experiment result

[yinhaibin]

hello, i used your codebase to do some experiment, however, i got a poor PSNR. Have you got a PSNR as the paper presented? thanks!

[kweisamx]

hi @yinhaibin
Because I use this for some production, so the test I scale up the image directly
If you want to test the PSNR, you need to scale down the image then use network to scale up,
it just can be calculate the PSRN.

[FBLeee]

hi @yinhaibin
Because I use this for some production, so the test I scale up the image directly
If you want to test the PSNR, you need to scale down the image then use network to scale up,
it just can be calculate the PSRN.

What's the reason for the big difference between the results of PSNR and those in the paper？Now, the results of the most star reappearance network in gitub are far lower than the PSNR results in the original paper! I am very confused!

[zhangcheng-007]

hi @yinhaibin
Because I use this for some production, so the test I scale up the image directly
If you want to test the PSNR, you need to scale down the image then use network to scale up,
it just can be calculate the PSRN.

The two methods show different PSNR values, your PSNR is high, and the other performs poorly. I commented out the following code of make_sub_data to complete the other method.

if not config.is_train:
    input_ = imread(data[i])
    input_ = input_ / 255.0
    sub_input_sequence.append(input_)
    return sub_input_sequence, sub_label_sequence