[vtkusers] vtkProcrustesAlignmentFilter problem

[Sebastian Ordas]

Hi Jonathan,
Thanks for the explanation. Let me further specify my problem:
Polydata members of each of my two groups have both, same number of
landmarks
and strict correspondence (i.e. between themselves). That's why I do not
need to use ICP. I want to align each
of these 2 groups separately using procrustes with similarity in order to
perform PCA (on each group) and do some comparisons.
My problem is that the normalization taking place within the algorithm,
depends on the number of landmarks (which is natural),
so I get two mean shapes of different size. Then, I was wondering if
something could be done in order to retrieve two mean
shapes (one for each group) with the same size.

Thanks again,
Sebastian

I do have the same number of landmarks in each
----- Original Message -----
From: "Jonathan.Bailleul" <Jonathan.Bailleul at greyc.ismra.fr>
To: "Sebastián Ordás" <sebordas at unizar.es>
Cc: <vtkusers at vtk.org>
Sent: Friday, January 09, 2004 12:16 AM
Subject: Re: [vtkusers] vtkProcrustesAlignmentFilter problem


> > Sebastián Ordás wrote:
> >
> > Hi,
> > I have a doubt regarding the vtkProcrustesAlignmentFilter
> > functionality. I have two sets of vtkPolyDatas which come from
> > a triangulation on the same object, but having a different number of
> > points. If I align each set as RigidBody
> > I get the same mean shapes (with GetMeanPoints()) , but when using the
> > Similarity option I get two mean shapes of different size. Is it
> > correct to expect getting mean shapes of the same size? If not, how
> > can I estimate the corresponding scaling factor?
> > Thanks in advance.
> >
 > Best,
> > Sebastian
>
>
> Dear Sebastian, (sorry for duplicate mail)
>
> The Procrustes Algorithms works with the assumption that all shapes have
> the same number of points. More than that, it expects that each point is
> implicitely "numbered" in order of appearance, designating the same
> anatomical locus (i.e point 0 will designate the tip of the index in
> every hand, if you have a tset of hands). Alignment will be performed so
> as to minimize a squared sum of distances between points of same number.
>
> So, if you just want to align meshed without correspondence and with
> variable point cardinality, consider using ICP instead (should make the
> trick).
>
> Best regards,
>
> --
> -----------------------------------
> Jonathan BAILLEUL, Doctorant
> GREYC Image - Université de Caen
> http://www.greyc.ismra.fr/~bailleul
>
>