The AUC on the DFD dataset is not as high as stated in the paper.

[IApple233]

I am trying to reproduce your results, but when I use the pretrained weights to test the AUC on the DFD (DeepFake Detection), It's only about 88%. I am quite sure that my method is working correctly. I would like to know if your DFD dataset is the official complete repository (approximately 300 real and 3000 fake), and if there were any special preprocessing methods applied.

My test results on other datasets(DFDC,CDF) are quite consistent with those in the paper, but the results on DFD differ significantly. This is my implementation of the Init_DFD function
real_path='/home/data/liqirui/FF_download/data/original_sequences/actors/c23/videos/*.mp4' real_videos=sorted(glob(real_path)) fake_path='/home/data/liqirui/FF_download/data/manipulated_sequences/DeepFakeDetection/c23/videos/*.mp4' fake_videos=sorted(glob(fake_path))

[mapooon]

Hi,
We used the "raw" videos (not "c23") for the evaluation.
Proprocessing for DFD is the same as the other datasets.

[Bytebender3]

@IApple233 Hello,I am reproducing this work recently. I want to ask that Is DFDC test dataset has 400 videos?and Can you give the labels.csv of it?Thank you very much.