This repository is a submodule of our paper "An Informative Path Planning Framework for Active Learning in UAV-based Semantic Mapping". The repository provides Pytorch Lightning implementations to train and evaluate our proposed general Bayesian ERFNet framework for semantic segmentation quantifying per-pixel model uncertainty using ensembles and Monte-Carlo dropout. The paper can be found here. If you found this work useful for your own research, feel free to cite it.
@article{ruckin2023informativeframework,
title={{An Informative Path Planning Framework for Active Learning in UAV-based Semantic Mapping}},
author={R{\"u}ckin, Julius and Magistri, Federico and Stachniss, Cyrill and Popovi{\'c}, Marija},
journal={arXiv preprint arXiv:2302.03347},
year={2023},
}
Our Bayesian ERFNet architecture for probabilistic semantic segmentation. We extend the network of Romera et al. with Monte-Carlo dropout (orange layers) to predict model uncertainty. Our network takes as input RGB (left) and outputs semantic labels (second from right) and pixel-wise uncertainty (first from right).
pip3 install -r requirements.txtRequires docker and docker-compose.
First, build the pipeline:
docker-compose buildTo start the training pipeline and tensorboard:
docker-compose upIn general, we follow the Python PEP 8 style guidelines. Please install black to format your python code properly. To run the black code formatter, use the following command:
black -l 120 path/to/python/module/or/package/
To optimize and clean up your imports, feel free to have a look at this solution for PyCharm.
Julius Rückin, [email protected], Ph.D. student at PhenoRob - University of Bonn
We would like to thank Jan Weyler for providing a PyTorch Lightning implementation of ERFNet. Our Bayesian-ERFNet implementation builds upon Jan's ERFNet implementation.
This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2070 – 390732324. Authors are with the Cluster of Excellence PhenoRob, Institute of Geodesy and Geoinformation, University of Bonn.