/faster-rcnn-graphics

Symbol detection in online handwritten graphics using Faster R-CNN

Primary LanguagePython

Symbol detection in online handwritten graphics using Faster R-CNN

faster-rcnn-graphics has been imported to ImageU. Future updates will only be done on that project

This repository contains the implementation of the models described in the paper "Symbol detection in online handwritten graphics using Faster R-CNN". A model is a Faster R-CNN network that takes an image of a handwritten graphic (flowchart or mathematical expression) as input and predicts the bounding box coordinates of the symbols that compose the graphic. The models are implemented using a fork of the the Tensorflow Object Detection API.

Symbol detection in flowchart

Symbol detetion in mathematical expression

Citing this work

In case you use this work, please consider citing:

@inproceedings{frankdas:2018,
  title={Symbol detection in online handwritten graphics using Faster R-CNN},
  author={Frank Julca-Aguilar and Nina Hirata},
  booktitle={13th IAPR International Workshop on Document Analysis Systems (DAS)},
  year={2018}
 }

Contents

  1. Installation
  2. Evaluating the models
  3. Training new models

Installation

  1. Clone the repository (with --recursive)
git clone --recursive https://github.com/vision-ime/faster-rcnn-graphics.git

The recursive option is necessary to download the fork version of the Tensorflow Object Detection API used in our experimentation.

  1. Follow the Tensorflow Object Detection API installation instructions to set up the API, which was cloned in the tf-models folder (the tf-models folder corresponds to the models folder described in the API installation instructions).

Evaluating the models

  1. Download the datasets. In the directory where you cloned this repository do:
./download_datasets.sh

The datasets will be saved in the datasets folder. Each dataset consist of Tensorflow's .record files, images of handwritten graphics, and xml metadata for each image. As described in the paper, the datasets were using the CROHME-2016 and flowcharts datasets.

  1. Download a model. Models can be download from: http://www.vision.ime.usp.br/~frank.aguilar/graphics/models/

For example, to download the model for symbol detection in flowcharts, trained with inception V2:

wget http://www.vision.ime.usp.br/~frank.aguilar/graphics/models/flowcharts/flowcharts_inceptionv2.tar.gzip

In order better to organize the different files, we can
save the model in the corresponding flowchart folder.

mkdir models/flowcharts/inceptionv2/trained
mv flowcharts_inceptionv2.tar.gzip models/flowcharts/inceptionv2/trained/
cd models/flowcharts/inceptionv2/trained
tar -xf flowcharts_inceptionv2.tar.gzip
  1. Execute the evaluation script. In the folder in which you cloned this work, to evaluate the model downloaded in step 2, you can do
python tf-models/research/object_detection/eval.py \
    --logtostderr \
    --pipeline_config_path=models/flowcharts/inceptionv2/pipeline.config \
    --checkpoint_dir=models/flowcharts/inceptionv2/trained \
    --eval_dir=models/flowcharts/inceptionv2/eval \ 
    --gpudev=0 \
    --run_once=True

The parameter gpudev indicates the GPU device that would be used to evaluate the model. A value -1 can be used to run over CPU. The rest of the parameters are defined as in the Object Detection API (here).

Training new models

New models can be trained using

python tf-models/research/object_detection/train.py \
--logtostderr \
--pipeline_config_path=models/math/inceptionv2/pipeline.config \
--train_dir=models/math/inceptionv2/new_trained \ 
--gpudev=0 &

As in the case of evaluation, the parameters are defined
here.