Time for training PWC net with dataset Flying chairs

[JingleiSHI]

Hi,
Thank you again for your code and your answer for my former questions. Now I am trying to train from initialization a PWC net with dataset Flying chairs, but I found that the EPE doesn't decrease (always about 11~12), do you have the same problem when training PWC net with dataset Flyingchairs or Sintel. If so, could you please tell me how to solve this problem? Thank you very much for your attention.

Yours sincerely,
Jinglei

[daigo0927]

Thanks, I also see the same problem at the training on FlyingChairs dataset in my environment.

I'm sorry that I have not figure out the cause for this, but I also confirmed that fine-tuning from the model_3007.ckpt seems to adapt to the FlyingChairs dataset.
If I found the solution, I will immediately update this repository.

Kindly regards.

[xianshunw]

@daigo0927 how is this problem solved?

[LiuzhuForFun]

@daigo0927 Thanks for your excellent job!
how is this problem solved? How the performance of PWCDCnet? How to train FlyingChairs for a good result? Thanks a lot!

[xianshunw]

@LiuzhuForFun it seems that the net is easy to get stuck in local minimum at the beginning. Try to train from different initial states. That's my solution.

[LiuzhuForFun]

@xianshunw,Thanks for your suggestion.Do you mean that train a lot times to find a suitable optimum？

[xianshunw]

@LiuzhuForFun I mean that if your net doesn't convergence after thousands of iterations then train from scratch again. This will help your net escape from this local minimum. This problem is mentioned in the original paper, see Optical flow estimator of Ablation Experiments(Section 4.2) in the original paper. The authors said that the

deeper optical flow estimator might may get stuck at poor local minima

But in my experience, the original optical flow estimator also have this problem.

[LiuzhuForFun]

@xianshunw Thank you very much.How about the performance of the net your trained?

[xianshunw]

@LiuzhuForFun without so many data argumentations as the original implementation, the training epe error can reduce to 1.1~(probably, I forget) and the validating epe can drop below 2 on FlyingChairs.

[LiuzhuForFun]

@xianshunw Thanks again! I got into trouble: the model performed well in the dataset I trained(for example FlyingChairs 1.23),but the result didn't perform in other dataset as well as the orgin paper put.Dou you have the same problem?

[xianshunw]

@LiuzhuForFun which dataset? FlyThings? Sintel or KITTI?

[LiuzhuForFun]

@xianshunw Sintel and Kitti

[Choneke]

For testing on other data set, you have to train it on FlyingChairs dataset. When you train on FlyingChairs Dataset, you can use the model to test it on any dataset.

[xianshunw]

@Choneke interesting, how is your model performing on sintel when it's only trained on chairs?

[lelelexxx]

@xianshunw Thanks for your sharing, I just reimplement the PWCnet using mxnet, During the training on MPI-Sintel, the flow_6 (the smallest scale) did not convergence, and remains in constant. while the optic flow from other scales seems good. Did you noticed this in your training?

[xianshunw]

@lelelexxx hope this issue helps you. NVlabs/PWC-Net#30