search

princeton-vl / SEA-RAFT

[ECCV2024 - Oral, Best Paper Award Candidate] SEA-RAFT: Simple, Efficient, Accurate RAFT for Optical Flow
BSD 3-Clause "New" or "Revised" License
327 stars 11 forks source link

Training Infrastructure #8

Closed ChristophReich1996 closed 4 months ago

ChristophReich1996 commented 4 months ago

Hi, great work! I'd like to know which compute infrastructure was used to train SEA-RAFT. More specifically, how many GPUs have been used to train SEA-RAFT, and how long did full training (w/ pre-training on TartanAir) take? Thanks for the help :)

MemorySlices commented 4 months ago

We trained SEA-RAFT on 8 L40s. I do not remember the full training time but I believe it can be done in <1 week. Specifically, the ablation setting (540 x 960 input, 12k steps) takes about <1.5 days for SEA-RAFT(M).

ChristophReich1996 commented 4 months ago

Amazing, thanks for the details!

emlcpfx commented 2 months ago

What could be done to train the model further? Is the problem the availability of great datasets?

I find that any optical flow model has quite a bit of drift over time. Is that a lack of good real-world training data?