search

polimi-ispl / icpr2020dfdc

Video Face Manipulation Detection Through Ensemble of CNNs
GNU General Public License v3.0
259 stars 100 forks source link

the proportion of true and false video is very unbalanced #53

Closed IItaly closed 3 years ago

IItaly commented 3 years ago

Hey, @CrohnEngineer the proportion of true and false video is very unbalanced, but it seems that you split it by 1:4. Doesn't it need to be balanced?

CrohnEngineer commented 3 years ago

Hey @IItaly ,

which dataset are you talking about? DFDC or FF++? Bests,

Edoardo

CrohnEngineer commented 3 years ago

Anyway, consider that even though both datasets are heavily unbalanced, we have constructed the FrameFaceIterableDataset so that the batches are balanced in the number of FAKE and REAL faces, and to stop the training if we have analyzed all the REAL faces. However, usually the networks overfit long before having analyzed all the REAL faces, and usually the training stops for reaching a plateau on the validation loss or the maximum number of iterations.

IItaly commented 3 years ago

Thanks for your reply. I'm talking about FF + + dataset. Because I found that other papers will enhance the real pictures to balance the proportion, so there is such a question.Now I have no problem.

CrohnEngineer commented 3 years ago

Glad this solved your doubts :) Bests,

Edoardo