IVOE: Infrared–Visible Over-Exposure Dataset

Paper · Project Page · Code

IVOE (Infrared–Visible Over-Exposure) is a real-world dataset of 447 aligned infrared–visible image pairs for image fusion under strong daylight, glare, and headlights. It provides object-detection labels and pixel-level segmentation masks for targets obscured in visible imagery but preserved in infrared.

Preview

Visible · Infrared · Detection ground truth · Segmentation ground truth

Dataset Structure

IVOE is released as a test-only dataset in IVOE.zip. After extraction, files belonging to the same sample share a common stem.

IVOE.zip
└── test/
    ├── vi/                       # Visible RGB images
    │   └── <sample_id>.png
    ├── ir/                       # Infrared images
    │   └── <sample_id>.png
    ├── Segmentation_labels/      # Pixel-level label maps
    │   └── <sample_id>.png
    ├── Segmentation_visualize/   # Colorized label previews
    │   └── <sample_id>.png
    └── Detect/                   # YOLO detection annotations
        └── <sample_id>.txt

All modalities in a sample share the same resolution. Visible images are RGB; infrared PNGs may be grayscale or RGB.

Annotations

Detection ID 0 1 2 Category person bicycle car

Pixel Value 0 1 2 3 Category background person bicycle car

Object Detection

Each test/Detect/<sample_id>.txt file uses normalized YOLO format:

<class_id> <x_center> <y_center> <width> <height>

Coordinates and box dimensions are normalized to [0, 1] relative to image width and height.

🔴 person · 🟢 bicycle · 🔵 car

Semantic Segmentation

Segmentation labels are single-channel, 8-bit PNG images. Foreground IDs are one greater than detection IDs because 0 is reserved for background.

Citation

If you use IVOE in your research, please cite the EPOFusion paper:

@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}
}