search

MrTornado24 / Next3D

[CVPR 2023 Highlight] Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars
https://mrtornado24.github.io/Next3D/
470 stars 29 forks source link

Wrong DECA rendering code #28

Open wamiq-reyaz opened 1 year ago

wamiq-reyaz commented 1 year ago

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 00000_inputs Landmarks 00000_landmarks2d Texture 00000

When you do not use the old landmarks

00180_inputs 00180_landmarks2d 00180

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)