments
We performed several experiments on the SVHN dataset. First, we tried to see whether our architecture is able to reach competitive results on the SVHN recognition challenge. Second, we wanted to determine whether our localization network can find a text distributed on a given grid. In our last experiment we created a dataset, where we randomly distributed the text samples on the image.
Datasets
This section describes what needs to be done in order to get/prepare the data. There is no need for creating the custom datasets by yourself, we also offer them for download. The information on how to create the datasets is included here for reference.
Original SVHN data
- Get the original SVHN datset from here.
- Extract the label data using the script
datasets/svn/svhn_dataextract_to_json.py
. - use the script
datasets/svhn/prepare_svhn_crops.py
to crop all bounding boxes, including some background from the SVHN images. Use the script like that:python prepare_svhn_crops.py <path to svhn json> 64 <where to save the cropped images> <name of stage>
. For more information about possible commands you can usepython prepare_svhn_crops.py -h
.
Grid Dataset
- Follow steps 1 and 2 of the last subsection in order to get all SVHN images and the corresponding groundtruth.
- The script
datasets/svhn/create_svhn_dataset_4_images.py
can be used to create the dataset. - The command
python create_svhn_dataset_4_images.py -h
shows all available command line options for this script
Random Dataset
- Follow steps 1 and 2 of the first subsection in order to get all SVHN images and the corresponding groundtruth.
- The script
datasets/svhn/create_svhn_dataset.py
can be used to create the dataset. - The command
python create_svhn_dataset.py -h
shows all available command line options for this script.
Dataset Download
You can also download already created datasets here.
Training the model
You can use the script train_svhn.py
to train a model that can detect and recognize SVHN like text.
The script is tuned to use the custom datasets and should enable you to redo these experiments.
Preparations
- Make sure that you have one of the datasets.
- For training you will need:
- the file
svhn_char_map.json
(you can find it in the folderdatasets/svhn
) - the ground truth files of the dataset you want to use
- the file
- Add one line to the beginning of each ground truth file:
<number of house numbers in image> <max number of chars per house number>
(both values need to be separated by a tab character). If you are using the grid dataset it could look like that:4 4
. - prepare the curriculum specification as a
json
file, by following this template:if you want to train using the curriculum learning strategy, you just need to add further dicts to this list.[ { "train": "<path to train file>", "validation": "<path to validation file>" } ]
- use the script
chainer/train_svhn.py
for training the network.
Starting the training
The training can be run on GPU or CPU. You can also use multiple GPUs in a data parallel fashion.
In order to specify which GPU to use just add the command line parameter -g <id of gpu to use>
e.g. -g 0
for using the first GPU.
You can get a brief explanation of each command line option of the script train_svhn.py
by running
the script like this: python train_svhn.py -h
You will need to specify at least the following parameters:
dataset_specification
- this is the path to thejson
file you just createdlog_dir
- this is the path to the directory where the logs shall be saved--char-map ../datasets/svhn/svhn_char_map.json
- path to the char map for mapping classes to labels.--blank-label 0
- indicates that class 0 is the blank label-b <batch-size>
- set the batch size used for training
FSNS Experiments
In order to see, whether our idea is applicable in practice, we also did experiments on the FSNS dataset. The FSNS dataset contains images of French street name signs. The most notable characteristic of this dataset is, that this dataset does not contain any annotation for text localization. This fact makes this dataset quite suitable for our method, as we claim that we can locate and recognize text, even without the corresponding ground truth for localization.
Preparing the Dataset
Getting the dataset and making it usable with deep learning frameworks like Chainer is not an easy task. We provide some scripts that will download the dataset, convert it from the tensorflow format to single images and create a ground truth file, that is usable by our train code.
The folder datasets/fsns
contains all scripts that are necessary for preparing
the dataset. These steps need to be done:
- use the script
download_fsns.py
for getting the dataset. You will need to specify a directory, where the data shall be saved. - the script
tfrecord_to_image.py
extracts all images and labels from the downloaded dataset. - We advise you to use the script
swap_classes.py
. With this script we will set the class of the blank label to be0
, as it is defined in the class to label mapfsns_char_map.json
. You can invoke the script like this:python swap_classes.py <gt_file> <output_file_name> 0 133
- next, you will need to transform the original ground truth, to the ground truth
format we used for training. Our ground truth format differs because we
found that it is not possible to train the model if the word boundaries are not
explicitly given to the model. We, therefore, transform the line based ground truth
to a word based ground truth. You can use the script
transform_gt.py
for doing that. You could call the script like that:python transform_gt.py <path to original gt> fsns_char_map.json <path to new gt>
.
Training the Network
Before you can start training the network, you will need to do the following preparations:
In the last section, we already introduced the transform_gt.py
script.
As we found that it is only possible to train a new model on the FSNS dataset,
when using a curriculum learning strategy, we need to create a learning curriculum
prior to starting the training. You can do this by following these steps:
- create ground truth files for each step of the curriculum with the
transform_gt.py
script.- start with a reasonable number of maximum words (2 is a good choice here)
- create a ground truth file with all images that contain max. 2 words by using the
transform_gt.py
script:python transform_gt.py <path to downloaded gt> fsns_char_map.json <path to 2 word gt> --max-words 2 --blank-label 0
- Repeat this step with 3 and 4 words (you can also take 5 and 6, too), but make sure
to only include images with the corresponding amount of words (
--min-words
is the flag to use)
- Add the path to your files to a
.json
file that could be calledcurriculum.json
This file works exactly the same as the file discussed in step 3 in the preparations section for the SVHN experiments.
Once you are done with this, you can actually train the network ๐
Training the network happens, by using the train_fsns.py
script.
python train_fsns.py -h
shows all available command-line options.
This script works very similarly to the train_svhn.py
script
You will need to specify at least the following parameters:
dataset_specification
- this is the path to thejson
file you just createdlog_dir
- this is the path to the directory where the logs shall be saved--char-map ../datasets/fsns/fsns_char_map.json
- path to the char map for mapping classes to labels.--blank-label 0
- indicates that class 0 is the blank label-b <batch-size>
- set the batch size used for training
FSNS Demo
In case you only want to see how the model behaves on a given image, you can use the fsns_demo.py
script.
This script expects a trained model, an image and a char map and prints you the predicted words in the
image + the predicted bounding boxes.
If you download the model provided here, you could call the script like this:
python fsns_demo.py <path to log directory> model_35000.npz <path to example image> ../datasets/fsns/fsns_char_map.json
It should be fairly easy to extend this script to also work with other models. Just have a look at how the different evaluators create the network
and how they extract the characters from the predictions and you should be good to go!
Text Recognition
Although not mentioned in the paper, we also provide a model with which, you can perform text recognition on already cropped text lines. We also provide code for training such a model. Everything works very similar to the scripts provided for SVHN and FSNS.
Dataset
Unfortunately, we can not offer our entire train dataset for download, as it is way too huge. But if you want to train a text recognition model on your own, you can use the "Synthetic Word Dataset" (download it here). After you've downloaded the dataset, you will need to do some post processing and create a groundtruth similar to the one for the FSNS dataset. We provide a sample dataset at the location, where you can also download the text recognition model (which is here).
Training
After you are done with preparing the dataset, you can start training.
Training the network happens, by using the train_text_recognition.py
script.
python train_text_recognition.py -h
shows all available command-line options.
This script works very similarly to the train_svhn.py
and train_fsns.py
script
You will need to specify at least the following parameters:
dataset_specification
- this is the path to thejson
file you just createdlog_dir
- this is the path to the directory where the logs shall be saved--char-map ../datasets/textrec/ctc_char_map.json
- path to the char map for mapping classes to labels.--blank-label 0
- indicates that class 0 is the blank label-b <batch-size>
- set the batch size used for training
Text Recognition Demo
Analog to the fsns_demo.py
script, we offer a demo script for text recognition named text_recognition_demo.py
.
This script expects a trained model, an image and a char map and prints you the predicted words in the
image + the predicted bounding boxes.
If you download the model provided here, you could call the script like this:
python text_recognition_demo.py <path to log directory> model_190000.npz <path to example image> ../datasets/textrec/ctc_char_map.json
It should be fairly easy to extend this script to also work with other models. Just have a look at how the different evaluators create the network
and how they extract the characters from the predictions and you should be good to go!
Pretrained Models
You can download our best performing model on the FSNS dataset, a model for our SVHN experiments and also a model for our text recognition experiments here.
General Notes on Training
This section contains information about things that happen while a network is training. It includes a description of all data that is being logged and backed up for each training run and a description of a tool that can be used to inspect the training, while it is running.
Contents of the log dir
The code will create a new subdirectory in the log dir, where it puts all data that is to be logged. The code logs the following pieces of data:
- it creates a backup of the currently used network definition files
- it saves a snapshot of the model at each epoch, or after
snapshot_interval
iterations (default 5000) - it saves loss and accuracy values at the configured print interval (each time after 100 iterations)
- it will save the prediction of the model on a given or randomly chosen sample. This visualization helps with assessing, whether the network is converging or not. It also enables you to inspect the training progress while the network is training.
Inspecting the training progress
If you leave the default settings, you can inspect the progress of the
training in real time, by using the script show_progress.py
. This script
is located in the folder utils
. You can get all supported command line arguments
with this command: python show_progress.py -h
. Normally you will want to start
the program like this: python show_progress.py
. It will open a TK window.
In case the program complains that it is not able to find TK related libraries,
you will need to install them.
Another approach is that you can use ChainerUI
, execute following commands to setup ChainerUI
:
chainerui db create
chainerui db upgrade
Create a project using the following command from the project directory:
chainerui project create -d ./ -n see-ocr
To check progress start server:
chainerui server
Creating an animation of plotted train steps
The training script contains a little helper that applies the current state of the model to an image and saves the result of this application for each iteration (or the way you configure it).
You can use the script create_video.py
to create an animation out of these images.
In order to use the script, you will need to install ffmpeg (and have the ffmpeg
command in your path)
and you will need to install imagemagick (and have the convert
command in your path).
You can then create a video with this command line call:
python create_video.py <path to directory with images> <path to destination video>
.
You can learn about further command line arguments with python create_video.py -h
.
Evaluation
You can evaluate all models (svhn/fsns/textrecognition) with the script evaluate.py
in the chainer
directory.
Usage
You will need a directory containing the following items:
- log_file of the training
- saved model
- network definition files that have been backed up by the training script
- set the gpu to use with
--gpu <id of gpu>
, the code does currently not work on CPU. - number of labels per timestep (typically max. 5 for SVHN and 21 for FSNS)
Evaluating a SVHN model
In order to evaluate a SVHN model, you will need to invoke the script like that:
python evaluate.py svhn <path to dir with specified items> <name of snapshot to evaluate> <path to ground truth file> <path to char map (e.g. svhn_char_map.json)> --target-shape <input shape for recogntion net (e.g. 50,50)> <number of labels per timestep>
Evaluating a FSNS model
In order to evaluate a FSNS model, you will need to invoke the script like that:
python evaluate.py fsns <path to dir with specified items> <name of snapshot to evaluate> <path to ground truth file> <path to char map (e.g. fsns_char_map.json)> <number of labels per timestep>
Evaluating a Text Recognition model
In order to evaluate a text recognition model, you will need to invoke the script like that:
python evaluate.py textrec <path to dir with specified items> <name of snapshot to evaluate> <path to ground truth file> <path to char map (e.g. ctc_char_map.json)> 23
Citation
If you find this code useful, please cite our paper:
@inproceedings{bartz2018see,
title={SEE: towards semi-supervised end-to-end scene text recognition},
author={Bartz, Christian and Yang, Haojin and Meinel, Christoph},
booktitle={Thirty-Second AAAI Conference on Artificial Intelligence},
year={2018}
}
Notes
If there is anything totally unclear, or not working, please feel free to file an issue. If you did anything with the code, feel free to file a PR.