Low confidence when detecting obviously deepfakes

[Daleyzw]

这个结果是： 0.2054801881313324

这个结果是：0.083944887

[HighwayWu]

Thank you for your question. I think there are two reasons:

• The model of this project is trained on pre-processed data (pre-cropped and aligned faces), so it may not be suitable for portraits with large distribution differences. It is recommended to crop the image to 512*512 with the face centered.
• Generalization (i.e. detecting cross-domain forgeries) has always been the core indicator of the Deepfake Detection task. Unfortunately, there is currently no well-defined solution to deal with the ever-changing forgeries. This is also the goal of this competition and this project, to build a powerful general deepfake detector.