If you were going to take Brad's approach with generating the list of
sample points once, you could try using the Subsamplers in Numerics/
Statistics to do the subsampling and then use the sorts available in
itkStatisticsAlgorithm. You could start with
itkUniformRandomSpatialNeighborSubsampler to choose samples within a
specified region of the image (see the corresponding test for an
example). In this approach, each sample is already a
NeighborhoodIterator. As Brad pointed out, it depends on what you're
doing for this overhead to be useful, but I'd be curious to see how
the timing works out and compars to what you've done already.
I'm not sure whether this will be useful for you, but I mention it in
case others are looking for ways to sample the images.
-Kris
Date: Mon, 13 May 2013 11:59:23 -0400
From: Bradley Lowekamp <[email protected]>
Subject: Re: [Insight-developers] fast image pixel access method
To: <[email protected]>
Cc: "[email protected]" <[email protected]>,
"Williams, Norman K" <[email protected]>
Message-ID: <[email protected]>
Content-Type: text/plain; charset="iso-8859-1"
One more thought.
If you are generating the list of sample points once and then using
them multiple times to sample there are perhaps other interesting
things which could be done.
CPU likes to access memory in a linear fashion. Depending on your
density, and the size of the neighborhood etc.. If you access your
random points in a sequential fashion you may see benefits. That is
1) generate a sequential set of random points 2) sample the points
in order. Step 2 should may see some benefits... Now under what
conditions does it out weight the first?
Brad
On May 13, 2013, at 11:49 AM, "Gao, Yi" <[email protected]> wrote:
Hi Brad,
Thank you for the suggestion!
Sorry in my previous code I had the mistake for randomly picking
position for the center point. Now I changed it to be in the inner
region.
index[0] = rg.lrand32(1, size[0] - 2);
index[1] = rg.lrand32(1, size[1] - 2);
index[2] = rg.lrand32(1, size[2] - 2);
And I also turn off the boundary condition by:
it.NeedToUseBoundaryConditionOff();
Now with the same parameters, nbhditerator is around 4 seconds (vs
direct access 5.5 seconds)!
I will modify my code to take advantage of this!
Thank you!
Best,
yi
On Mon, May 13, 2013 at 11:34 AM, Bradley Lowekamp <[email protected]
> wrote:
Hello,
On May 13, 2013, at 11:03 AM, "Gao, Yi" <[email protected]> wrote:
When using NeighborhoodIterator, i'm not sure if the boundary
checking is automatically turned on. I guess if that is by default
on, then there is not much speed gain with it.
Because the region you specify to the Iterator is on the boundary
it will turn on the boundary checking. You can either turn off the
condition:
http://www.itk.org/Doxygen/html/classitk_1_1ConstNeighborhoodIterator.html#a438cb0146b802b04a771a2461952cff8
or you can change the region size you initialize the Iterator with.
You can just use the operator<< to check what the iterator is doing.
Also consider using a large type for the accumulation type, such as
itk::NumericTraits<T>::AccumulatorType;
And consider using std::accumulate(it.Begin(), it.End(), 0 )
Brad
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-developers/attachments/20130513/8b26666b/attachment-0001.htm
>
------------------------------
_______________________________________________
Insight-developers mailing list
[email protected]
http://www.itk.org/mailman/listinfo/insight-developers
End of Insight-developers Digest, Vol 109, Issue 13
***************************************************
_______________________________________________
Powered by www.kitware.com
Visit other Kitware open-source projects at
http://www.kitware.com/opensource/opensource.html
Kitware offers ITK Training Courses, for more information visit:
http://kitware.com/products/protraining.php
Please keep messages on-topic and check the ITK FAQ at:
http://www.itk.org/Wiki/ITK_FAQ
Follow this link to subscribe/unsubscribe:
http://www.itk.org/mailman/listinfo/insight-developers
