Questions about the use of PWC-Net. Looking forward to your reply.

[JQJ001]

It's an excellent work! I have two questions to ask:
(1) Did you directly load the weight of pytorch-pwc(https://github.com/sniklaus/pytorch-pwc) to estimate optical flow? Have you fine-tuned on your own dataset?
(2) When I warp the image with the output optical flow estimated by pytorch-pwc, there are duplicates (as shown below). Have you encountered this problem?

Thanks! Looking forward to your reply.

[cszhilu1998]

Thanks for your attention to this work.
(1) We directly load the weight from pytorch-pwc(https://github.com/sniklaus/pytorch-pwc) to estimate the optical flow and do not fine-tune it.
(2) I guess the possible reason for this problem is that the image size is too large. It is recommended to use the down-sampled image to calculate the optical flow, and then up-sample the optical flow and multiply the factor of the up-sampling.

[JQJ001]

Thank you very much!
I tried to use the down-sampled image to calculate the optical flow, and then up-sample the optical flow and multiply the factor of the up-sampling, but the problem was still there.
Duplicates usually occurs when the foreground object moves a lot, and I'm not sure if the problem is due to occlusion.

[cszhilu1998]

Maybe you can try a large down-sampling factor. In the case of occlusion, the optical flow may indeed be difficult to exert a good effect.

[JQJ001]

OK, I'll try.
Thanks again.