MAD_Competition

Matlab code for MAD Competition

Initial versions of all code provided by Zhou Wang via email in July 2020, except for the following:

• ssim_inex.m downloaded from Lab for Computational Vision website in July 2020
• contents of the msssim folder downloaded from Zhou Wang's website in August 2020.

Download data

Images used to generate synthesized images found on the OSF. If you have wget installed on your system (you probably do: run which wget from the command line and, if it prints a path, you do), you can use the included download_data.sh script: from the command line, run ./download_data.sh, and it will download the data, unzip the images, and remove the zip file.

Requirements

• MATLAB (not sure of version right now, but works at least for R2018b)
• matlabPyrTools

Add matlabPyrTools to your path before running any of these functions

Contents

This file contains the files to calculate SSIM (standard and weighted) and MS-SSIM, as well as to run MAD Competition with the weighted version SSIM, as shown in the paper. We primarily are using it to check the outputs of plenoptic's implementation of SSIM and to synthesize images for comparison.

Use

The primary function is fungds_mse_ssimweighted.m, which synthesizes a complete set of MAD Competition images (four: two that fix SSIM and min/max MSE, two that fix MSE and min/max SSIM) and saves the output. Its call signature is

fungds_mse_ssimweighted(mse, iter, img, save_path)

where mse is the amount of noise added (equivalent to initial_noise in plenoptic, this is the fixed MSE value), iter is a 1d matrix with 4 values (e.g., [10 10 10 10]) that gives the number of iterations for each image (in order: fix MSE max SSIM, fix MSE min SSIM, fix SSIM min MSE, fix SSIM max MSE; note that for SSIM max is best and min is worst, while it's the opposite for MSE), img is the path to the target image, and save_path is the path to save the results at (as a .mat) file.

synthesis_imgs.m is a wrapper function, which will create each of these for the 10 sample images we provide and noise levels going from 8 to 1024 (log-spaced), with 100 iterations per image.

plenoptic_analysis.m saves several SSIM values (both weighted and standard) for use in plenoptic's tests. We then upload this manually to the OSF page.