Code for HED-UNet, a model for simultaneous semantic segmentation and edge detection.
This model was originally developed to detect calving front margins in Antarctica from Sentinel-1 SAR imagery.
As the original dataset isn't publicly available, this repository contains an adaption of the model for building footprint extraction on the Inria Aerial Image Labeling dataset. Here are some example results:
In order to use this for your project, you will need adapt either the get_dataloader function in train.py or the methods in data_loading.py.
If you find our code helpful and use it in your research, please use the following BibTeX entry.
@article{HEDUNet2021,
author={Heidler, Konrad and Mou, Lichao and Baumhoer, Celia and Dietz, Andreas and Zhu, Xiao Xiang},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={HED-UNet: Combined Segmentation and Edge Detection for Monitoring the Antarctic Coastline},
year={2021},
volume={},
number={},
pages={1-14},
doi={10.1109/TGRS.2021.3064606}
}