Incorrect modelfitting with different mesh

[Masu de Ridder]

Dear Scalismo community,

First I would like to thank you for the great SSM course and the additional support.

I am trying to create a script that results in a good model fitting for any input mesh that needs no further manual adjustment.

It works for the target meshes you provided, Thomas Vetter and Paola (see attachment) but not for any other mesh. I tested it with other stl files that were online available (see other attachment for example).

I used predefined landmarks on the model, that are only in the centre of the face as is visible on the paola image. The fitting with Vetter and Paola is nearly identical but as you can see, the fitting with the man's face is not. The grey face is the input mesh and the blue face is the resulting mesh.

Can you help me with this?

I am also curious as to why Vetter and Paola are even identical in areas such as the neck and ears, because those are not included in the landmark locations so I would expect that those areas look more like the model.

Kind Regards,

Masu

[Marcel Luethi]

Dear Masu

Are you using the model from the online course or did you download the full Basel Face Model?

If you are not using the model from the course you should expect large differences when you apply it to real data. It is very limited in its flexibility, as we built it from only 10 publicly available datasets. The two faces are most likely part of the model, and therefore even points that were not considered in the fitting will closely coincide.

In case you are using the full Basel Face Model, you need to make sure that you ignore the cone in its mouth interior during the fitting. Otherwise the algorithm tries to minimize the distance between this cone and the target.

Best regards,

Marcel

--
You received this message because you are subscribed to the Google Groups "scalismo" group.
To unsubscribe from this group and stop receiving emails from it, send an email to scalismo+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo/98ada38a-e16c-4a11-937c-a46018cb927cn%40googlegroups.com.

[Masu de Ridder]

Dear Marcel,

Thank you for your quick response. I will try out your suggestions. Thank you.

Kind regards,

Masu

[Masu de Ridder]

Dear Marcel,

I tried out your suggestions: I made sure that I used the full model from 2019 and I excluded the tongue cone from the model. I used a pointId per 10 pointIds. However, the result is still nog correct. I used Paola again. I attached the figures for clarification.

Do you have any tips?

Kind regards,

Masu de Ridder

[Marcel Luethi]

Dear Masu,

Registration is a really subtle problem and there are many things that can go wrong if one is not careful.

One thing that pretty much always improves the result is to start registration with a truncated model (Maybe only the 2-3 components) and a large noise term, and only in later iterations use the full model. This reduces the chance of getting stuck in bad local minima (as seems to be the case for you). Another strategy that can improve the results is to initialize the registration with a few landmarks, to get the pose right (you are adjusting for pose, right?) If you use landmarks, you could actually also build a posterior model and fit the resulting posterior.

Best regards,

Marcel

To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo/cae1f19c-1e95-474c-8604-b20f5c261a0dn%40googlegroups.com.

[Masu de Ridder]

Dear Marcel,

 

I used rigid ICP alignment before the model fitting, and based on the resulting visualizations this seems correct. That is what you mean by “adjusting for pose” right?

I will try out your new suggestions, thank you!

 

Kind regards,

Masu