Variational Simultaneous Stereo Matching and Defogging in Low Visibility

This repository stores the MATLAB implementation of the method described in

Variational Simultaneous Stereo Matching and Defogging in Low Visibility

Yining (Ted) Ding, Andrew Wallace and Sen Wang

In BMVC, 2022

Paper

Supplementary material

Poster

Video presentation

If you find our work useful in your research please consider citing our paper:

@inproceedings{ding2022variational,
    title = {Variational Simultaneous Stereo Matching and Defogging in Low Visibility},
    author = {Ding, Yining and Wallace, Andrew and Wang, Sen},
    booktitle = {BMVC},
    year = {2022}
}

Environment

The code was lastly tested in MATLAB R2022a on a laptop (running Ubuntu 22.04) equiped with a dedicated NVIDIA GPU that is supported by MATLAB. Please check MATLAB's GPU Computing Requirements.

Directory Stucture

├── classes
├── data/pad
├── util
└── demo.m

• classes stores various MATLAB class definitions.
• data/pad is to store the data needed (see Prepare data).
• util stores various MATLAB functions.
• demo.m is a MATLAB script that you are able to run to test our method on a sample scene.

Usage

Prepare data

1. Download the Pixel Accurate Depth Benchmark and unzip it.
2. Unzip the following: intermetric_rgb_left.zip, rgb_left_8bit.zip and rgb_right_8bit.zip. You can unzip others but only these three are needed by our code.
3. Copy directories intermetric_rgb_left (dense ground truth depth data), rgb_left_8bit (left intensity images) and rgb_right_8bit (right intensity images), and paste them to data/pad.
4. Load each .npz file in data/pad/intermetric_rgb_left and save it as a .mat file which can be loaded in MATLAB. The python code below should do the job. You can then manually delete all .npz files if you wish.

import os
import glob

import numpy as np
from scipy.io import savemat


npz_files = glob.glob("data/pad/intermetric_rgb_left/*.npz")    # you may need to adjust this path
for npz_file in npz_files:
    mat_file = os.path.splitext(npz_file)[0] + '.mat'
    data = np.load(npz_file)
    savemat(mat_file, data)
    print(f'generated {mat_file} from {npz_file}')

Run our method

Simply run demo.m. This will apply our method to a sample scene. In your MATLAB workspace you will find variables including a dense disparity map, a defogged intensity image and various metrics. You can try out other scenes, samples and visibilites.

Acknowledgements

Some of the code is replicated or adapted from the following repositories. We thank the authors for making their code open source.

fog_simulation-SFSU_synthetic by Christos Sakaridis

non-local-dehazing by Dana Berman

tofmark by David Ferstl

PixelAccurateDepthBenchmark by Tobias Gruber