Awesome

Pytorch implementation of MA3E

1. Pre-training Dataset Preparation

Download the MillionAID dataset. The data structure should be as follows:

├── MillionAID
│   ├── train
│   │   ├── agriculture
│   │   │   ├── arable_land
│   │   │   │   ├── dry_field
│   │   │   │   │   ├── P0011232.jpg
│   │   │   │   │   ├── ...
│   │   │   │   ├── ...
│   │   │   ├── ...
│   │   ├── ...
│   ├── test
│   │   ├── xxx.jpg
│   │   ├── xxx.jpg
│   │   ├── ...
│   ├── ...

Since the testing set of the MillionAID dataset does not release labels, to generate random pseudo-labels for pre-training

python create_pretraining_labels.py

• Please modify the dataset path in line 6 and 7

2. Pre-training MA3E

To pre-train ViT-Base with multi-gpu distributed training, run the following on 1 nodes with 8 GPUs each:

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 main_pretrain.py \
    --batch_size 128 \
    --model ma3e_vit_base_patch16 \
    --norm_pix_loss \
    --rcrop \
    --crop_size 96 --nums_crop 1 --r_range 45 \
    --b_ratio 0.75 --r_ratio 0.75 \
    --epochs 800 \
    --warmup_epochs 40 \
    --blr 1.5e-4 --weight_decay 0.05 \
    --dataset millionaid
    --data_path ${Your Million-AID_DIR}

• Here the effective batch size is 128 (batch_size per gpu) * 1 (nodes) * 8 (gpus per node) = 1024. If memory or # gpus is limited, use --accum_iter to maintain the effective batch size, which is batch_size (per gpu) * nodes * 8 (gpus per node) * accum_iter.

• This repo is based on timm==0.3.2, for which a fix is needed to work with PyTorch 1.8.1+.

3. Scene Classification Dataset Preparation

Download scene classification datasets: NWPU-RESISC45, AID, and UC Merced. The data structures should be as follows:

├── NWPU-RESISC45
│   ├── airplane
│   │   ├── airplane_001.jpg
│   │   ├── airplane_002.jpg
│   │   ├── ...
│   ├── airport
│   │   ├── airport_001.jpg
│   │   ├── airport_002.jpg
│   │   ├── ...
│   ├── ...

├── AID
│   ├── Airport
│   │   ├── airport_1.jpg
│   │   ├── airport_2.jpg
│   │   ├── ...
│   ├── BareLand
│   │   ├── bareland_1.jpg
│   │   ├── bareland_1.jpg
│   │   ├── ...
│   ├── ...

├── UC Merced
│   ├── agricultural
│   │   ├── agricultural00.tif
│   │   ├── agricultural01.tif
│   │   ├── ...
│   ├── airplane
│   │   ├── airplane00.tif
│   │   ├── airplane01.tif
│   │   ├── ...
│   ├── ...

-For NWPU-RESISC45, randomly sample 20% of the images for each class as the training set and the remaining 80% as the testing set.

-For AID and UC Merced, these two ratios is all 50%.

4. Fine-tuning

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 main_finetune.py \
    --batch_size 64 \
    --model vit_base_patch16 \
    --finetune ${MA3E_PRETRAIN_CHKPT} \
    --epochs 200 \
    --blr 5e-4 --layer_decay 0.65 \
    --weight_decay 0.05 --drop_path 0.1 --reprob 0.25 --mixup 0.8 --cutmix 1.0 \
    --dist_eval --data_path ${Your DIR of NWPU-RESISC45 (2:8) or AID (5:5) or UC Merced (5:5) Dataset}

5. Linear probing

OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 main_linprobe.py \
    --batch_size 256 \
    --model vit_base_patch16 --cls_token \
    --finetune ${MA3E_PRETRAIN_CHKPT} \
    --epochs 100 \
    --blr 0.1 \
    --weight_decay 0.0 \
    --dist_eval --data_path ${Your DIR of NWPU-RESISC45 (2:8) or AID (5:5) or UC Merced (5:5) Dataset}

6. Other downstream tasks

For Rotated Object Detection and Semantic Segmentation. Please refer to Wang et al. repo.

Visualization of reconstruction

Acknowledgments

This repo is a modification on the MAE repo. Installation and preparation follow that repo.

Citation

Please cite this paper if it helps your research:

@inproceedings{li2025masked,
  title={Masked angle-aware autoencoder for remote sensing images},
  author={Li, Zhihao and Hou, Biao and Ma, Siteng and Wu, Zitong and Guo, Xianpeng and Ren, Bo and Jiao, Licheng},
  booktitle={European Conference on Computer Vision},
  pages={260--278},
  year={2025},
  organization={Springer}
}