CLIP-Guided Source-Free Object Detection in Aerial Images

💥 News

This paper is accepted by IGARSS2024 as a oral representation and is accessible here.

🎨 Overview

Overview of the proposed SFOD framework.

🎮 Getting Started

1. Install Environment

conda create --name sfod python=3.8
conda activate sfod

pip install torch===1.7.1+cu110 torchvision===0.8.2+cu110 torchaudio===0.7.2 -f https://download.pytorch.org/whl/torch_stable.html --no-cache

git clone https://github.com/Lans1ng/SFOD-RS.git
cd SFOD-RS
pip install -r requirements.txt

2. Prepare Dataset

• Download the DIOR dataset (Google Drive or BaiduNetDisk).

• Download the corruption images with cloud (DIOR-Cloudy) from Google Drive.

Note : For more detail about other corruptions and original cloudy image, please refer to DOTA-C.

• Organize the dataset as follow:

  dataset/
  ├──DIOR/
  │   ├── Annotations 
  │   ├── JPEGImages
  │   ├── ImageSets
  │   ├── Corruption/
  │   │   ├── JPEGImages-brigtness
  │   │   ├── JPEGImages-cloudy
  │   │   ├── JPEGImages-contrast
  │   │   ├── ...
  │   │   ├── JPEGImages-zoom_blur  

  • JPEGImages contains all images in DIOR dataset.
  • ImageSets contains all train/val/test splits in DIOR dataset.
  • Corruption contains all kinds of corrupted images on DIOR dataset.

3. Download Checkpoints

Before training，please download the pretrained Oriented-RCNN model weights baseline.pth trained from DIOR training set to baseline fold.

4. Training

Take DIOR-Cloudy dataset as examples:

python train.py configs/unbiased_teacher/sfod/unbiased_teacher_oriented_rcnn_selftraining_cga.py --cfg-options corrupt="cloudy"

Note: If you want to retrain the Oriented-RCNN baseline model, please refer to mmrotate.

5. Testing

Take DIOR-Cloudy dataset as examples:

python test.py configs/unbiased_teacher/sfod/unbiased_teacher_oriented_rcnn_selftraining_cga.py work_dirs/unbiased_teacher_oriented_rcnn_selftraining_cga/latest.pth --eval mAP --cfg-options corrupt="cloudy"

💡 Acknowledgement

• DOTA-C
• mmrotate
• SSOD
• SOOD
• VFA
• SoftTeacher

🖊️ Citation

If you find this project useful in your research, please consider cite:

@INPROCEEDINGS{10642376,
  author={Liu, Nanqing and Xu, Xun and Su, Yongyi and Liu, Chengxin and Gong, Peiliang and Li, Heng-Chao},
  booktitle={IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium}, 
  title={Clip-Guided Source-Free Object Detection in Aerial Images}, 
  year={2024},
  volume={},
  number={},
  pages={7964-7967},
  doi={10.1109/IGARSS53475.2024.10642376}}