[Insight-users] 2nd try: 3D point-set registrations

[Luis Ibanez]

Hi Radu,

1) The process for doing PointSet to Model registration
    with ITK, requires that you define a Metric returning
    a value that measure how well the point set matched
    the model. Once you write such metric, you can use all
    the other components of the registration framework (e.g.
    Transforms, Optimizers...) in order to complete a
    registration process.

2) For pointset to pointset registration without defined
    correspondences you can use the IterativeClosestPoint
    Metric. Note that this metric has been renamed in the
    CVS repository as "EuclideanDistancePointMetric", the
    name change is due to the fact that there is nothing
    iterative in the way this metric is computed.

    For examples on the use of this metric, you may want
    to look at the files

    Insight/Examples/Registration/
             IterativeClosestPoint1.cxx
             IterativeClosestPoint2.cxx
             IterativeClosestPoint3.cxx

     If you are using the CVS version, note that these files
     have been moved to the Patented directory:

           Insight/Examples/Patented

     Due to the similarity between these examples, and an
     existing patent that covers the definition of the
     Iterative Closest Point (ICP) registration method.


3)  It is *NOT* correct to conclude that

       "3D model-to-mesh-points is not possible"

     The correct conclusion is that in order to make
     3D model-to-mesh-points registration, you must
     implement the Metric of your choice and connect
     it to the other components of the registration
     framework.


4)  Using the sum of squared euclidean distance to
     the closest point is a classical Metric for
     PointSet to PointSet registration. You could
     however implement more formal metrics, Eg. those
     considering not only the point distance but also
     the coherence between points in one set and points
     in the other set.

     The simple ICP is usually generating wrong
     correspondences during the first iterations.
     The implicit assumption of the method is that
     despite the wrong correspondences, the intermediate
     steps are still moving the point set in the right
     direction. You could imaging Metrics more elaborate
     than the simple sum of squared distance to closest
     points, that will provide better correspondences
     between points in both sets.




    Regards,


       Luis




------------------
Radu C. Popa wrote:

> Hi,
> 
> Please help in finding the "final" answers to these questions:
> 
> 1. is it possible in ITK to fit-register a 3D **model** (e.g., an ellipsoid)
> to a "cloud" of **points** (i.e., the nodes of a 3D mesh)?
> 
> 2. the same question for a 3D cloud of points to a 3D cloud of points, when
> the points are not given as pairs as per landmark registration
> 
> 
> Is it correct to conclude that?:
> 
> ->1. the 3D model-to-mesh-points is not possible; instead, only the
> model-to-image (gray levels) solution is readily available
> 
> ->2. the solution for mesh-points-to-mesh-points is to use the metric
> IterativeClosestPointMetric
> 
> Thank you in advance.
> 
> Regards,
> 
> Radu
> 
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
> 
>