What is the Training set of those pre-trianed models that provided by your dropbox link about icpr2020dfdc HOT 1 CLOSED

Hey @Chancing0 ,

for the DFDC, we relied on the raw data and did not use the HQ nor the LQ data.
We proceeded in this way for two reasons:

1. we extracted the faces from each video and saved them as JPEG (even though with a high quality factor, 95);
2. during training, as data-augmentation we perform JPEG compression with ranging quality factor to simulate a video compression operation (even though surely video compression is a more hard operation on data quality).

For the FF++, we used instead original video and c23 compressed videos.
Hope this helps!
Bests,

Edoardo

from icpr2020dfdc.