Why is the effect of naked training motiondeblur poor?

[aiaini66]

I according to the parameter settings in train_motiondeblur.sh, the datasets used GroPro, but the test effect is far worse than the effect of your opened model。 (https://mailustceducn-my.sharepoint.com/:u:/g/personal/zhendongwang_mail_ustc_edu_cn/EfCPoTSEKJRAshoE6EAC_3YB7oNkbLUX6AUgWSCwoJe0oA?e=jai90x) Is there a problem with the training parameter settings？ Or some other reason？ as showed follow，left is your model result， right is my model result。 Hope you reply。 thanks for lot !!! @vinthony @ZhendongWang6