Semantic Segmentation of Radargrams via Unsupervised Random Walks and User-Guided Label Propagation

This is the repository of the work An Approach to Semantic Segmentation of Radar Sounder Data Based on Unsupervised Random Walks and User-Guided Label Propagation. The project can be easily replicated using the associated Docker container. To create and launch the container use:

bash launch_docker.sh <name> <tag>

and choose a name and a tag for the container. Within the running container, you can use:

python train.py <args>

to launch the unsupervised training step described in the paper above. Be sure to set all the paths to datasets and create the datasets in needed. Mind that this is a developer container, hence one would likely modify part of the source code to use a new dataset.

Part of the tests and plots used for the paper can be found within the \test\ folder. One would slightly modify the scripts to obtain a personalized script for performing inference with a trained model.

For comments, clarifications and issues, please write to jordy.dalcorso@unitn.it

For citations, use the following:

@INPROCEEDINGS{10641860,
  author={Corso, Jordy Dal and Bruzzone, Lorenzo},
  booktitle={IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium}, 
  title={Radargrams as Sequences: A Method for The Semantic Segmentation of Radar Sounder Data}, 
  year={2024},
  volume={},
  number={},
  pages={8179-8183},
  keywords={Representation learning;Radar remote sensing;Visualization;Semantic segmentation;Semantics;Object segmentation;Manuals;Semantic segmentation;Radar sounder;Sequence;Label propagation;MCoRDS},
  doi={10.1109/IGARSS53475.2024.10641860}}

For an earlier work addressing horizontal correlation within radargrams see also Radargrams as Sequences: A Method for The Semantic Segmentation of Radar Sounder Data.