[Insight-users] Re: Insigh-journal: KNN classifier

[Christopher Wyatt]

Mark,

I regularly use the KNN implementation on single channel (scalar) images
up to 512x512x450 on a machine with 4GB of RAM. However, I often have to
split the image up and apply the KNN to the regions separately and then
reassemble the image. This changes the classification results of course,
but with spatially distributed sampling (e.g. every other voxel) you can
often get acceptable results with less of a memory footprint.

-- chris.
clwyatt at vt.edu

> Date: Mon, 02 Jul 2007 17:20:31 +0200
> From: Mark Bouts <mark at invivonmr.uu.nl>
> Subject: [Insight-users] Insigh-journal: KNN classifier
> To: insight-users at itk.org
> Message-ID: <468917BF.5010200 at invivonmr.uu.nl>
> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
> 
> Hi,
> 
> in the insight-journal an article and its implementation on the
> KNN-classifier can be found.The article: Computational Geometry
> Computation and KNN Segmentation in ITK
> (http://insight-journal.org/InsightJournalManager/view_reviews.php?back=public
> ations.php%3Fjournalid%3D5%26order%3Drating%26revision_display%3Dcombined&pubi
> d=94) 
> describes a means for segmenting images based on 2D feature data.
> Currently, I'm working on an implementation which exceeds the 2D feature
> set. However I seem to get stuck at the memory allocation for the
> voronoifilter. I can't seem to figure out why this allocation should be
> such an issue. (Right now a map of elements should be created of
> 256x256x256 = 16777216 elements) Has anyone tried to extend this code to
> larger nD (for instance n=3)? Or has anyone suggestions of finding a
> good and simple solution to this problem. Even pointing me to another
> implementation of the Knn-classifier would be of great help!
> 
> Cheers!
> Mark
>