Can't reproduce the ideal performance

ray075hl commented 5 years ago

hello @assafshocher, I run the code with this command on Set5

python run_ZSSR.py X2_ONE_JUMP_IDEAL_CONF 0

The LR images processing step as follow First I used opencv convert Set5 HR images to YCbCr format, only saved Y channel of this images to calculate PSNR with SR images. Then, resized the Y channel with tool "assafshocher/Resizer", saved resized image as LR images, which are the inputs of ZSSR. The final average PSNR of Set 5 I obtained is 35.3 much lower than 37.3 Is that my processing wrong? Very appreciate for your help.

assafshocher commented 5 years ago

Hi @ray075hl , There are a few possible solutions:

We use full RGB and not just luminance for running ZSSR.
Try using python run_ZSSR.py X2_GRADUAL_IDEAL_CONF
Evaluation is done by a common MATLAB script by Kime Et Al. (basically taking only luminance, and cropping some deges)

ray075hl commented 5 years ago

@assafshocher Thanks a lot! I will try it.

ray075hl commented 5 years ago

I used opencv function cv2.cvtColor(img, cv2.COLOR_RGB2YCrCb) to get Y channel have bit differents with matlab code convert rgb to Y. Now I get the average psnr of Set5: 37.03.
Thanks.

assafshocher / ZSSR

Can't reproduce the ideal performance #2