ocrd_publaynet

convert PubLayNet data into METS/PAGE-XML

Introduction

This offers OCR-D compliant (i.e. METS-XML/PAGE-XML based) conversion for PubLayNet or similar, MS-COCO-based, ground-truth data.

Installation

System packages

Install GNU make and wget if you wish to use the Makefile.

# on Debian / Ubuntu:
sudo apt install make wget

Install Python3 regardless:

# on Debian / Ubuntu:
sudo apt install python3 python3-pip python3-venv

Equivalently:

# on Debian / Ubuntu:
sudo make deps-ubuntu

Python packages

It is strongly recommended to use venv. You can create and install a virtual environment of your own (which the Makefile will re-use when activated), or have the Makefile do that for you.

pip install -r requirements.txt
pip install .

Equivalently:

make install

Usage

command-line interface ocrd-import-mscoco

Once installed, the following executable becomes available:

Usage: ocrd-import-mscoco [OPTIONS] COCOFILE DIRECTORY

  Convert MS-COCO JSON to METS/PAGE XML files.

  Load JSON ``cocofile`` (in MS-COCO format) and chdir to ``directory``
  (which it refers to).

  Start a METS file mets.xml with references to the image files (under
  fileGrp ``OCR-D-IMG``) and their corresponding PAGE-XML annotations (under
  fileGrp ``OCR-D-GT-SEG-BLOCK``), as parsed from ``cocofile`` and written
  using the same basename.

Options:
  --help  Show this message and exit.

apply on PubLayNet

To apply on the validation subsection:

ocrd-import-mscoco publaynet/val.json publaynet/val

This will create a METS publaynet/val/mets.xml and PAGE files publaynet/val/*.xml for all image files.

To apply on the training subsection:

ocrd-import-mscoco publaynet/train.json publaynet/train

This will create a METS publaynet/train/mets.xml and PAGE files publaynet/train/*.xml for all image files.

Equivalently (including download/extraction if necessary):

make convert

Note: PubLayNet itself requires approximately 103 GB of disk space. If you already have it (elsewhere), but still wish to use the Makefile to convert the files, make sure to symlink it here, so it does not get downloaded twice: ln -s your/path/to/publaynet publaynet

Note: PubLayNet's train.json is 1.6 GB on disk and takes about 10 GB in (resident!) memory to load. Any incremental/stream-based method would be magnitudes slower than plain json.load(). Also, MS-COCO cannot be split because it basically defines a (humongous) annotations dict with pointers to a (large) images dict – sequentially. Another problem is that we cannot parallelize this, since everything needs to be in one final METS file. So this may take days. Just grin and bear it!

all Makefile targets

Rules to install ocrd-import-mscoco, and to use it on
PubLayNet (by downloading, extracting and converting).

Targets:
	help: this message
	deps-ubuntu: install system dependencies for Ubuntu
	all: alias for `install download convert`
	install: alias for `pip install .`
	download: alias for `publaynet.tar.gz`
	convert: alias for `publaynet/val/mets.xml publaynet/train/mets.xml`
	uninstall: alias for `pip uninstall ocrd_publaynet`
	clean-xml: remove results of conversion (METS and PAGE files in `publaynet`)
	clean: remove `publaynet` altogether

Variables:
	VIRTUAL_ENV: absolute path to (re-)use for the virtual environment
	PYTHON: name of the Python binary
	PIP: name of the Python packaging binary