Best optimizer for landmark registration with only 3 points

[Charlotte Curtis]

I'd like to use landmark registration to get a rough first guess prior to
image registration. When I run IterativeClosestPoint2 with 4 landmarks, it
complains about the number of landmarks:

vnl_least_squares_function: WARNING: unknowns(6) > residuals(4)
vnl_levenberg_marquardt: Number of unknowns(6) greater than number of data
(4)
Solution = [0, 0, 0, 0, 0, 0]

I'm just wondering if there's an optimizer that would let me get away with a
minimum of three landmarks. Since 3 points would define a unique location
in 3D space, it seems to me that there shouldn't need to be 6. Indeed, if I
simply duplicate all of my landmarks and fool it into thinking it has 8
unique points, it comes up with a reasonable solution. This is my backup
plan if nothing else works, but it seems a little messy. Any suggestions?
Thanks,

Charlotte

[Alireza Salamy]

Hi Charlotte

If you only have Three points, you could rather use the class:

itkLandmarkBasedTransformInitializer


if you do not have the correspondence between the points, well...
this methods is so fast with only Three points that you could
actually try all the combinations and settle for the one with
the smallest final squared errors,otherwise your own suggestion of duplicating the landmarks should work.

I hope it works.
Alireza

Charlotte Curtis <curtisc-dxuVLtCph9gsA/***@public.gmane.org> wrote:
I'd like to use landmark registration to get a rough first guess prior to image registration. When I run IterativeClosestPoint2 with 4 landmarks, it complains about the number of landmarks:

vnl_least_squares_function: WARNING: unknowns(6) > residuals(4)
vnl_levenberg_marquardt: Number of unknowns(6) greater than number of data (4)
Solution = [0, 0, 0, 0, 0, 0]

I'm just wondering if there's an optimizer that would let me get away with a minimum of three landmarks. Since 3 points would define a unique location in 3D space, it seems to me that there shouldn't need to be 6. Indeed, if I simply duplicate all of my landmarks and fool it into thinking it has 8 unique points, it comes up with a reasonable solution. This is my backup plan if nothing else works, but it seems a little messy. Any suggestions? Thanks,

Charlotte _______________________________________________
Insight-users mailing list
Insight-users-***@public.gmane.org
http://www.itk.org/mailman/listinfo/insight-users



---------------------------------
Get easy, one-click access to your favorites. Make Yahoo! your homepage.

[Charlotte Curtis]

Thanks for the suggestion, Alireza. I actually tried the
LandmarkBasedTransformInitializer, and the result wasn't very good, which
kind of surprised me as it seems that with the same landmarks it should
converge on approximately the same solution as the iterative closest point
method. I do have correspondence between the points - does the
TransformInitializer not use the correspondence? That could explain the
difference in the two solutions.

Thanks for your help,

Charlotte

Post by Alireza Salamy
Hi Charlotte
itkLandmarkBasedTransformInitializer
if you do not have the correspondence between the points, well...
this methods is so fast with only Three points that you could
actually try all the combinations and settle for the one with
the smallest final squared errors,otherwise your own suggestion of
duplicating the landmarks should work.
I hope it works.
Alireza
I'd like to use landmark registration to get a rough first guess prior to
image registration. When I run IterativeClosestPoint2 with 4 landmarks, it
vnl_least_squares_function: WARNING: unknowns(6) > residuals(4)
vnl_levenberg_marquardt: Number of unknowns(6) greater than number of data (4)
Solution = [0, 0, 0, 0, 0, 0]
I'm just wondering if there's an optimizer that would let me get away with
a minimum of three landmarks. Since 3 points would define a unique location
in 3D space, it seems to me that there shouldn't need to be 6. Indeed, if I
simply duplicate all of my landmarks and fool it into thinking it has 8
unique points, it comes up with a reasonable solution. This is my backup
plan if nothing else works, but it seems a little messy. Any suggestions?
Thanks,
Charlotte _______________________________________________
Insight-users mailing list
http://www.itk.org/mailman/listinfo/insight-users
------------------------------
Get easy, one-click access to your favorites. Make Yahoo! your homepage.<http://us.rd.yahoo.com/evt=51443/*http://www.yahoo.com/r/hs>