This is the code for our paper:
- LSRFormer: Efficient Transformer Supply Convolutional Neural Networks with Global Information for Aerial Image Segmentation in IEEE Transactions on Geoscience and Remote Sensing 2024.
We repeated the experiment with 5 different random seeds. The average and best results of the 5 repetitions are as follows:
| Method | Dataset | mIoU (Average) | mIoU (Best) |
|---|---|---|---|
| ConvLSR-Net | iSAID | 70.8±0.11 | 70.89 |
| ConvLSR-Net | Vaihingen | 84.56±0.06 | 84.64 |
| ConvLSR-Net | Potsdam | 87.80±0.08 | 87.91 |
| ConvLSR-Net | LoveDA | 54.77±0.08 | 54.86 |
Due to some random operations in the training stage, reproduced results (run once) may slightly different from the reported in paper.
Please follw the GeoSeg to preprocess the LoveDA, Potsdam and Vaihingen dataset.
Please follow the mmsegmentation to preprocess the iSAID dataset.
"-c" means the path of the config, use different config to train different models.
python train_supervision.py -c ./config/isaid/convlsrnet.pypython train_supervision_dp.py -c ./config/potsdam/convlsrnet.pypython train_supervision_dp.py -c ./config/vahingen/convlsrnet.pypython train_supervision_dp.py -c ./config/loveda/convlsrnet.pyiSAID
python test_isaid.py -c ./config/isaid/convlsrnet.py -o ~/fig_results/isaid/convlsrnet_isaid/ -t "d4"Vaihingen
python test_vaihingen.py -c ./config/vaihingen/convlsrnet.py -o ~/fig_results/convlsrnet_vaihingen/ --rgb -t "d4"Potsdam
python test_potsdam.py -c ./config/potsdam/convlsrnet.py -o ~/fig_results/convlsrnet_potsdam/ --rgb -t "d4"LoveDA (Online Testing)
My LoveDA results: LoveDA Test Results
python test_loveda.py -c ./config/loveda/convlsrnet.py -o ~/fig_results/convlsrnet_loveda --rgb -t "d4"If you find this project useful in your research, please consider citing our papers：
- R. Zhang, Q. Zhang and G. Zhang, "LSRFormer: Efficient Transformer Supply Convolutional Neural Networks With Global Information for Aerial Image Segmentation," in IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1-13, 2024, Art no. 5610713, doi: 10.1109/TGRS.2024.3366709.
@ARTICLE{10438484,
author={Zhang, Renhe and Zhang, Qian and Zhang, Guixu},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={LSRFormer: Efficient Transformer Supply Convolutional Neural Networks With Global Information for Aerial Image Segmentation},
year={2024},
volume={62},
number={},
pages={1-13},
doi={10.1109/TGRS.2024.3366709}}If you encounter any problems while running the code, feel free to contact me via [email protected]. Thank you!
Our training scripts comes from GeoSeg. Thanks for the author's open-sourcing code.