Next3D
Next3D copied to clipboard
Wrong DECA rendering code
Hey, I am trying to perform the preprocessing on FFHQ, but the DECA estimation code that you have is simply wrong.
https://github.com/MrTornado24/Next3D/blob/bdab84fc12cb61bd1002a3efb2c0a707a1139d97/dataset_preprocessing/ffhq/deca/decalib/deca.py#L170 basically changes the jaw rotation to be neutral. I don't know why.
And the line https://github.com/MrTornado24/Next3D/blob/bdab84fc12cb61bd1002a3efb2c0a707a1139d97/dataset_preprocessing/ffhq/deca/decalib/deca.py#L178
uses the original landmarks.
Using the original landmarks makes the outputs look like
Input
Landmarks
Texture
When you do not use the old landmarks
My question is this. When sampling which vertex positions/landmark positions should one use? (Yours)
The ones in the world space, or the ones generated by the orthographic projection. (The correct ones)