This is the repo for our wekaly supervised building segmentation, accepted at IGARSS 2019.
We use the disaster response dataset, released as the mapping challenege:. You should be able to get the data form this website. If you have any trouble in acquiring the dataset, please contact us.
There are several files, comments at the top of each file explain the purpose.
All the settings are stored in config.py. It is expected that you will forst train a model and then you can run visualization code.
To train a model, specify dataset path, training settings (level of supervision, loss function, batch size etc) in config.py. You will also specify a directory in which the trained model will be saved. Once the training finishes, a log file and loss curves will be saved in that folder.
After training, you can run the visualize_trained.py file, which will load the trained model and save some visual results in that folder.
If you find this paper or code helpful, please cite this paper:
M. Usman Rafique, Nathan Jacobs, "Weakly Supervised Building Segmentation From Aerial Images", In: IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2019.
Please feel free to contact us for any question or comment.
The code is provided for academic purposes only without any guarantees.