[Insight-users] quick question: discrete or recursive gaussianfilter

Gunnar Farnebäck gunnar at imt.liu.se
Mon Aug 29 07:43:09 EDT 2005


Jim wrote:
> Someone once did an analysis of the two techniques. Not sure where
> that is. I google search may find it in the mail archives.

Well, it would help if the Insight-developers archive wasn't closed.

> I think they compared accuracy vs tims. We should try to track that
> down and put it on the ITK Wiki or maybe a short paper in the
> Insight Journal.

http://www.itk.org/mailman/private/insight-developers/2003-May/004684.html
and followups.

> I think in general, the DiscreteGaussian is more accurate than the
> RecursiveGaussian. I think this is particularly true for small sigma.

Yes, this is the general behavior.

> For large sigma, the RecursiveGaussian is much much much faster. I
> think the RecursiveGaussian is constant time wrt sigma.

Correct. In terms of the number of computations involved the recursive
Gaussian filter is equivalent to a discrete Gaussian filter with a
spatial support of 16 pixels, regardless of sigma. Due to
implementation differences the break-even filter size with respect to
actual speed may be different, however.

/Gunnar


More information about the Insight-users mailing list