Weakly Supervised Building Segmentation from Overhead Images

This is the code for our weakly supervised building segmentation paper, accepted at IGARSS 2019.

Paper: Weakly Supervised Building Segmentation From Aerial Images

Dataset

We use the disaster response dataset released as the Mapping Challenge. You should be able to get the data from this website. If you have any trouble acquiring the dataset, please contact us.

Setup

Create and activate the conda environment:

conda env create -f environment.yml
conda activate wbseg

How to Use

All settings are stored in wbseg/config.py. Edit that file to set the dataset path (ROOT_DIR), output directory (DIRECTORY), supervision mode, loss function, batch size, and other training options.

Train a model:

python -m wbseg.train

This trains the U-Net model and saves the trained weights, loss curves, and metrics to the directory specified by DIRECTORY in config.py.

Visualize results:

python -m wbseg.visualize_trained

This loads the trained model and saves prediction figures to the same output directory.

Supervision modes

Set SUPERVISION in wbseg/config.py to one of:

Value	Description
Gaussian	Dense masks derived from bounding boxes using a bivariate Gaussian (default)
Naive	All pixels inside bounding boxes set to foreground
GrabCut	OpenCV GrabCut applied within each bounding box
Full	Full supervision using ground-truth segmentation masks (upper bound)

Loss functions

Set LOSS_FN in wbseg/config.py to one of:

Value	Description
Proposed_OneSided	Proposed one-sided loss (default)
CE	Standard binary cross-entropy

Citation

If you find this paper or code helpful, please cite:

M. Usman Rafique, Nathan Jacobs, "Weakly Supervised Building Segmentation From Aerial Images",
In: IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2019.

People

Please feel free to contact us with any questions or comments.

M. Usman Rafique

Nathan Jacobs

License

The code is provided for academic purposes only without any guarantees.