Multigrid method to train AVA

facebookresearch / SlowFast

PySlowFast: video understanding codebase from FAIR for reproducing state-of-the-art video models.

Apache License 2.0

6.65k stars 1.22k forks source link

Multigrid method to train AVA #253

Open ZZZZZZZZJ opened 4 years ago

ZZZZZZZZJ commented 4 years ago

Hi, thank you for sharing! I read the recent paper in CVPR 2020, "A Multigrid Method for Efficiently Training Video Models". The paper shows that the method has extreme performance for action recognition task. But I also curious about how it works on action detection tasks, such as AVA dataset. Would it work or not?

chaoyuaw commented 4 years ago

Hi, I think in principal it should work (at least to a certain extent), as I'm not aware of anything that would fundamentally stop multigrid strategy from working for detection. That said, I don't have empirical evidence to share yet, and I don't know how well the existing schedule would work for detection, but I'd be interested in knowing how it goes if you give it a try :)

coldmanck commented 4 years ago

@ZZZZZZZZJ Have you tried on it? If yes, could you give some feedback on this? Thanks!

abhaygargab commented 4 years ago

Hello authors,

Can you please let us know how to modify the data loaders to use Multi-grid to train AVA?

ThankYou