Questions about dividing dataset and network architecture

[JimHeo]

Hi, @soomean
I appreciate your great work.

I have some questions about datasets and network.

1. Are there any criteria to divide the datasets 'generate' and 'classify'?
   According to your paper, you divided the databases into homogeneous and heterogeneous datasets.
   Are the 'generate' datasets equivalent to homogeneous and the 'classify' dataset equivalent to heterogeneous?
   Then, do the 'generate' dataset mean Xia's datasets and the 'classify' dataset mean Aberman's datasets?
   I am wondering if all of Xia's datasets are used in training or are used in both training and testing.

2. In the mapping network, the unshared layer is not used.
   Isn't there any difference in performance even if I add the unshared layer like Stargan v2?

Thanks for your work again.

[soomean]

1. The homogeneous dataset (Xia's) is divided into 'generate' and 'classify' datasets, used for training and testing, respectively. There are no criteria to divide them into such two as they were randomly sampled. In this work, the heterogeneous dataset (Aberman's) is used just for testing style transfer into long-term heterogeneous motions.

2. We observed that the mapping network performed better without branches, which made a small difference with Stargan v2.

[JimHeo]

I understand that.
Thanks.

[JimHeo]

Hi, @soomean
Sorry to reopen the issue.

Could you tell the file index of Xia's dataset used to each 'generate' and 'classify'?
In my case, for example, kicking motion is not working actively.
I think the cause is imbalanced data when dividing by myself.
I would like to train from the scratch including data preprocessing.

Thanks.