spatial-correlation-sampler

[jing98tian]

i am new to this area, would you relsese the code with "spatial-correlation-sampler" function for me to learn, thank you !

[ltkong218]

I have added the file ./models/FastFlowNet_.py that supports higher versions of CUDA and PyTorch. Please see the updated README.md.

[jing98tian]

thank you for your reply, i try it in cuda10.2/pytorch1.6 and it works.

[jing98tian]

I find that the network you proposed has poor generalization performance for untrained scenarios. Are you concerned about this problem? and The network seems to have problems in optical flow estimation of distant objects

[ltkong218]

All the supervised optical flow networks have this problem, since they are trained on synthesized datasets which have large domain gap with real scenes. I suggest you to train FastFlowNet with your video sequences in a self-supervised way, where you can load the released pre-trained model parameters as initialization. For unsupervised learning, you can refer to UnFlow, DDFlow, SelFlow and UFlow.

[jing98tian]

thank you very much ,I'll try your suggestion.