About training strategy and pre-trained model.

[Leeeshuz]

I am very interested in your paper and codes, so I want to run the code and I have preprocessed the pose files according to FD-GAN. However, I am confused with the training strategy. What is the difference between the train_cross_cont.sh and train_cross_stage2.sh? How to train the model for stage 1, or would you mind providing the pretrained models used in the .sh files? Thank you very much for your help!

[yujheli]

Please use train_cross_stage2.sh for use, train_cross_cont.sh is for visualization. I will update the Readme to make it more clear.

[Leeeshuz]

Please use train_cross_stage2.sh for use, train_cross_cont.sh is for visualization. I will update the Readme to make it more clear.

Thx！ Looking for your update!

[QingzeYin]

Hi, Have you got the same results as the FD-GAN proposed about pose files on Market? I use the rtpose model from 'pytorch_Realtime_Multi-Person_Pose_Estimation' which cannot get the same results on Market. And the results I got were very weak. Thanks for your help which is very important to me.

[yujheli]

FD-GAN has its stage-III pre-trained model which has its reported result (https://github.com/yxgeee/FD-GAN). Yet, I could not reproduce their finalized result following stage-II and stage-III as well so I directly load their pre-trained weights for the source domain in our work.

[QingzeYin]

Thanks for your answer.
Could you please give me a hint about the evaluation in the final test which was highlight below. Do you mean the test-loader is the target-test-loader?
Thanks a lot!

[yujheli]

Oh, you are right. That is target-test-loader.

[QingzeYin]

Got it. Thanks for your confirmation.