JingyunLiang / RVRT

Recurrent Video Restoration Transformer with Guided Deformable Attention (NeurlPS2022, official repository)
https://arxiv.org/abs/2206.02146
Other
354 stars 33 forks source link

Training Vimeo90K with 14 Frames #24

Open gauenk opened 1 year ago

gauenk commented 1 year ago

Hello :wave: Thank you for your awesome code. I want to understand how you train a network using 14 frames when Vimeo90K only provides 7 frame sequences.

I notice many sequences continue consecutively with subfolder ID (i.e. 00001/0003 and 00001/0004 seem from the same sequence). However, many sequences do not (i.e. 00001/0001 and 00001/0002 are from the different sequences).

Did you use a script to identify only sequences with 14 frames for grouping? Or did you group pairs of adjacent 7-frame sequences together using an odd+even ID value? Or maybe another method I've not thought of?

Thank you for your thoughts.