All python scripts needed for benchmarking Document Analysis algorithms are contained in this repository.

[Python 3.4.4] (https://www.python.org/downloads/) or newer

• Install required modules:

pip install pyyaml

• Open this directory (containing the setup.py) in a command line and type

python setup.py install

Database contains scripts to crawl your hard disk and reduce the number of specific file types. See:

lib\database\cdb.py --help

Example:

lib\database\cdb.py --outfile C:/temp/db.txt --copyto C:/temp/reduced C:/temp 5

This command locates all images in C:/temp, reduces the set to 5 images and copies these images to C:/temp/reduced. Then, C:/temp/db.txt is created which contains the paths to the original images of the reduced set.

Call this function from python using:

import database
rf = database.cdb.index_and_reduce_database("C:/temp", 13)
print("\n".join(rf))

Another script split.py splits a folder into train, evaluation and test set with user set proportions:

lib\database\split.py C:/temp C:/temp/split --train 0.2 --eval 0.1

This command copies 20% of the images contained in C:/temp into C:/temp/split/train, 10% into C:/temp/split/eval and all other images into C:/temp/split/test. You can set --eval 0 if you don't need an evaluation set.

Writer contains scripts to analyze XML files in the mets format. It extracts the and his corresponding pages out of the XML files and copies it to a new directory

First you have to create a database with all the xml files using the database crawler:

lib\database\cdb.py --outfile c:\tmp\db.txt --ext xml c:\tmp\myDatabase 1000000

From these xml files the writers are than extracted using

lib\writer\analyze.py --infile c:\tmp\db.txt --outfile c:\tmp\writers.txt --dumpfile c:\tmp\writer.pkl

and then the dataset is created using

lib\writer\create.py --infile c:\tmp\writer.pkl --outdir c:\tmp\writer-database \\
--minnum 0 --maxnum -1 --mapfile c:\tmp\writer-database\authorName-to-id-mapping.txt

minnum is the minimal number of pages, and maxnum is the maximal number of pages (if -1 than all pages are taken from this writer)

• CVL