I am confused about FAN landmark generation.

[ChangYuance]

Should I use the mediapipe version or not? Are they only (i.e. https://github.com/hhj1897/face_alignment and mediapipe ) the different result of FAN landmark ?
I wonder whether the code could train with mediapipe version's result.
Also, I have downloaded the MEAD dataset. Is these precess commands available for MEAD dataset?
And I also wonder the file structure of MEAD dataset. What should my files look like. Is there somebody achieved training knows how to organize the files structure?

[ChangYuance]

And this is a wonderful job by the way!