I am trying to run the model on a few samples from DeeperForensics and LRW samples. The predictions from the model are heavily skewed towards fake.
{'Deepfakes': 1.0, 'LRW': 0.5454545617103577, 'Real': 0.5}
The ACC observed on a small dataset of 10 real LRW videos is 0.54.
The ACC observed on DeeperForensics is 1.0.
The ACC observed on custom dataset of real samples also shows a similar trend where it calls 50% of samples as fake.
Wondering if anyone else observed the same?
NOTE:
When I look at the face crop outputs generated they seem to be pretty jittery. Wondering if that is expected?
I am using face-alignment to generate the 68 landmarks and using the extract_faces.py script to generate the face crops.
I am trying to run the model on a few samples from DeeperForensics and LRW samples. The predictions from the model are heavily skewed towards fake.
{'Deepfakes': 1.0, 'LRW': 0.5454545617103577, 'Real': 0.5}
Wondering if anyone else observed the same?
NOTE: When I look at the face crop outputs generated they seem to be pretty jittery. Wondering if that is expected? I am using
face-alignment
to generate the 68 landmarks and using theextract_faces.py
script to generate the face crops.