sndnshr / DeepFuse

Pytorch implementation of DeepFuse paper published in ICCV 2017 (Unofficial)

DeepFuse

This repository contains an unofficial Pytorch implementation of the DeepFuse network for image fusion of extreme exposure image pairs, published in ICCV 2017.

Prerequisits

This release of DeepFuse implementation was tested in Google Colab with

• Python = 3.8
• PyTorch = 1.10.0
• torchvision = 0.11.1

Dataset

The dataset used for this experiment is the subset of the SICE dataset given here: https://github.com/ytZhang99/CF-Net

The SICE dataset can be found here: https://github.com/csjcai/SICE

Training

Folder structure

Place ground truth, over exposed, and under exposed images for training in the following folder structure.

  - SICE_subset
      - test_data
          - GT
          - OE
          - UE
      - train_data
          - GT
          - OE
          - UE
      - val_data
          - GT
          - OE
          - UE

Training

    python Main.py --train True --use_cuda True --trainset "./SICE_subset/train_data/"

Testing

Set the use_cuda flag to False if necessary.

    python Main.py --train False --use_cuda True --testset "./SICE_subet/test_data/"

References

[1] K. Ram Prabhakar, V Sai Srikar, R. Venkatesh Babu. DeepFuse: A Deep Unsupervised Approach for Exposure Fusion with Extreme Exposure Image Pairs, ICCV2017, pp. 4714-4722

Acknowledgements

Thanks to Kede Ma et al. for their PyTorch implementation of the MEFSSIM metric from https://github.com/makedede/MEFNet