Your EDVR_REDS_deblur_Stage2.pth cannot reproduce the results in your paper

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

Your EDVR_REDS_deblur_Stage2.pth cannot reproduce the results in your paper #104

Closed ysyscool closed 4 years ago

ysyscool commented 4 years ago

Thanks for your source code and excellent work. I am testing the performance of your pre-trained models. The results of the EDVR_REDS_deblur_L.pth on REDS4 is the same as the paper. While the result of EDVR_REDS_deblur_Stage2.pth is only 25.84 dB not 36.49dB. Can you help me check it? Thanks!

xinntao commented 4 years ago

Did you use the outputs of EDVR_REDS_deblur_L.pth as the inputs of EDVR_REDS_deblur_Stage2.pth?

ysyscool commented 4 years ago

Did you use the outputs of EDVR_REDS_deblur_L.pth as the inputs of EDVR_REDS_deblur_Stage2.pth? Thanks! But I come with a new problem. Should the inputs of EDVR_REDS_deblur_Stage2.pth be the flip_test results of EDVR_REDS_deblur_L.pth? I put the outputs of the stage1 model without flip_test as the inputs. The PSNR of the stage2 is 35.71dB.