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

Train dataset #212

Open xubin1994 opened 4 years ago

xubin1994 commented 4 years ago

Hi,I saw on the sceneflow dataset website that you mentioned that you have customized a Flything3D subset. In this subset you ommitted some extremely hard samples. So my question is what is the standards for the "extremely hard samples"?And If using the original Flything 3D subset, what impact will it have on the network? image

nikolausmayer commented 4 years ago

IIRC "extremely hard" samples have disparity magnitudes > 300 pixels (I am not sure whether there was also a flow criterion, but I do not think so). I think training on the original full set should still work well, maybe a bit slower.

xubin1994 commented 4 years ago

IIRC "extremely hard" samples have disparity magnitudes > 300 pixels (I am not sure whether there was also a flow criterion, but I do not think so). I think training on the original full set should still work well, maybe a bit slower.

I traversed the Flything3D subset, and the subset still has a few samples containing more than 25% of the pixels, which disparity magnitudes > 300。This makes me more curious about the screening.