search

jiangsutx / SRN-Deblur

Repository for Scale-recurrent Network for Deep Image Deblurring
http://www.xtao.website/projects/srndeblur/srndeblur_cvpr18.pdf
MIT License
709 stars 184 forks source link

Can get the same PSNR ans SSIM results as reported in the paper #56

Closed yJJtiG closed 3 years ago

yJJtiG commented 3 years ago

I can't install tensorflow 1.4 version, and I tested in the tensorflow 1.15 version according to the code and model provided by the authors, ans used the Matlab functions psnr() and ssim() to calculate the qualitative results. The results are PSNR = 27.34 dB, SSIM = 0.8952 on GOPRO testing dataset when using --model=lstm, which are far from the results PSNR = 30.19 dB, SSIM = 0.9334 provided by the authors. I don't know if it's because of different tensorflow versions?

Matlab code is as follows

% im_gt: Ground Turth, im_out: Output Image
psnr(im_gt, im_out)
ssim(im_gt, im_out)
yJJtiG commented 3 years ago

Or could the authors provide the Matlab code for calculating qualitative results?

jiangsutx commented 3 years ago

First, please check whether the produced results sharper and clearer than input images ? Make sure it produces reasonable results, rather than artifacts and noise.

Do the output and GT image correspond to the same image, without shift.

I remembder GOPRO dataset contains linear & non-linear version of images. Please try and select the correct one for the pretrained model. We use linear data (no gamma correction) for both training and testing.

yJJtiG commented 3 years ago

Thank you for replying quickly. I did use the wrong testing data and got the correct results after modification. Thank you so much.

ldlshizhu commented 3 years ago

Thank you for replying quickly. I did use the wrong testing data and got the correct results after modification. Thank you so much.

hello, sorry to ask you such a basic question. Excuse me, how to calculate psnr with matlab, is the input a picture? How many sets of pictures are used to calculate psnr and ssim?