FreqSal

Deep Fourier-embedded Network for RGB and Thermal Salient Object Detection [IEEE][arXiv]

• Nov. 4, 2025
  The paper has been accepted by TCSVT2025
• Apr. 29, 2024
  The paper is undergoing peer review. The code will be released upon acceptance of the paper.
• 
• In this project, we proposed the deep Fourier-embedded network, namely FreqSal, a purely Fourier-based model aimed at solving the high-resolution bimodal inputs and feature fusion while minimizing memory consumption of GPU, outperforming existing state-of-the-art bimodal salient object detection (SOD) models on four RGB-T, five RGB-D, and one RGB-D-T SOD benchmark datasets. To the best of our knowledge, this is the first Fourier-based supervised model in a series of SOD tasks.
• Please cite our paper if you find it useful for your research.

@article{lyu2025deep,
  title={Deep Fourier-embedded Network for RGB and Thermal Salient Object Detection},
  author={Lyu, Pengfei and Yu, Xiaosheng and Yeung, Pak-Hei and Wu, Chengdong and Rajapakse, Jagath C},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2025},
  publisher={IEEE}
}

Requirements

List of prerequisites or required libraries for the project to run:

• Pytorch 2.0.0
• Cuda 11.8
• Python 3.8 or higher
• tensorboardX
• opencv-python
• timm == 0.5.4
• thop
• numpy

Datasets

• We conducted experiments to evaluate our model on the VT821, VT1000, VT5000, and VI-RGBT1500 datasets for the RGB-T SOD task, and on the NLPR, NJUD, DUT-RGBD, SIP, and STERE datasets for the RGB-D SOD task. Please click for the corresponding dataset.

Pre-trained Weights of FreqSal

Resolution | Backbone | Tpye | weights ---- | ----- | ------ | ------ 384 x 384 | CDFFormer-m36 | RGB-T | Link 512 x 512 | CDFFormer-m36 | RGB-T | Link 384 x 384 | CDFFormer-m36 | RGB-D | Link

Results

• The RGB-T and RGB-D results of our model can be found at link.
• 
• 

Evaluation Metrics Toolbox

• The Evaluation Metrics Toolbox is available here: link.

Acknowledgements

• Thanks to all the seniors and projects (e.g., DFFormer, FFL, UHDFour, MGAI, and SwinNet).

Contact Us

If you have any questions, please contact us (lvpengfei1995@163.com).