Regularization of Building Boundaries in Satellite and Aerial Images

This repository contains the implementation for our publication "Machine-learned regularization and polygonization of building segmentation masks", ICPR 2021. If you use this implementation please cite the following publication:

@inproceedings{zorzi2021machine,
  title={Machine-learned regularization and polygonization of building segmentation masks},
  author={Zorzi, Stefano and Bittner, Ksenia and Fraundorfer, Friedrich},
  booktitle={2020 25th International Conference on Pattern Recognition (ICPR)},
  pages={3098--3105},
  year={2021},
  organization={IEEE}
}

and

@inproceedings{zorzi2019regularization,
  title={Regularization of building boundaries in satellite images using adversarial and regularized losses},
  author={Zorzi, Stefano and Fraundorfer, Friedrich},
  booktitle={IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium},
  pages={5140--5143},
  year={2019},
  organization={IEEE}
}

Explanatory video of the approach:

Dependencies

• cuda 10.2
• pytorch >= 1.3
• opencv
• gdal

Running the implementation

After installing all of the required dependencies above you can download the pretrained weights from here.

Unzip the archive and place saved_models_gan folder in the main projectRegularization directory.

Please note that the polygonization step is not yet available!

Evaluation

Modify variables.py accordingly, then run the prediction issuing the command

python regularize.py

Training

Modify variables.py accordingly, then run the training issuing the command

python train_gan_net.py