lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

Solver.prototxt for sintel and kitti pretrained models. #124

Closed TusharNimbhorkar closed 6 years ago

TusharNimbhorkar commented 6 years ago

Hey

I was wondering what solver.prototxt files you used for sintel and kitti models.

And are these models got only trained on this two dataset.

nikolausmayer commented 6 years ago

Our published models for Sintel and KITTI are pretrained on FlyingThings3D, then finetuned on the respective dataset’s training set.

The solver.prototxt file is exactly the same as the one for pretraining—with the exception of the crop parameters. Sintel and KITTI have different image sizes and aspect ratios, and the crop size for FlyingThings3D would not work.

TusharNimbhorkar commented 6 years ago

Thanks for the reply.

One more thing. Do you have anything to check epe while training?

nikolausmayer commented 6 years ago

We just print loss values from the loss layers, nothing more fancy

TusharNimbhorkar commented 6 years ago

Most of the statistics in the Flownet2 paper are given in EPE. Do you calculate that separately?

nikolausmayer commented 6 years ago

FlowNet2 is trained directly on EPE as loss, so we don't have to do that separately.

TusharNimbhorkar commented 6 years ago

Ah okay, so the Loss which is shown during training is basically EPE.

nikolausmayer commented 6 years ago

Yes

TusharNimbhorkar commented 6 years ago

Thanks again closing the issue.