delair-ai / DISIR

Deep Image Segmentation with Interactive Refinement

Presentation

This repository contains the code of DISIR: Deep Image Segmentation with Interactive Refinement. In a nutshell, it consists in neural networks trained to perform semantic segmentation with human guidance. You may refer to our paper for detailed explanations.

This repository is divided into two parts:

• train which contains the training code of the networks (README)
• qgs_plugin which contains the code of the QGIS plugin used to perform the interactive segmentation (README)

Install Python dependencies

conda create -n disir python=3.7 rtree gdal=2.4 opencv scipy shapely -c 'conda-forge' 
conda activate disir
pip install -r requirements.txt

To use

Please note that this repository has been tested on Ubuntu 18.4, QGIS 3.8 and python 3.7 only.

1. Download a segmentation dataset such as ISPRS Potsdam or INRIA dataset.
2. Prepare this dataset according to Dataset preprocessing in train/README.md.
3. Train a model and convert it to a torch script still following train/README.md.
4. Install the QGIS plugin following qgs_plugin/README.md.
5. Follow How to start in qgs_plugin/README.md and start segmenting your data !

References

If you use this work for your projects, please take the time to cite our ISPRS Congress conference paper:

@Article{isprs-annals-V-2-2020-877-2020,
AUTHOR = {Lenczner, G. and Le Saux, B. and Luminari, N. and Chan-Hon-Tong, A. and Le Besnerais, G.},
TITLE = {DISIR: DEEP IMAGE SEGMENTATION WITH INTERACTIVE REFINEMENT},
JOURNAL = {ISPRS Annals of Photogrammetry, Remote Sensing and Spatial Information Sciences},
VOLUME = {V-2-2020},
YEAR = {2020},
PAGES = {877--884},
URL = {https://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/V-2-2020/877/2020/},
DOI = {10.5194/isprs-annals-V-2-2020-877-2020}
}

Licence

Code is released under the MIT license for non-commercial and research purposes only. For commercial purposes, please contact the authors.

See LICENSE for more details.

Authors

See AUTHORS.md

Acknowledgements

This work has been jointly conducted at Delair and ONERA-DTIS.