FreDFT: Frequency Domain Fusion Transformer for Visible-Infrared Object Detection

This is an official PyTorch implementation for our FreDFT. ArXiv paper can be downloaded in FreDFT. TMM paper can be downloaded in FreDFT.

Our paper has been accepted by IEEE Transactions on Multimedia for publication as a regular paper. The final version of the paper will be updated soon.

Create a conda virtual environment and activate it.

1. conda create --name MOD python=3.9
2. conda activate MOD
3. pip install -r requirements.txt

Download these datasets and create a dataset folder to hold them.

1. FLIR dataset: FLIR
2. LLVIP dataset: Official Website or Baidu Netdisk
3. M3FD dataset: M3FD

Download our FreDFT weights and create a weights folder to hold them.

1. FLIR dataset: FreDFT_FLIR.pt
2. LLVIP dataset: FreDFT_LLVIP.pt
3. M3FD dataset: FreDFT_M3FD.pt

Dataset path, GPU, batch size, etc., need to be modified according to different situations.

python train.py

python test.py

If you find FreDFT helpful for your research, please consider citing our work.

@article{Wu2026,
  author={Wu, Wencong and Zhang, Xiuwei and Yin, Hanlin and Dai, Shun and Zhang, Hongxi and Zhang, Yanning},
  journal={IEEE Transactions on Multimedia}, 
  title={FreDFT: Frequency Domain Fusion Transformer for Visible-Infrared Object Detection}, 
  year={2026},
  pages={1-12},
  doi={10.1109/TMM.2026.3712630}
}

@article{Wu2025,
  title={FreDFT: Frequency Domain Fusion Transformer for Visible-Infrared Object Detection}, 
  author={Wencong Wu and Xiuwei Zhang and Hanlin Yin and Shun Dai and Hongxi Zhang and Yanning Zhang},
  journal={arXiv preprint arXiv:2511.10046},
  year={2025}
}