bernet.23.icdar

Experimental protocol and results for the paper "Linear Object Detection in Document Images using Multiple Object Tracking" accepted at ICDAR 2023 by Bernet et al.

✨ Resources

📝 Paper	🖼️ Poster	🎥 Teaser	👀 Online Demo

---

Scribo Line Detector

Examples of possible applications of the line detector

Old directory (cropped): Separator detection for document rotation.

Original	Vectorial line detection	Rotation

Staff lines: Staff lines line detection for music score recognition.

Original	Staff lines line detection

Old map (cropped): Line detection for line removal of geographical lines.

Original	Line detection	Removed geographical lines

All line detection outputs are reproductible in the demo/demo.ipynb notebook.

---

Python API

The line detector is available on the PyPI server and can be simply installed with pip by executing the following command:

pip install pylena

A complete documentation and examples are available here. Call examples are present in the demo notebook demo/demo.ipynb.

C++ API

The line detector is included in the Pylene library.

A complete documentation on the installation and examples are available here. Call example is present in line_detectors/pylene-ours/main.cc.

Datasets

Datasets are available here :

Set the path in the file lsd_path_dataset.py

Evaluation

The line detection evaluation is available in the evaluation folder. Evaluation is performed using Python.

There are two different evaluation:

• vectorial evaluation: the line detection is evaluated using the vectorial representation of the lines.
• pixel evaluation: the line detection is evaluated as a segmentation task.

Benchmarks

Requirements

python -m venv pylene_line_detection_venv
source pylene_line_detection_venv/bin/activate
pip install -r requirements.txt

Line detectors C++ builds

The pylene library uses the conan (version >= 2.0) package manager to install dependencies. You can install it using pip:

pip install conan

Opencv C++ is required. Due to a conflict with pylene and opencv on libpng version using conan, we recommend to install using your package manager.

sudo apt-get install libopencv-dev

Build line detectors

cd line_detectors/build
conan remote add lrde-public https://artifactory.lrde.epita.fr/artifactory/api/conan/lrde-public
conan install .. --output-folder=. -s build_type=Release -s compiler.cppstd=20 -s compiler.libcxx=libstdc++11 --build missing --build freeimage* --build openjpeg*
cmake .. -DCMAKE_TOOLCHAIN_FILE=./conan_toolchain.cmake
make -j

Demo interactive

The notebook demo/demo_interactive.ipynb contains some code to test the C++ builded line detectors. Using this notebook, you will be able to compare the line detectors on your own images. You can add your own image in the demo/image folder. Image must be 8b-grayscale.

Run the benchmarks

Use respecting the order the following notebooks:

• benchmark_run.ipynb: to run the benchmark
• benchmark_score.ipynb: to score the benchmark
• benchmark_visualize.ipynb: to output the graphs

Citation

If you use or found this code, evaluation, and/or dataset useful, please cite:

@InProceedings{bernet.23.icdar,
  author    = {Philippe Bernet and Joseph Chazalon and Edwin Carlinet and Alexandre Bourquelot and Elodie Puybareau},
  title     = {Linear Object Detection in Document Images using Multiple Object Tracking},
  booktitle = {Proceedings of the 17th International Conference on Document Analysis and Recognition},
  year      = {2023},
  address   = {San José, California, USA},
  month     = {08},
  publisher = {Springer},
  note      = {To appear.},
}