...Moreover, we proposed a pipeline that harnesses the capabilities of powerful large vision-language models (VLMs) as image encoders, establishing new baselines for FS-RSI-SC on commonly used datasets under standard experimental settings.

Please follow the official setup of CLIP here: https://github.com/openai/CLIP Please follow the official setup of BLIP here: https://github.com/salesforce/BLIP

pytorch-template/
├── run_classifier_fromFeatures.sh - the main file to run classifier_fromFeatures.py
├── saveFeatures_subfolder_aid_nwpu_whu.py - extract features from images and save
├── classifier_fromFeatures.py
│
│
├── nwpu-resisc45/ - extracted features, class-wise split following the common setup
│   ├── train - train split of the common setup
│   ├── test - 
│   ├── val - 
│   
├── acc.txt - evaluation results

To reproduce the results in Table 4, please download datasets. The train-val-test split is following a common one in litearature, and the training set is not actually used for training models.

First extract feautres using saveFeatures_subfolder_aid_nwpu_whu.py, and then set the path in run_classifier_fromFeatures.sh, and run: bash run_classifier_fromFeatures.sh and results will be saved in the folder "acc.txt".

If you use this code for your research, please cite: