Pseudo Visible Feature Fine-Grained Fusion for Thermal Object Detection (CVPR-25)

Environmental Requirements

We utilize YOLOX, implemented via the MMDetection for environment setup instructions.

Our development environment includes the following dependencies:

python==3.9
torch==1.12.1+cu116
torchvision==0.13.1+cu116
mmcv-full==1.7.2
mmdet==2.26.0

Additionally, install the official Mamba library by following the instructions in the hustvl/Vim repository. After installation, replace the mamba_simpy.py file in the installation directory with the version available in the mamba block directory of the Pan-Mamba repository.

About the Code

This repository contains only the modifications made to the MMDetection codebase. For example:

• Add the code in mmdetection/mmdet/datasets/FLIR.py to your MMDetection.
• Ensure all newly added classes are registered in __init__.py.

Dataset and Models

• Datasets and model checkpoints can be downloaded from this cloud link, with the extraction code: PFGF.
• Download the Pearl-GAN pretrained weights from https://github.com/FuyaLuo/PearlGAN/. Place them into configs/graphmamba/pearlgan_ckpt/FLIR_NTIR2DC/.

Inference

To evaluate the FLIR dataset, run the following command:

python tools/test.py configs/graphmamba/yolox_l_tirgraphmamba_1x8_200e_FLIR_r.py work_dirs/flir.pth --eval mAP

Training

To train the model on the FLIR dataset, use the command:

python tools/train.py configs/graphmamba/yolox_l_tirgraphmamba_1x8_200e_FLIR_r.py

Acknowledgement

This project is based on mmdetection, DATFF, Pan-Mamba, Cas-Gnn. Thanks for their wonderful works.

Citation

If you find our PFGF framework useful, please consider citing our paper:

@inproceedings{li2025pseudo,
  title={Pseudo Visible Feature Fine-Grained Fusion for Thermal Object Detection},
  author={Li, Ting and Ye, Mao and Wu, Tianwen and Li, Nianxin and Li, Shuaifeng and Tang, Song and Ji, Luping},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={6710--6719},
  year={2025}
}