can not reproduce SNN result on kodak dataset

[guochengqian]

Hi SNN team,
Thanks for releasing the code.
However, when I run the code under NLM. I cannot reproduce the result shown in Fig. 1 in the paper. For example, NLM_{0.8}^{16} achieves 31.01 PSNR as shown in Fig. 1. But the result I have is only 30.8.
Did I misunderstand something related to the code?

I revise the NLM/nlmTest.m by adding the imread code here. I used the default parameters.

    img = imread(fullfile(src_path, 'kodim04.png'));
    img = double(img)/255.;
    img_n = img + randn(size(img)) * sigma;

The PSNR I have is 30.8 using SNN.
Also, there are a lot of color artifacts in the eyes (which is different from the paper).

[ifrosio]

Real images are clipped between 0 and 1 (and noise statistics is consequently altered because of saturation), can you try with:

img_n = max(0, min(1, img + randn(size(img)) * sigma));

[guochengqian]

Real images are clipped between 0 and 1 (and noise statistics is consequently altered because of saturation), can you try with:

img_n = max(0, min(1, img + randn(size(img)) * sigma));

Hi. @ifrosio. Thanks very much for your quick reply.
However, I tried your suggestion. It did not work. The result is the same.

[ifrosio]

If I remember well, PSNR reported in Fig. 1 should be referred to the patches shown in Fig. 1, not the entire image. In the additional material, you may find full images, but we are not reporting PSNR.

[guochengqian]

okay. But the code I use is the correct one, right?
Since the obtained denoised images still have obvious color artifacts, which are not shown in the paper. This confuses me a lot.
Also, can you provide the link to the supplementary file? It is not available on the IEEE website.

[ifrosio]

Supplemental material is in the media section on IEEE: https://ieeexplore.ieee.org/document/8463600/media

As for the residual color artifacts, please check this part of the code:

sigma = 0.3;                % noise std
halfPatchSize = 3;          % half size of the patch
windowSearchHalfSize = 6;   % half size for searching the neighbors
N_n = 16;                   % numer of neighbors
h = 0.3 * sigma^2;          % nlm filtering parameter

parameters in nlmTest.m were optimized for the testing script, not necessarily for the Kodak dataset. You can try changing these parameters to see if the result improves. Sorry I do not have time right now to go back in my files looking for the optimal parameters - I may have more time in the next days.

[guochengqian]

Supplemental material is in the media section on IEEE: https://ieeexplore.ieee.org/document/8463600/media

As for the residual color artifacts, please check this part of the code:

sigma = 0.3;                % noise std
halfPatchSize = 3;          % half size of the patch
windowSearchHalfSize = 6;   % half size for searching the neighbors
N_n = 16;                   % numer of neighbors
h = 0.3 * sigma^2;          % nlm filtering parameter

parameters in nlmTest.m were optimized for the testing script, not necessarily for the Kodak dataset. You can try changing these parameters to see if the result improves. Sorry I do not have time right now to go back in my files looking for the optimal parameters - I may have more time in the next days.

awesome! Thank you. Looking forward to it.
I will try different ones by myself at first.