This is an unofficial Python implementation of the Mertens et. al Exposure Fusion method using PyTorch. The method combines multiple images with different exposures to create a single image with an extended dynamic range. This implementation is 100% PyTorch-based, making it compatible with GPU acceleration for faster processing. Additionally, it should supports gradient backpropagation, enabling seamless integration into deep learning pipelines.

Input Image A	Input Image B	Input Image C	Exposure Fusion Result

• PyTorch

import torch
from exposure_fusion import exposure_fusion


# Load your input images as PyTorch tensors
# Replace ... with your image tensor, of size [N, C, H, W], where
# N is the number of frames, C the color channels, H the height and W the width
burst = torch.tensor(...)

# Perform exposure fusion
result = exposure_fusion(burst)

# Use the resulting tensor as needed (e.g., display or further processing)

The example.py script contains an example. It may require a few more modules to read and save the images. Feel free to try it on the few bursts provided in the data folder.

import os
import glob
import torch as th
import cv2
import matplotlib.pyplot as plt
from pathlib import Path

from exposure_fusion import exposure_fusion

# Read image burst
burst_path = Path("data/mask")

im_path_list = glob.glob(os.path.join(burst_path.as_posix(), '*.jpg'))
im_path_list += glob.glob(os.path.join(burst_path.as_posix(), '*.png'))
assert len(im_path_list) != 0, 'At least one .jpg or .png file must be present in the burst folder.'

burst = []
for im_path in im_path_list:
    burst.append(
        cv2.cvtColor(
            cv2.imread(im_path, cv2.IMREAD_UNCHANGED),
            cv2.COLOR_BGR2RGB)
        ) # flag to keep the same bit depth as original

# Normalise the burst between 0 and 1
burst = th.Tensor(burst)/255
burst = burst.movedim(-1, 1) # batch, channel, H, W format


out = exposure_fusion(burst)

out = out.clamp(0, 1)
out = out.movedim(0, -1).cpu().numpy() # [H, W, C] format for matplotlib
plt.imsave("out/result.png", out, vmin=0, vmax=1)

The original implementation uses 5x5 gaussian filters for downsampling and upsampling the stages of the gaussian pyramid. I used instead a 3x3 binomial filter. This modification is chosen for its efficiency, speed, and the observation that it does not significantly seem to impact the quality of the results.

If you encounter any issues, bugs, or have questions about the PyTorch Exposure Fusion implementation, feel free to reach out for assistance. You can:

• Open an Issue: If you believe you've identified a bug or have a feature request, please open an issue on the GitHub repository.

• Email: For private inquiries or specific concerns, you can reach me via email at [email protected].

I appreciate your feedback and will do my best to address any concerns promptly.

This implementation is provided under the MIT License