[Insight-developers] Bug in itkShrinkImageFilter

[Bill Lorensen]

Hans and Brad.

I have fixed the shrink filter bugs and checked in the changes
yesterday. There is something strange happening on 64-bit platforms
with the itkShrinkImagePreserveObjectPhysicalLocations test. I'm still
tracking that down.

My checked in fixes are for the Center Location bug
http://public.kitware.com/Bug/view.php?id=8275 and the Streaming bug
http://public.kitware.com/Bug/view.php?id=8245 . I won't close them
until I figure out this 64 bit issue.

Bill

On Wed, Dec 10, 2008 at 12:19 PM, Bill Lorensen <bill.lorensen at gmail.com> wrote:
> Hans,
>
> I'll look at the ShrinkImage problem and also look at Brad's bug.
>
> Bill
>
> On Wed, Dec 10, 2008 at 11:20 AM, Hans Johnson <hans-johnson at uiowa.edu> wrote:
>> Bill,
>>
>> THANKS!  These comments are exactly what I was hoping for.
>>
>> So this now becomes two separate issues:
>> 1) Change multiresolution pyramid filter to use ResampleImageFilter.
>> 2) Make the Shrink image filter preserve object physical locations better,
>> but without using interpolation.
>>
>> ====================
>>
>> I'll be working on a solution to #1 above today, but since this eliminates
>> my immediate need to solve problem #2 with the ShrinkImageFilter I'll likely
>> not be addressing that problem anytime soon.  Could that be assigned to
>> someone else?
>>
>> ====================
>> With regards to point number 2 above:
>>
>> I feel quite strongly that an 8x8 white square in the middle of a 32x32
>> space should retain it's center of mass after the shrink image is applied
>> with a shrink factor of 2,4, or 8.  This is one of the test cases in the new
>> failing test I added last night.  In the current implementation, the center
>> of mass is not preserved.
>>
>> In the case of shrinking a 64x64 image to an 8x8 image, the index locations
>> map as follows
>>
>> Fill     Current    Proposed    Object Space Preserving
>> Output   Input      Input       Input ContinuousIndex
>> [0,0]    [0,0]      [4,4]       [3.5,3.5]
>> [1,1]    [8,8]      [12,12]     [11.5,11.5]
>> ...
>> ...
>> [7,7]    [56,56]    [60,60]     [59.5,59.5]
>>
>> By reading from the current input there is a bias towards shifting the
>> objects towards the origin in the current implementation.  Perhaps the best
>> way to deal with this would be to add the difference between columns 3 and 4
>> in the table above to the origin.
>>
>> * My proposed solution will need to better checking of boundary conditions
>> (seg fault would occur on images of size 60x60 being sub-sampled to and
>> image space of 8x8).
>>
>>
>> Thanks,
>> Hans
>>
>> On 12/10/08 7:21 AM, "Bill Lorensen" <bill.lorensen at gmail.com> wrote:
>>
>>> Hans,
>>>
>>> Another issue if you make your changes. If the new Shrink is run on a
>>> labeled image, I believe your code will interpolate the new pixels.
>>> This behavior is not desirable for subsampling a labeled image.
>>>
>>> Perhaps we should leave this filter alone and change the
>>> multiresolution pyramid code to use ResampleImageFilter with the
>>> proper values set to do subsampling.
>>>
>>> Bill
>>>
>>> On Wed, Dec 10, 2008 at 12:05 AM, Bill Lorensen <bill.lorensen at gmail.com>
>>> wrote:
>>>> Hans,
>>>>
>>>> The documentation for ShrinkImageFilter says that it does simple
>>>> subsampling. So it is not really a bug according to the documentation.
>>>> However, I think your implementation is more useful (I assume it is
>>>> correctly done). You should probably fix the documentation in the .h
>>>> file when you commit these changes.
>>>>
>>>> Also, beware that we have enabled a KWStyle commit check for some
>>>> directories. These directories are Testing/Code/Review, Code/Review,
>>>> Code/IO and Code/BasicFilters. All of these directories are currently
>>>> clear of any style defects.
>>>>
>>>> Bill
>>>>
>>>> On Tue, Dec 9, 2008 at 11:55 PM, Bill Lorensen <bill.lorensen at gmail.com>
>>>> wrote:
>>>>> Hans,
>>>>>
>>>>> Two of the tests are seg faulting.
>>>>>
>>>>> Bill
>>>>>
>>>>> On Tue, Dec 9, 2008 at 11:38 PM, Hans Johnson <hans-johnson at uiowa.edu>
>>>>> wrote:
>>>>>> Luis and Bill,
>>>>>>
>>>>>> I think I've uncovered a bug in the itkShrinkImageFilter that has
>>>>>> ramifications in any use of the itkMultiResolutionPyramidImageFilterTest.
>>>>>>
>>>>>> I was writing a program that aligns the mid-sagital plane with the center
>>>>>> of
>>>>>> an image, and found that as I went from low-resolution estimates to
>>>>>> high-resolution estimates, the center of mass of the object in the
>>>>>> different
>>>>>> scale images would change.
>>>>>>
>>>>>> A new test (currently failing) has been added, and the patch that makes
>>>>>> that
>>>>>> test pass has been included in the Bug report:
>>>>>> http://public.kitware.com/Bug/view.php?id=8275
>>>>>>
>>>>>> This change, however, will cause any of the tests that depend on the
>>>>>> MultiResolutionPyramidImage (and prehaps others) to start failing.  I'd
>>>>>> like
>>>>>> your advice on how to get this bug fix resolved in the least obtrusive way.
>>>>>>
>>>>>> Thanks,
>>>>>> Hans
>>>>>>
>>>>>>
>>>>>
>>>>
>>
>>
>