[Insight-users] Mean of transformation parameters

[Luis Ibanez]

Hi Eve,


1) No we can't guarantee that transform would not deviate from the mean.
    If the optimizer has unstable parameters, a transform can end up very
    far from a reasonable registration.

    If you need to guarrantee a certain bound, then, that restriction
    must be added into the Transfrom or into the Optimizer.


2) You probably want to explore the concepts of robust-estimation,
    http://en.wikipedia.org/wiki/Robust_statistics
    http://mathworld.wolfram.com/RobustEstimation.html
    where you first get rid of outliers, and then you use the remaining
    results for generating a robust mean that is representative of the
    stable population.

3) If you use a set of Affine transforms, you are correct with the
    intuition that you can only averaget their parameters *if* they
    are using the same center of rotation.

    If the center of rotation is different on every case, then you
    should compute the average of the offset, and the average of the
    rotation matrix. In this way, the center of rotation will be fused
    into the offset computation.



   Regards,


       Luis


-----------
Eve wrote:
> Hi Users,
> 
> Given N sets of transformation parameters, can we guarantee that any one set
> would not deviate much from the mean set? Say, if one obtains N registration
> results based on stochastic optimization, how can one report the average
> result? Further, if an affine transform is used, how can we compute the mean
> parameters? Is it true that we simply compute the mean, as long as the
> center of rotation is fixed in N registrations? What if the center of
> rotations are different in each case?
> 
> Thanks,
> Eve
>