search

zzh-tech / ESTRNN

[ECCV2020 Spotlight] Efficient Spatio-Temporal Recurrent Neural Network for Video Deblurring
MIT License
314 stars 39 forks source link

Average PSNR drops when testing a sequence checkpoint #9

Closed Sazoji closed 3 years ago

Sazoji commented 3 years ago

when testing the best checkpoint, which when the sequence is in /valid the PSNR is 36.038 during training, the same sequences (renamed to /test) will have a PSNR of 27.366(the same or even blurrier as what's being fed in as the blur) it is applying some change, but it's not sharpening the image. Input 00000088_input ESTRNN 00000088_estrnn GT 00000088_gt

zzh-tech commented 3 years ago

First, the PSNR is calculated using cropped data (256x256) for the training and validation steps, while the entire input is used for the test step. However, 36.038dB is very high and the restored result should be sharp. Could you please send me the checkpoint (including log) and the data sequence? (Email: zzh.tech@gmail.com) I may check it for you sometime.

Sazoji commented 3 years ago

results are sharpening above 28psnr but require a lot more epochs (>800) for some scenes than the default 500, although that may depend more or less on the type of sharpening being done (resolving real blur vs generated) and dataset size.