DocParser: Hierarchical Structure Parsing of Document Renderings

Codes for the system presented in "DocParser: Hierarchical Structure Parsing of Document Renderings"

paper

Installation and requirements

Tested for Ubuntu 18.04/20.04.

Use of a GPU significantly speeds up generation of detection outputs, but it is possible to run the inference demo code on CPU.

To setup via Anaconda, please follow these steps:

1. Install anaconda. Up-to-date instructions can be found at: https://docs.anaconda.com/anaconda/install/

2. Set up python 3.6 environment: conda create -n docparser python=3.6

3. Activate the environment: source activate docparser

4. Install all requirements: pip install -r requirements.txt

   • (for GPU-enabled installation: pip install -r requirements_gpu.txt)

5. Install Mask R-CNN library:

   • We used a slightly modified version of https://github.com/matterport/Mask_RCNN, though the original version should still be usable, possibly with minor adaptions.
   • Clone repository from https://github.com/j-rausch/Mask_RCNN
   • Change into mask rcnn directory
   • type python setup.py install

6. Install docparser:

   • Change into DocParser directory
   • type python setup.py develop

7. Prepare the datasets:

   • Download arxivdocs-target from https://github.com/DS3Lab/arXivDocs
   • To run the ICDAR demo, download the prepared files from: https://drive.google.com/file/d/1SdGTq80eUGqUJBA6kdVQBO9L6a_ijAcN/view?usp=sharing
   • Extract datasets to the DocParser subdirectory
     • (resulting in structure: DocParser/datasets).

8. Prepare the trained models:

   • Download from URL: https://drive.google.com/file/d/1Hi4-tg4Zmtx8zYiCg6IBi47R88PdmAW4/view?usp=sharing
   • Extract the pretrained models to the default_models subdirectory in DocParser/docparser/
     • (resulting in structure DocParser/docparser/default_models/).
   • For convenience, we include the COCO pre-trained weights from from https://github.com/matterport/Mask_RCNN/releases/download/v2.0/mask_rcnn_coco.h5 in the zip file

9. For running the ICDAR demo:

   • Please note that, in order to run the ICDAR 2013 evaluation script provided by the competition organizers, a Java installation is necessary. We used openjdk 11.0.7 2020-04-14 in our experiments.
   • If necessary, update permissions for the evaluation script (on linux systems): chmod a+x DocParser/docparser/utils/dataset-tools-20180206.jar

10. From the DocParser directory, execute: python demos/demo_inference.py plus one or more of the following command line arguments:

    • --page
    • --table
    • --icdar
    • e.g. python demos/demo_inferencey.py --page --table

Evaluations

arXivDocs

The results of our current system on arXivDocs-target is likely to perform better than the one evaluated in the last version of the paper, mostly due to further improvements to postprocessing.

ICDAR 2013, Table Structure Recognition

Updated Results. We corrected a read-out error on the outputs of the provided evaluation script for documents with multiple tables.

System	F1*	F1
DocParser Baselie	0.8443	0.8209
DocParser WS	0.8117	0.8056
DocParser WS+FT	0.9292	0.9292

(PDF-based system F1: 0.9221)

Credits

Parts of our code is based on: https://github.com/rafaelpadilla/Object-Detection-Metrics

https://github.com/matterport/Mask_RCNN

Reference

Rausch, J., Martinez, O., Bissig, F., Zhang, C., & Feuerriegel, S. (2019). DocParser: Hierarchical Structure Parsing of Document Renderings. http://arxiv.org/abs/1911.01702