/dpn_rs

Implementations for "An Advanced Dirichlet Prior Network for Out-of-Distribution Detection in Remote Sensing"

Primary LanguagePython

An Advanced Dirichlet Prior Network for Out-of-Distribution Detection in Remote Sensing

Out-of-Distribution detection for unseen classes, changing sensors, and changing locations.

The proposed method uses in-distribution and out-of-distribution training data in order to learn a gap between (familiar) in-distribution data and (partially unknown) out-of-distribution data.

To run the code please download the datasets and put them in datasets folder.
Then cd to Experiments directory and run the code for the UCM dataset by using following commands:
python experiment_sensor_shift.py --data ucm --approach dpn_rs --seed 42
python experiment_class_shift.py --data ucm --approach dpn_rs --seed 42
data, approach and seed can be adapted.

Citation

If you find this code useful, please consider citing:

@article{gawlikowski2022Andvanced,
  title={An Advanced Dirichlet Prior Network for Out-of-distribution Detection in Remote Sensing},
  author={Gawlikowski, Jakob and Saha, Sudipan and Kruspe, Anna and Zhu, Xiao Xiang},
  journal={IEEE Transactions in Geoscience and Remote Sensing}
  year={2022},
  publisher={IEEE}
}