BFM 2019 shape parameters

[Catherine G]

Dear Scalismo group and hello everyone,

I hope this email finds you well.

I want to express my immense gratitude towards your generous sharing of the BFM models. They are very useful. I am currently using BFM 2019 version, and I used the following formula to generate random faces (in terms of shape, expression, and textures):

The formula seems to be working pretty well. I just have three questions:

1. Did I use the right formula? 

2. For the dataset dictionary, there is ['model']['pcaVariance'], and I used the square root of it, because then it corresponds to the formula above. Is this the right way?

2. I found that all the faces have similar width and height, despite their differences in facial features. Is it normal? I multiplied the shape coefficients (each drawn from a normal distribution, a~N(0,1)) by a factor of 800, and now the shapes start to show bigger variations in terms of width, height, and possibly some facial features.

Is multiplication by a constant factor to the shape coefficients a reasonable thing to do? I don't feel good about this because I think the eigenvectors and eigenvalues should already take into accounts these variations among the individual participants. I fear that I am using the wrong formula.

Note: I didn't have to multiply any constant for expression or textures, only shape. 

Thank you for your time.

Best wishes,

Catherine

[Andreas Morel-Forster]

Hi Catherine

We are glad that you find our work useful.

1. The formula is correct. This holds as long as you look at what is stored in the scalismo file format. The "mean" in the h5 files is always the mean mesh.
This is different from our theoretical work, e.g. Morphable Face Models - An Open Framework looking at the formulas (1) and (6), where we usually model deformations as a GP and hence have a mean deformation. By deforming the reference using the mean deformation you get the mean mesh.

2. Yes you have to take the square root of the variance.

3. A constant factor should not be needed. Without a picture of generated faces it is hard to tell you more what could be the issue. Maybe you simply have a too high expectation how much the faces should differ in size. You have to know also that our model as a bias. While we tried to have as much diversity in our model as possible, it has a bias based on which people we could motivate to provide their data for building the model (see also reference [24] in the above paper).

To get a feeling what you can expect from the model, you could e.g. use our parametric face image generator from https://github.com/unibas-gravis/parametric-face-image-generator to create some images. If I am right, then you could also directly use the BFM 2019 model by only changing the configuration files and using the provided JAR file.

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/d77a8971-835e-4835-b8da-529518c96c06n%40googlegroups.com.

[Catherine G]

Dear Andreas,

Thank you so much for the explanations! I do suspect that I have the wrong expectations of the face sizes. 

I will look at your parametric face image generator pipeline and generate some images to check my understanding. 

Best,

Catherine

[Bernhard Egger]

Dear Catherine,

the model viewer might also be helpful:

https://github.com/unibas-gravis/basel-face-model-viewer

you can export an rps file that gives you the shape parameters and you can export a ply that gives you a mesh. Same parameters should lead to same mesh to check your implementation.

Best

Bernhard

--

0 1 0
0 0 1
1 1 1

To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo-faces/57a5875a-0aba-4491-a983-22b0abfe5fedn%40googlegroups.com.

[Catherine G]

Thank you for the great suggestion, Bernhard! I was able to install this model viewer and check my implementation.