[Insight-users] point cloud registration

[Luis Ibanez]

Hi Dennis,


You haven't quite read the full Registration Chapter.


It seems that you missed the

            itk::Similarity3DTransform

that is described in section 8.8.14, pdf-page 428,
along with the 2D example of its use in section
8.6.3, pdf-page 387.

A Similarity Transform performs rotation, translation
and scaling without any shearing.


You may find useful to read the entire Registration
chapter.


The Euclidean distance metric will work regardless
of the Transform that you choose because the metric
is computed after the points have been transformed,
so it is always a measure of the residual error after
transformation.



   Regards,


     Luis



---------------------
Dennis Wenzel wrote:
> Hello Luis,
> 
>> You *MUST* read the Registration chapter of the ITK
>> Software Guide
>>
>>       http://www.itk.org/ItkSoftwareGuide.pdf
>>
> 
>> that will help you clarify all the concepts. For example,
>> you will see that you can select any Transformation
>> for performing this registration. It may be rigid, or
>> affine, or even a deformable transform such as BSplines.
> 
> 
> I have been through that chapter.  I see that you have an affine transform
> and a transform for uniform scaling, but I need a transform that only does
> translate, rotate, and scale (no shear).  That should be one with 9
> parameters instead of the 12 which are solved for using the affine
> transform.  How can I get a transform which restricts in this manner?  Am I
> missing something?  If not, do you think it is easy enough to implement?
> But then, I noticed that you have optimizers which can take advantage of
> certain types of transforms.
> 
>>
>>
>> You may want to forget about PointSet-PointSet registration
>> and look more into PointSet to Image registration:
> 
> 
> Actually, I still have a situation I need to support where I have two point
> clouds of the same surface from separate contact probe digitizations.  So,
> it seems I will still need to do the pointset-to-pointset registration.  I
> have a situation where I may never have an "image" to which to register.
> 
> With only a Euclidean distance metric available for pointset-to-pointset
> registration, will this work for the case that I have non-uniform scaling?
> 
> Thanks...
> 
> Dennis
> 
> 
>