Landmark correspondence for regression data

[Maia R.]

Dear community,

It is unclear for me if to establish landmark correspondence in order to use them as regression data (e.g. for shape completion), do we use ICP for rigid alignment as in Tutorial 10 or a non-rigid one.

Thank you,

Best regards,

Maia 

[Marcel Luethi]

Hi Maia

In Scalismo, we use the landmarks to compute  deformation vectors. These deformation vectors should represent only the shape change from the reference and not include any rotation/translation/scale.

Therefore, if the shapes are not aligned, you should first standardize the pose by e.g. using rigid ICP or a procrustes alignment. Once pose has been standardized, you can use the landmark data also in a non-rigid setting,

either by computing a posterior model or by including them as a likelihood as in the Metropolis-Hastings tutorial

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/30ba98cc-feb7-4338-9d78-1af7fcced611n%40googlegroups.com.

[Maia R.]

Hi Marcel,

Thank you for the explanation.

To standardize the pose, we need landmarks, either defined manually or automatically and then use rigid ICP or a Procrustes alignment: do we need to define those landmarks on the model mean or on the model reference for the shape completion ? I think on the model mean as the non-rigid icp starts with the model mean as initial instance but I am not sure ...

Also, do we need a lot of points for the rigid alignment (ex: in Tutorial 10, there are 50 points and in femur project, only few landmarks are used). If I have a priori known anatomical landmarks correspondence (very few ones), do I still need to do rigid icp with more points in order to diminish the pose error ? I suspect that I have (small) misalignment error in my app due to either 1) landmarks error (noise or whatever) or 2) because the anatomical landmarks I use are defined on the model reference and not on the model mean.

Sorry to bother you with that since it was done long before. My question is more about the methodology to get independently the landmarks for shape completion (as in Tutorial 9). For shape completion, we need to have more correspondences in order to obtain more similar instances to the target. In Tutorial 9, 200 corresponding landmarks were used as input to the model posterior (full distribution). Tutorial 9 mentions that the method used to obtain those points (used as landmarks) is in Tutorial 10, which deals with rigid icp only from dense sparse points, that's why I am a bit confused .

Thank you very much,

Best regards

Maia

[Marcel Luethi]

HI Maia

We need to carefully distinguish landmark points from other points on the surface. Landmark points denote specific anatomical points on the surface, which can be consistently identified in all examples in the population. As we can identify the same landmark in all examples of the population, we automatically have correspondence. These landmark points are different from the points used in the rigid ICP. There we just have a number of arbitrarily chosen points on the surface, for which we do not know the correspondence.

The landmarks always need to lie on the surface that you wish to align. In the tutorial we want to align the model mean with the target, hence we have landmarks on the target.

When you have landmarks, there is usually no need to run a rigid ICP. Rigid ICP is useful for computing a rigid alignment, when you don't have given landmark points.

When we do shape completion, we usually also don't work with landmarks. Rather, we use in a first step non-rigid icp (or another fitting method) to find corresponding points in the model and the (partial) target. We then use these points to compute the completion.

Does this clear up your confusion?

Best regards,

Marcel

To view this discussion on the web visit https://groups.google.com/d/msgid/scalismo/3f9ca9be-321f-490f-98a9-fa75d0516570n%40googlegroups.com.

[V R]

Hello Marcel,

Thank you very much for the explanation.

I managed to get the same results as yours. 

Among my improvements is to define landmarks on the model mean and not on the reference mesh and align the noseless mesh to the model mean as the model mean is the initial instance for the shape completion (by non-rigid icp). I think I missed this information before.

Second, for the posterior of Tutorial 9, I get random points (200 as in this tutorial) from the target and the corresponding points on the reference model (as the deformation is from the reference to the target) and I get the same results as yours.

I think I better understand now.

Best regards,

Maia