search

NVIDIA / flownet2-pytorch

Pytorch implementation of FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
Other
3.12k stars 739 forks source link

FlowNetC does not converge #62

Open MatthewD1993 opened 6 years ago

MatthewD1993 commented 6 years ago

I have studied different network for optical flow, including FlowNet and its descents, SPYNet, PWCNet. I met a serious problem with correlation layer, https://github.com/NVIDIA/flownet2-pytorch/tree/master/networks/correlation_package. I tried to train FlowNetC and PWCNet (my own implementation) with this correlation implementation, but it never works. However, after I change the correlation to square error to construct the cost volume, for both networks the training converge very soon after 20 epoch or so.

​ I want to ask if you have trained FlowNetC by yourself, if so, how do you get it converge? And do you have any thoughts why Square Error is better in my case? Thanks so much if you could squeeze some time and reply my message. Looking forward to hearing from you!

duduheihei commented 3 years ago

Meet the same problem , FlowNetC does not converge and could not find the solution.