Are the results the same as those in the original paper?

[logicred]

Hello, thanks for your code implementation. I believe the results of in-domain (train and test data from the same dataset) in the original paper are correct, while the out-of-domain (train and test data from different datasets) results are too suprising to me. Previous works, which uses temporal information, can achieve about 0.7-0.8 AUC in CDF (up to ICCV 2021). I really hope that you can answer my question after some experiment.

By the way, few researchers can afford the raw version of FF++ videos : ).

[jtchen0528]

Unfortunately, I stopped working on this and have not produced out-of-domain experiments on many datasets.

The problem now mainly lies in I2G face generation. The paper did not show how exactly they cropped out the different portion of the face, for different manipulation methods such as DF, FF, FS, NT. I tried to train generated DF faces and tested on DF test set, it worked fine, but it failed hard on NT test set. It is because that the model learned only the borders of the whole face while NT only manipulated the mouth part. I tried some methods such as randomly taking out face parts (mouth, eyes, etc) (code part is here), but now, they did not work well on dealing with multiple manipulation methods at the same time.

Let's hope the paper later reveals how they generated the faces to deal with out-of-domain/unseen/multiple manipulate methods, or if somebody figure this out in the future. I am not working on this project now.