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

Purpose of setting black border in BlackAugmentationLayer #140

Closed wang-yijie closed 6 years ago

wang-yijie commented 6 years ago

Hi, BlackAugmentationLayer randomly sets corresponding borders of image and flow to zero. What is the purpose of this augmentation? In testing, input images are resized to the desired shape, containing no black borders. Why black borders are added during training?

nikolausmayer commented 6 years ago

Hi,

the black border is a type of augmentation. We use it to teach the network how to deal with images that have, well, black borders ;) The small-displacements network tends to treat image borders like this wrongly, so we teach it explicitly.

wang-yijie commented 6 years ago

Thx!