Awesome

Cross-Scale MAE (NeurIPS 2023)

Authors: Maofeng Tang · Andrei Cozma · Konstantinos Georgiou · Hairong Qi

This is a PyTorch implementation of our NeurIPS 2023 paper: Cross-Scale MAE: A Tale of Multi-Scale Exploitation in Remote Sensing

Pretraining

The run the pretraining on a single node, you can use use the the train.sh. Make sure you modify its contents to match your environment. Alternatively, you can use the main_pretrain.py directly.

For multi-gpu training, use the train_distributed.sh instead.

Finetuning & Linear Probing

To run finetuning on a single node, you can use the finetune.sh. Make sure you modify its contents to match your environment. Alternatively, you can use the main_finetune.py directly.

For linear probing, use the linprobe.sh and main_linprobe.py instead.

Model Weights

Pretrained weights for the models used in the paper can be found here:

You would need to download the weights and place them in a folder named weights in the root of the repository.

Acknowledgements

Code from this repository is inspired from the following repositories:

• SatMAE (NeurIPS 2022)
• Masked Autoencoders (MAE)

Citation

If you found our project helpful, please cite our paper:

@inproceedings{tang2023cross,
  title={Cross-Scale MAE: A Tale of Multiscale Exploitation in Remote Sensing},
  author={Tang, Maofeng and Cozma, Andrei Liviu and Georgiou, Konstantinos and Qi, Hairong},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}

License

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.