[ITK-users] Smallest region containing the physical space covered by another image's region

[Matt McCormick]

Hi Cyril,

Excellent, thanks for the update!

Yes, avoiding the circular dependency between ImageBase and
ImageRegion could be tricky.  Please push what you have.  I may take a
shot at getting around the circular dependency.

Thanks,
Matt

On Thu, Apr 16, 2015 at 5:03 AM, Cyril Mory
<cyril.mory at creatis.insa-lyon.fr> wrote:
> Hi Matt,
>
> Thank you for your detailed comments. I fixed most of the issues you spotted
> and amended the patch. Unfortunately, I was unable to move the method to
> itkImageRegion: from there I could not use the itkImageBase methods
> TransformIndexToPhysicalPoint and TransformPhysicalPointToIndex.
>
> I am note familiar with forward declarations, friend classes and how they
> are meant to be used, so maybe it was possible indeed and I just did it
> wrong.
>
> Cyril
>
>
> On 04/15/2015 07:22 PM, Matt McCormick wrote:
>>
>> Hi Cyril,
>>
>> Thanks for contributing the patch!
>>
>> I added a comments for some minor things on Gerrit.  If it is more
>> convenient to do the computations manually than use BoundingBox, that
>> is fine. In terms of where to put it, the ImageRegion class may be a
>> better spot than ImageBase since there already are the PadByRadius,
>> ShrinkByRadius, and Crop methods. In terms of the name, perhaps
>> EnlargeOverBox or similar may be better. This is not a Get method like
>> other Get'ers, we avoid the passive 'ing verb, and Box or something
>> similar indicates we have done the computation in physical space
>> instead of index space, which "Region" may imply.
>>
>> Thanks,
>> Matt
>>
>> On Wed, Apr 15, 2015 at 7:23 AM, Cyril Mory
>> <cyril.mory at creatis.insa-lyon.fr> wrote:
>>>
>>> Hi Matt,
>>>
>>> I have added such a method to itkImageBase, it is called
>>> GetSmallestRegionContainingRegion. The patch is submitted on a branch
>>> with
>>> the same name. I made my best, but I think it will require more
>>> improvements
>>> than the previous one before it can get merged.
>>>
>>> In particular, I haven't used the itk::BoundingBox class, because from
>>> its
>>> description it seems to handle only itk::Point. Should I convert indexes
>>> to
>>> points, run the BoundingBox calculation, and convert back ?
>>>
>>> I'll wait for the comments
>>> Cyril
>>>
>>>
>>> On 04/13/2015 09:12 PM, Matt McCormick wrote:
>>>>
>>>> Hi Cyril,
>>>>
>>>>> I am preparing a patch for the itkWarpImageFilter, and I came to
>>>>> realize
>>>>> that computing the requestedRegion of the inputs from the output's
>>>>> requested
>>>>> region is non trivial.
>>>>
>>>> Great job on the patch submission! :-)
>>>>
>>>>
>>>>> - the displacement in the deformation field can be very large, in any
>>>>> direction, ... therefore there is no choice but to load the full input
>>>>> image
>>>>> into memory (unless we want to go through the whole deformation field
>>>>> to
>>>>> evaluate the maximum displacement, but in a method like
>>>>> GenerateInputRequestedRegion we do not want such heavy processing)
>>>>> => Input image's requested region should be the largest possible region
>>>>
>>>> Yes, this is a good assessment.
>>>>
>>>>
>>>>> - the requested region in the deformation field, on the other hand, can
>>>>> be
>>>>> computed. But it is not trivial. If the output and the DF have
>>>>> identical
>>>>> information (spacing, origin, direction), then the requested region in
>>>>> the
>>>>> DF should be copied from the output. If their information differs,
>>>>> computing
>>>>> the requested region would require a method to compute the smallest
>>>>> rectangular region that contains the physical space covered by an other
>>>>> image's region, the other image having (potentially) a different
>>>>> spacing,
>>>>> origin and/or direction. Does such a method exist somewhere in ITK ?
>>>>
>>>> A possible approach here could transform all corner indexes of the
>>>> output to physical space points, transform physical points to index
>>>> space of the DF, and compute a bounding box of these index points.
>>>>
>>>> The methods to apply could be
>>>>
>>>>     TransformIndexToPhysicalPoint
>>>>     TransformPhysicalPointToIndex
>>>>
>>>> on the ImageBase class and the itk::BoundingBox class.
>>>>
>>>>
>>>> HTH,
>>>> Matt
>>>
>>>
>