Fitting sparse and/or fragmentary data
Ennnasus
Dear statismo team,
I have some questions about fitting sparse and fragmentary data using statismo and would appreciate your help very much.
I have been working with the statismo library, building up a model based on approximately 300 shapes, each having about 20000 data points. The fitting always worked well and I never had to use any landmarks. The model looks fine and seems to be able to represent a lot of shape variations.
Now I would like to fit this model to a data file with much less data points (about 2000), which cover only a part of the area of the shape. There are no points at all at large parts of the shape. I pre-rotate the data file to the model mean before doing the fit, so the shapes are already very close to each other. However, it seems to me that without using landmarks, the shape isn't recognized at all and the fit is way off. When I start using landmarks, the fit gets much better, but still not very accurate.
My question is the following: in which way do the landmarks influence
the fitting process? What exactly is the difference between landmarks
and the 'normal' data points? If I use landmarks, are the other data
points still taken into account?
I've been experimenting a bit with the number of landmarks, trying out different constellations between three and twenty landmarks. Do you have any recommendations here? And how important is it to choose the right point-to-point correspondances here?
Thanks in advance for your help!
Best wishes,
Susanne
Thomas Albrecht
I may have an answer for this: the fitting works iteratively:
For each model point, the closest target point is sought, and then the model deforms so that at least on average every point moves closer to its closest target point.
This works best when both surfaces are similarly and densely sampled, the closest point is typically orthogonal to the model surface.
Now when the target surface is partial or sparsely sampled, the closest point may be quite far away or lie tangential to the model point. The information that gives to the next iteration of model fitting is confusing at best.
I once wrote a workshop paper on trying to tackle this. I think it involved in going back and forth between the two surfaces to find more reasonable point pairs for the next iteration.
The paper is available here: https://gravis.dmi.unibas.ch/publications/2012/MeshMed12_Albrecht.pdf
I am afraid there is no source code online though. I hope this can help as a first pointer at least.
Best regards
Tom