the problem of PCD_Align

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

the problem of PCD_Align #193

Closed browneyes1222 closed 3 years ago

browneyes1222 commented 3 years ago

I wonder why L3_offset *2 before concat with L2_offset. thanks

ckkelvinchan commented 3 years ago

Since the offsets are mimicking optical flow, we should multiply the offsets by 2 when we do upsampling (motions should be twice larger when you upsample the images by 2x.) This should not make a large difference, feel free to try without it.