[EAAI 2025] Feature transformation and statistical calibration for cross-domain few-shot classification

This is the implementation of our EAAI 2025 paper: Feature transformation and statistical calibration for cross-domain few-shot classification.

Datasets

Source Domain: NWPU45

Target Domains: AID, PatternNet, RSI-CB128, GID, Place25

Please download the datasets from the official websites.

Data Preparation

After downloading the datasets, remove overlapping classes following Table A.6 in our paper to ensure proper cross-domain few-shot learning setup.

Meta-Training

We provide pre-trained backbones trained on the NWPU45 source domain for ResNet-18 backbone and Conv-4 backbone. Download the models from the models/ folder or train your own backbones.

Meta-Testing

For FETAC with Conv-4 backbone on AID dataset (1-shot):

python test_target.py \
    --data.dataset_dir /home/fetac/data/l_AID \
    --data.shot_num 1 \
    --data.total_file 4420 \
    --data.test_start 13 \
    --test.use_lightFiLM True \
    --test.use_rsa True \
    --test.rectify lla \
    --model.name conv4 \
    --model.weight_path models/42conv4.pth.tar

To run experiments with different settings, modify the parameters in config.py. The key parameters are described below:

data args:

• data.dataset_dir : the path of the test dataset
• data.shot_num : the number of shot
• data.total_file : the number of the test dataset images
• data.test_start : the number of the test dataset classes

test args:

• test.use_lightFiLM : whether use lightFiLM

• test.use_rsa : whether use the rsa

• test.rectify : use what types of rectification

  For FETAC:

  --test.use_lightFiLM True
  --test.use_rsa True
  --test.rectify lla

  For TSA+LLA:

  --test.use_lightFiLM False
  --test.use_rsa True
  --test.rectify lla

model args:

• model.name : use what types of backbones
• model.weight_path : the path of pretrained backbone

Acknowledgments

Our code builds upon URL and STF repositories.

Citation

If you find FETAC useful in your research, please cite:

@article{liu2025feature,
  title={Feature transformation and statistical calibration for cross-domain few-shot classification},
  author={Liu, Jiafan and Deng, Jin and Cui, Jinrong and Luo, Wei},
  journal={Engineering Applications of Artificial Intelligence},
  volume={157},
  pages={111181},
  year={2025},
  publisher={Elsevier}
}