[Insight-developers] precision truncation of 'delta' in ResampleImageFilter

Tue Mar 22 11:22:24 EDT 2011

You should look at the non-threaded resample filter code, and check if 
the threaded version is generating identical output.

It might be better if it used ImageFunction::IsInsideBuffer() instead of 
reimplementing those checks,
and it isn't clear to me it is correctly incrementing across the 
continuous space. It doesn't seem to have the boundary calculations right.

For your specific question, the continuous index version of 
IsInsideBuffer, correctly returns whether or not a continuous index is 
inside the image.  It accounts for the half-voxel-width of continuous 
space along each dimension of the voxel.
You can call: ImageFunction:IsInsideBuffer(255.0000002)
and be sure it returns the correct answer.

--Simon

On 03/22/2011 11:06 AM, M Stauffer (V) wrote:
> Thanks Simon. Must the rounding code you mention be used explicitly, or
> is it used under the hood somehow? If this delta value here ends up
> yielding an index of 255.000000002 for a range of [0,255], what will
> happen if it's used in, e.g., ImageFuncition::IsInsideBuffer() ?
>
> Cheers,
> Michael
>
>> -----Original Message-----
>> From: insight-developers-bounces at itk.org
>> [mailto:insight-developers-bounces at itk.org] On Behalf Of Simon Warfield
>> Sent: Tuesday, March 22, 2011 10:16 AM
>> To: insight-developers at itk.org
>> Subject: Re: [Insight-developers] Insight-developers Digest,
>> Vol 83, Issue 38
>>
>>
>> This code looks like it predates the development of consistent
>> imaging bounds, voxel coordinates and interpolators.
>> It looks like it works around issues that have since been resolved.
>>
>> In any case, the consistent voxel coordinates and
>> cross-platform rounding code that we added prior to version
>> 3.20 are the way to address the concern the code comments are
>> expressing.
>>
>> --Simon
>>
>>
>> On 03/22/2011 10:03 AM, insight-developers-request at itk.org wrote:
>>> Date: Mon, 21 Mar 2011 21:50:58 +0000
>>> From: "Johnson, Hans J"<hans-johnson at uiowa.edu>
>>> Subject: Re: [Insight-developers] precision truncation of 'delta' in
>>>           ResampleImageFilter
>>> To: Michael Stauffer<mstauff at verizon.net>, ITK
>>>           <insight-developers at itk.org>
>>> Message-ID:<C9AD3259.715E%hans-johnson at uiowa.edu>
>>> Content-Type: text/plain; charset="us-ascii"
>>>
>>> My changes were almost certainly cosmetic.  I don't
>> recognize this code.
>>> Hans
>>>
>>>
>>> On 3/21/11 3:28 PM, "Michael Stauffer"<mstauff at verizon.net>   wrote:
>>>
>>>> Hi,
>>>>
>>>> I'm cleaning up ResampleImageFilter, removing redundant and
>>>> unnecessary code. I have a question about lines 637-661,
>> copied below. 'git blame'
>>>> shows the code was originally from Dirk Padfield, and most recently
>>>> edited by David Cole and Hans Johnson.
>>>>
>>>> This code truncates values to a set precision to avoid accumulated
>>>> rouding errors while repeatedly adding a small delta. The comments
>>>> here suggest it should be truncating the 'delta' variable, but it's
>>>> operating on 'inputIndex'. As best I can understand, it should be
>>>> operating on 'delta' and not 'inputIndex'. As it stands, 'delta'
>>>> doesn't get truncated to the desired precision at all. 'inputIndex'
>>>> is initialized before this code in order to calculate delta, then
>>>> it's set again in the subsequent while loop, to what looks like the
>>>> same value in the first loop iteration, so this precision
>> adjustment here actually doesn't do anything.
>>>> The code is repeated again later in the subsequent while
>> loop, where
>>>> it operates on 'inputIndex', which looks correct.
>>>>
>>>> Modules/Filtering/ImageGrid/include/itkResampleImageFilter.txx
>>>> Lines 637-661:
>>>>
>>>>    // Delta is precise to many decimal points, but this precision
>>>>    // involves some error in the last bits.  This error can
>> accumulate
>>>>    // as the delta values are added.
>>>>    // Sometimes, when the accumulated delta should be inside of the
>>>>    // image, it will be slightly
>>>>    // greater than the largest index in the image, like
>> 255.00000000002
>>>>    // for a image of size 256.  This can cause an empty
>> column to show up
>>>>    // at the right side of the image. If we instead
>>>>    // truncate this delta value to some precision, this
>> solves the problem.
>>>>    // Therefore, the following routine uses a
>>>>    // precisionConstant that specifies the number of relevant bits,
>>>>    // and the value is truncated to this precision.
>>>>    for ( unsigned int i = 0; i<   ImageDimension; ++i )
>>>>      {
>>>>      IndexValueType roundedInputIndex = (IndexValueType)(
>> inputIndex[i] );
>>>>      if ( inputIndex[i]<   0.0&&   inputIndex[i] !=
>>>> (double)roundedInputIndex )
>>>>        {
>>>>        --roundedInputIndex;
>>>>        }
>>>>      double inputIndexFrac = inputIndex[i] - roundedInputIndex;
>>>>      double newInputIndexFrac = (IndexValueType)( precisionConstant
>>>>                                         * inputIndexFrac )
>>>>                                 / precisionConstant;
>>>>      inputIndex[i] = roundedInputIndex + newInputIndexFrac;
>>>>      }
>>
>> --
>> Simon
>>
>> _______________________________________________
>> 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.html
>>
>> 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

-- 
Simon K. Warfield, Ph.D.
Professor of Radiology
Harvard Medical School
Director of Radiology Research
Director Computational Radiology Laboratory
Department of Radiology
Children's Hospital Boston

http://www.crl.med.harvard.edu

Administrator:
   Laura Alice
   laura.alice at childrens.harvard.edu
   617-355-2755