[Insight-developers] ImageToImageFilter: dimension reductionp roblem

[Miller, James V (Research)]

I just checked in a series of changes to the architecture that allow
the output of a filter to be a different dimension than the input.
The difficulty in supporting this is that 

ImageToImageFilter::GenerateInputRequestedRegion()

needs to be able to compile regardless of whether the input and output 
dimensions match. The default implementation of this routine tries to
call

input->SetRequestedRegion( output->GetRequestedRegion() );

This line of code will not compile if the input region and the output 
region are not the same dimension.

In the new model, the line above is now

InputImageRegionType inputRegion;
m_RegionCopier(inputRegion, output->GetRequestedRegion() );
input->SetRequestedRegion( inputRegion );

m_RegionCopier is an ivar that is a function object. The call operator
dispatches to an overloaded function that is in NOT a member function 
of ImageToImageFilter but rather is a standard function in a separate
namespace (thanks Brad!).  Thus, the RegionCopier dispatches to one routine
if the input dimension matches the output dimension, another if the input
dimension is less than the output dimension, and a third if the input
dimension is greater than the output dimension.  The default implementations
are rather simplistic.  If the input dimension is greater than the output 
dimension, the output region information is copied to the first part of the 
input region.  Zeros fill the remainder of the StartIndex and Size.  If the
input dimension is less than the output dimension, the first part of the 
output region is copied to the input region.

If a filter needs a behavior different from this (for instance the
ExtractImageFilter should support a different behavior), the filter can 
pass a different function object to ImageToImageFilter.  This would
probably be done in the constructor of the filter.  It would call the protected
method SetRegionCopier().

The function object would have to provide the call operator and dispatch
to a different set of routines to handle the different behavior.

To support these operations, there is a new file called 
itkImageToImageFilterDetail.h that has a series of types defined to facilitate
dispatching to different routines based on dimensions.

Brad also pointed out that the dispatch pattern currently deployed in ITK
in a few places will have a problem with explicit instantiation.  I'll be
converting the current dispatch implementations to this new model.

Wilson, you should be able to implement your filter to remove an image dimension.
I would suggest that functionality be placed in the ExtractImageFilter.





-----Original Message-----
From: Wilson Chang [mailto:wmcst6+@pitt.edu]
Sent: Thursday, April 04, 2002 4:21 PM
To: Miller, James V (CRD)
Subject: Re: [Insight-developers] ImageToImageFilter: dimension
reductionproblem


Does it look like a fix to the base classes will be doable?  If it isnt, I
could do what luis was suggesting by having the dimension reduction filter
inherit from process object.

thanks,
wilson


----- Original Message -----
From: "Miller, James V (CRD)" <millerjv@crd.ge.com>
To: "'Luis Ibanez'" <luis.ibanez@kitware.com>; "Wilson Chang"
<wmcst6@pitt.edu>
Cc: "Insight Developers" <Insight-developers@public.kitware.com>
Sent: Monday, April 01, 2002 8:36 AM
Subject: RE: [Insight-developers] ImageToImageFilter: dimension
reductionproblem


> I think there are a few ways that we can handle this.
>
> There already is a non-templated Region base class so that
> may be an avenue.  Another option is to function overloading
> to force a default implementation when the dimensions
> match and require subclasses provided a method when the
> the dimensions do not.
>
> I'll put together a fix.  Wilson, could you send me the
> snippet of your test program (or check in the test without
> putting it in the TestDriver yest) that is causing this
> error (just to save me from reinventing the wheel).
>
>
>
> -----Original Message-----
> From: Luis Ibanez [mailto:luis.ibanez@kitware.com]
> Sent: Friday, March 29, 2002 5:15 PM
> To: Wilson Chang
> Cc: Insight Developers; Miller, James V (CRD)
> Subject: Re: [Insight-developers] ImageToImageFilter: dimension
> reduction problem
>
>
>
> Hi Wilson,
>
> The problem does not seem to be related with the
> run-time behavior of the virtual functions.
>
> What seems to happen is that regardless of how you overload
> the method in your code, the code of ImageToImageFilter<TI,TO>
> has to be compiled anyways and since the TInputImage
> and TOutputImage that are being used are not compatible
> the error will always appear at compilation time.
>
> In fact, the streaming itself is built on the assumption
> that dimension match.  (Jim, please correct me if I'm wrong here)
> The only way around would be to have a RegionBase class, non
> templated from which Region derives, and then do the dimension
> checking at run time as opossed to compile time. The drawback
> here is the we will lose the performace of inlining Region related
> code and got the lower speed of virtual calls...
>
> Another option could be to add another ImageToImageFilter
> (with another name...) which just does not have methods
> relying on matching dimensions between the input and
> output.
>
> This is similar to the problem we encounter when we
> generalize the Transforms. Everything was fine until we
> got to the Perspective transform where the dimension
> of the output points do not match the dimension of the
> input points.... The solution taken in that case was to
> add the TransformBase (non-templated) class and
> afford virtual calls each time a point is transformed.
>
> One very short term solution for you could be to not derive
> from ImageToImageFilter but from ProcessObject.....
> but that may probably be only a temporary solution.
>
>
>    Luis
>
> ====================================
>
> Wilson Chang wrote:
>
> > Hi, We have been working on a visualization project with itk and vtk
> > that requires having all of our images in 3-D or less.  Therefore, I
> > have been working on a filter that reduces the number of dimensions of
> > an image.  This filter inherits from the ImageToImageFilter class.
> > The snag that I have run into is the GenerateInputRequestedRegion
> > function in the ImageToImageFilter class:
> >
> >
> >
> >
> >
> > template<class TInputImage, class TOutputImage>
> > void
> > ImageToImageFilter<TInputImage,TOutputImage>
> > ::GenerateInputRequestedRegion()
> > {
> > ...
> >
> >
input->SetRequestedRegion(this->GetOutput()->GetRequestedRegion());
> > ...
> >
> >
> >
> > This line of code basicly requires that the input and output image
> > have the same number of dimensions.  When I try to override this
> > GenerateInputRequestedRegion in my filter, I still get a compile error
> > saying:
> >
> >
> >
> > cannot convert from 'const class itk::ImageRegion<2>' to 'const class
> > itk::ImageRegion<3>'
> >
> >
> >
> > Is it possible that the superclass' definition of this function is
> > being called regardless of whether i've redefined this function?
> >
> >
> >
> > ideas?
> >
> >
> >
> > thanks,
> >
> > wilson
> >
>
>
> _______________________________________________
> Insight-developers mailing list
> Insight-developers@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-developers
>