new obj

[luoww1992]

i have a new obj file with 6172 points from the default obj file,
Q1:what is the meaning of the file face_mean and face_std and the two txt with smoothing ? Is the middle face and the hyperbole face ?
Q2: how to make the face_mean and face_std and the smooth txt file?

[alexanderrichard]

Hi,

1. face_mean contains the mean vertex positions over all training data and face_std contains the standard deviation of each vertex computed over all training data. It's used for a typical z-score normalization as you would apply in most neural network based systems.
2. Assume all you training meshes are stored in a tensor X of shape N x V x 3 (N: #training samples, V: #vertices). Then, face_mean = th.mean(X, dim=0) and face_std = th.std(X, dim=0)

I am not sure which file you are referring to when you talk about smooth.txt.

[luoww1992]

@alexanderrichard
sorry,
the smooth file are forehead_mask.txt and neck_mask.txt,

[alexanderrichard]

I created them manually but loading the neutral face mesh and selecting vertices in the respective areas. Meshlab is pretty easy to use and does the job just well.

[songtoy]

@alexanderrichard Does the face_std need to be scaled? I tried with no-scaling vector, but the model doesn't seem to converge (the audio code grows exponentially, and gradient for encoder vanishes.

[alexanderrichard]

Scaling should not be required. You'd want to make sure that there are no zeros in the face_std if there are vertices that never move. I can't really speak for you here as I don't know which data you use for training.