xinntao / EDVR

Winning Solution in NTIRE19 Challenges on Video Restoration and Enhancement (CVPR19 Workshops) - Video Restoration with Enhanced Deformable Convolutional Networks. EDVR has been merged into BasicSR and this repo is a mirror of BasicSR.
https://github.com/xinntao/BasicSR
1.48k stars 320 forks source link

Unable to get expected performance on stage 2 model #129

Open qqi0O0 opened 4 years ago

qqi0O0 commented 4 years ago

Hi, I downloaded the pretrained stage 1 + 2 models for the blur track (EDVR_REDS_deblur_L.pth and EDVR_REDS_deblur_Stage2.pth), ran stage 1 on REDS4 blur track to get a new set of images, ran stage 2 on the new set of images, then calculated the PSNR of the result against the ground truth using the provided script. I was able to get results that are better than using stage 1 alone, yet worse than what was reported on the paper:

INFO: Folder 000 - Average PSNR: 37.438530 dB. Center PSNR: 37.526899 dB. Border PSNR: 35.317672 dB. INFO: Folder 011 - Average PSNR: 35.941583 dB. Center PSNR: 36.022184 dB. Border PSNR: 34.007162 dB. INFO: Folder 015 - Average PSNR: 36.293789 dB. Center PSNR: 36.470988 dB. Border PSNR: 32.041014 dB. INFO: Folder 020 - Average PSNR: 33.187654 dB. Center PSNR: 33.301712 dB. Border PSNR: 30.450266 dB. Total Average PSNR: 35.715389 dB for 24 clips. Center PSNR: 35.830446 dB. Border PSNR: 32.954029 dB.

I was able to reproduce stage 1 results without issue, but these stage 2 results are better than the 34.80 of stage 1, yet worse than 36.37 reported for stage 2. Do you know why this might be? Thank you!