Basel Face Model 2019 Point Coordinates

[Wenxuan Guo]

Hello,

I'm writing a Python program that uses the Basel Face Models to generate faces. Thank you so much for sharing the data! I found that the 2009 model has the attribute parameters. Does 2019 also have the relevant attribute parameters to change shapes?

I also found that the coordinates of the shape parameters are different in the 2009 vs. 2019 versions.

For the 2019 model, the mean shape parameters are:

For the 2009 model, the mean shape params are:

Can someone explain to the me the difference, please? Is it just a scale difference? Thank you again for sharing the data.

Best wishes,

Wen

[Andreas Morel-Forster]

Hi Wen,

Glad to here that the data is usefull to you.  Unfortunately the attributes are not published or calculated for the 2019 model.

While the models are simmilar, there are few differences between the models:

• The 2019 model uses millimeters, hence there is a general factor of 1000 difference between the models.
• The 2019 model has a different surface parametrization, so a different ordering of the values within the datastructure.
• The 2019 model is not built from exactly the same data. While both use 200 identities to estimate the face-space, not all 200 ids are identical.

Best, Andreas

--
You received this message because you are subscribed to the Google Groups "scalismo-faces" group.
To unsubscribe from this group and stop receiving emails from it, send an email to scalismo-face...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo-faces/d8b8abb7-7096-4da7-b86c-415bc32ee4c6n%40googlegroups.com.

[Wenxuan Guo]

Dear Andreas,

Thank you so much for the direction. I found that the 2009 BFM model in the .h5 format also uses millimeter just like the 2019 BFM model, though the 2009 model in .mat format uses micrometer. And I was able to use the attribute.mat file now for different attributes like age. Though the coefficient of the attribute parameters required trials and errors, now they look great!

For the 2019 model, I assume that in order to model the deformation fields with the 'pcaVariance' in the file, we have to use the square root of the eigenvalue (Karhunan-Loeve expansion), but the shape variance is not that big, so I have to multiply by some constant factor to increase the variance/standard deviation. However, since we don't have the nonlinearity of these attribute parameters, I can't increase the whole vector by too much, or it will produce a weird shape. They look like this:

I have one more question about the expression parameters. To model the face, we have a texture map and a shape map. And we have three sets of parameters in the model, shape, expression, and texture. I use texture for the texture map. For the shape map, I am applying this formula:

shape_map = shape mean values + (shape PCA basis * square root of eigenvalues of the shape * some constant that increases the variance) 

                         + (expression PCA basis * square root of the eigenvalues of the expression**an exponent that increases the variance)

Do we just add the expression parameters directly to the shape parameters? I didn't add the mean value of the expression because it looks weird. Everyone would have their tongues out...Or perhaps I'm using the wrong formula?

Thank you for the great help!

Best wishes,

Wen