search

NVIDIA / flownet2-pytorch

Pytorch implementation of FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
Other
3.09k stars 739 forks source link

Model difference #200

Open lhao0301 opened 4 years ago

lhao0301 commented 4 years ago

Thanks for the wonderful work!

What's the model difference between flownet2-SD and flownet2-SD within flownet2? Was flownet2-SD trained on flyingthing3d+ChairSD or only ChairSD? And how the flownet2-SD network within flownet2 was trained, end-to-end? I test the flownet2-SD(with the released flownet2-SD weights) and the flownet2-SD within flownet2(with the flownet2 weights), the results looks very different. @fitsumreda

lhao0301 commented 4 years ago

From https://github.com/lmb-freiburg/flownet2/issues/112, I get that flownet2-SD is trained on ChairSD only, and the flownet2-SD's weights within flownet2 is identical. However, I load the two models respectively and find that the weights of flownet2-SD is different from the weights of flownet2-SD within flownet2, which is not consistent with https://github.com/lmb-freiburg/flownet2/issues/112. Also, the optical flow's prediction of them looks different.