Inconsistency between FAN detected landmarks and DECA predicted landmarks

[Miaoing]

Hi, I notice that the landmarks detected by FAN or openpose or dlib, have shorter eyebrows, while the eyebrows predicted by DECA are longer. How do you deal with this inconsistency, when computing landmark loss in the training process? Thank you for your response.

[adevra]

I have just came across this issue, especially lip corners are totally off. I wish someone could explain how to use our own generated 68pts landmarks for reconstruction.