search

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

About 2 Flyingthings-3D dataset #196

Closed WormCoder closed 5 years ago

WormCoder commented 5 years ago

I found two flyingthings-3D dataset in your page (https://lmb.informatik.uni-freiburg.de/resources/datasets/SceneFlowDatasets.en.html)

And I downloaded the "DispNet/FlowNet2.0 dataset subsets", which has images about 35G.
In your paper, it seems that you used the other one, which has images about 37G. What's the difference between the two datasets? To train flownet2, must I download the other one? It's really large and long time for downloading. :)

nikolausmayer commented 5 years ago

The "subset" datasets represent what we used to train FlowNet2. They omit some very hard samples.

The original datasets have significantly more data. This could also be of use.

To reproduce our FlowNet2, use the "subset" datasets.

WormCoder commented 5 years ago

@nikolausmayer Thanks for your reply. I am new to optical flow and confused about the data. I found the folders, left and right in the Image_clean folder of "subset" dataset. What's the difference between the right and left? And Which one should be used for training?

"val/image_clean/left/0000108.png" and "val/image_clean/left/0000109.png" are from two different videos, so what's "val/flow/left/into_future/0000108.flo" is for?

nikolausmayer commented 5 years ago

"left" and "right" are stereo views.

You could use val/flow/left/into_future/0000108.flo for future flow prediction, but for optical flow training it is generally not useful. We just include it because it is available.

I added information about the lengths of individual sub-videos to the datasets page.

nikolausmayer commented 5 years ago

(closed due to inactivity)