RAFT
RAFT copied to clipboard
princeton-vl
RAFT small
Hi, Do you maybe have some results (epe on chairs test or sintel) after training your --small model just on FlyingChairs ? Does additional training on FlyingThings has big impact ?
I don't know if it helps you, since I reimplemented part of the code myself (although it should still use the same procedure as this original RAFT). After training RAFT small on FlyingChairs only, and then using 12 iterations during inference, I observed the following EPE results in the training sets:
- Sintel Clean: 2.7
- Sintel Final: 3.9
- KITTI 2012: 5.3
- KITTI 2015: 11.8
Thank you
Šalje: Henrique Morimitsu @.> Poslano: 11. srpnja 2021. 9:52 Prima: princeton-vl/RAFT @.> Kopija: Antonio Pavliš @.>; Author @.> Predmet: Re: [princeton-vl/RAFT] RAFT small (#93)
I don't know if it helps you, since I reimplemented part of the code myself (although it should still use the same procedure as this original RAFT). After training RAFT small on FlyingChairs only, and then using 12 iterations during inference, I observed the following EPE results in the training sets:
- Sintel Clean: 2.7
- Sintel Final: 3.9
- KITTI 2012: 5.3
- KITTI 2015: 11.8
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://github.com/princeton-vl/RAFT/issues/93#issuecomment-877757395, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AK4SXQPKOJRYQNCJ5ZVFOT3TXFETLANCNFSM42HRHMMQ.
I don't know if it helps you, since I reimplemented part of the code myself (although it should still use the same procedure as this original RAFT). After training RAFT small on FlyingChairs only, and then using 12 iterations during inference, I observed the following EPE results in the training sets:
- Sintel Clean: 2.7
- Sintel Final: 3.9
- KITTI 2012: 5.3
- KITTI 2015: 11.8
Could you provide the parameters you use to train the small model on the FlyingCharis dataset?Such as learning rate, Batchsize, imagesize and so on.Thank you very much.
