EvTradeNet (ICCVw 2025)

Implementation of Monochromatic Event Guided Image Deblurring with Event-triggering-aware Decomposition (ICCV 2025 MIPI Workshop)

Environment

conda create -n evtradenet python=3.10
conda activate evtradenet
pip install torch torchvision
pip install einops numpy opencv-python scikit-image scipy tensorboardX tqdm lpips "huggingface_hub[cli]"

Data Preparation

Real-world evaluation (EvRGB-Deblur)

The EvRGB-Deblur dataset is hosted on HuggingFace. You can use download.py to download dataset.

• Download link: EvRGB-Deblur Dataset

Train

python train.py --TrainImgPath `Path of Train Image` --TrainEvePath `Path of Events` --TrainGTPath `Path of Ground Truth Image` --TestImgPath `Path of Test Image` --TestEvePath `Path of Test Events` --TestGTPath `Path of Test Ground Truth Image`

Evaluation

Pretrained weights for both the luminance and color stages are also available on Google Drive.

• Download link: EvTradeNet Pretrained Weights
• Place model_LuminDeblur_best.pth and model_ImgColorNet_best.pth into the pretrained/ directory.

python test.py \
  --ckp_ld pretrained/model_LuminDeblur.pth \
  --ckp_cc pretrained/model_ImgColorNet.pth \
  --RealImgPath EvRGB_Deblur/blur \
  --RealEvePath EvRGB_Deblur/event \
  --RealGTPath EvRGB_Deblur/sharp \
  --RealSavePath results_evtradenet

The script writes restored frames to RealSavePath

References

@InProceedings{TengEvTradeNet2025,
    author    = {Teng, Minggui and Li, Boyu and Yang, Yixin and Zhou, Chu and Chen, Yan and Ren, Jimmy S. and Shi, Boxin},
    title     = {Monochromatic Event Guided Image Deblurring with Event-triggering-aware Decomposition},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops},
    year      = {2025},
    pages     = {3876-3885}
}