Exposure-aware Progressive Optimizationfor Infrared and Visible Image Fusion

Zhiwei Wang1,2 · Defeng He1 · Li Zhao3 · Xiaoqin Zhang1 · Yuxing Li2 · Edmund Y. Lam2

1 College of Information Engineering, Zhejiang University of Technology
2 Department of Electrical and Computer Engineering, The University of Hong Kong
3 College of Information Science and Technology, Zhejiang Shuren University
Paper Code 🗂 IVOE
EPOFusion architecture: exposure-aware guidance, iterative decoding fusion, and multi-scale context fusion
The architecture of EPOFusion.

Infrared–Visible Over-Exposure benchmark (IVOE)

IVOE sequence FLIR_05511
IVOE sequence FLIR_03901
IVOE sequence FLIR_09399

Each loop runs consecutive frames as four panels — visible, infrared, detection GT, segmentation GT. The annotations sit on targets the visible image has lost to saturation while the co-registered infrared still resolves them.

447real over-exposed pairs
3,722detection boxes, 3 classes
51.7%frames with a severely blown-out target
74.1%boxes where IR carries more detail than RGB
person 1,993 bicycle 186 car 1,640 masks 447 · 3 classes

Citation

@article{wang2026epofusion,
  title={EPOFusion: Exposure-aware Progressive Optimization Method for Infrared and Visible Image Fusion},
  author={Wang, Zhiwei and He, Defeng and Zhao, Li and Zhang, Xiaoqin and Li, Yuxing and Lam, Edmund Y},
  journal={arXiv preprint arXiv:2603.16130},
  year={2026}
}