search

XPixelGroup / BasicSR

Open Source Image and Video Restoration Toolbox for Super-resolution, Denoise, Deblurring, etc. Currently, it includes EDSR, RCAN, SRResNet, SRGAN, ESRGAN, EDVR, BasicVSR, SwinIR, ECBSR, etc. Also support StyleGAN2, DFDNet.
https://basicsr.readthedocs.io/en/latest/
Apache License 2.0
6.47k stars 1.15k forks source link

Why is there a gap between the PSNR in paper and the results of official pre-trained model? #634

Open CJ0109 opened 10 months ago

CJ0109 commented 10 months ago

The following are the results of official pre-trained model EDSR_Lx4 (https://drive.google.com/drive/folders/1rtJCHuOAEixB1OWmUVbbVm158vzC3kTt) image

and the following is reported in EDSR paper image

For x4 SR, the model trained on DIV2K. The test results on Set5 and Set14 seem to be very different from those written in the paper, but the results on the DIV2K validation set are similar. Why is this? Are more pixels cropped when computing PSNR on Set5 and Set14?

CJ0109 commented 10 months ago

test results/paper Set5: 30.5582/32.46 Set14: 27.0147/28.80 DIV2K val: 29.2770/29.25

hanzhangshen03 commented 2 days ago

In your test configuration (yml) file, you need to set test_y_channel to true under val/metrics. Hope this helps.