SeungjunNah/DeepDeblur-PyTorch

The performance result on GoPro_Large dataset

pp00704831 opened this issue · 2 comments

Hello,

I downloaded your GoPro_Large training results and tested the PSNR.
it is able to achieve 30.4 in PSNR , but why it is about 29.08 in PSNR released in your paper?

Hi @pp00704831,

The accuracy difference comes from the implementation changes.
RGB range change, loss change from L2 to L1, more epochs, etc.
Those changes are more like training tricks rather than academic contributions but matters in terms of accuracy.

Hi @SeungjunNah ,
I got it, thanks you for your fast reply and devotion to deblurring!