Hi folks,
Let’s get together again to discuss progress on the ITK Performance efforts.
When: Friday, October 20th, 11 AM EST
Where: Google Meet https://meet.google.com/nue-tugx-pxt
What: Discuss ITK Performance progress
Notes from the previous discussion on 2017-07-28 can be found in the weekly confab document. We can also add notes there.
There was an interesting conversation held on the ITK Community mailing list regarding the time it takes to initialized the image registration framework:
https://public.kitware.com/pipermail/community/2017-September/013662.html
The conclusion was that a significant amount of time is consumed by the slow DiscreteGuassianImageFilter, the SmoothingRecursiveGaussianImageFilter can provide improved performance, but slightly ( insignificant? ) different results.
I would also like to note, that when a gaussian filter, then a “shrink” filter is done, more efficiency can be gained by doing the gaussian in 1D, then shrink in 1D, repeat. Consider a case where you are shrinking by 10x10x10, This amount of work done by the gaussian can be reduced by over 100x.