[vtk-developers] vtkImageData, C versus Fortran ordering

[David Gobbi]

On Wed, Jul 19, 2017 at 10:55 AM, Marcus D. Hanwell <
marcus.hanwell at kitware.com> wrote:

> On Wed, Jul 19, 2017 at 12:35 PM, David Thompson <
> david.thompson at kitware.com> wrote:
>
>> Hi Marcus,
>>
>> > We work on Tomviz, and one of its primary goals is to work with volumes
>> in an intuitive way making heavy use of the volume rendering code. It also
>> needs to interact with several libraries that operate on volumes, ideally
>> using our Python NumPy integration and zero copy approaches to process data
>> in place.
>> >
>> > One of the challenges has been preserving the Fortran ordering of the
>> single component scalar data. We will also be adding support for
>> multi-component data soon. What are people's thoughts on potentially adding
>> some API to vtkImageData to mark it as C ordered?
>> >
>> > That would largely just geometrically transform how the volume is
>> rendered, but ensure a contour is correctly extracted and aligned. It would
>> then let us use the default C ordering, simplifying our application code,
>> and the current order could remain the default. I would suggest considering
>> the same for the vtk.js work too.
>> >
>> > It is feasible to make these changes in the application logic too, but
>> this feels like a fairly common problem that others would also hit from
>> time to time. Maybe I am missing more corner cases, so I am posting this in
>> hopes of feedback before writing any code (or asking someone else to).
>>
>> I think a common problem would be that many image filters choose a
>> particular order to process data because access is much faster in one plane
>> than others. Are you thinking that those filters would also adjust their
>> algorithms/traverse order to accommodate FORTRAN ordering when it is
>> present on the input (and presumably output)?
>
>
> My ideal scenario is to mark a vtkImageData as C ordered (default it to
> Fortran/what it is now). It may be easier to manage more of that at an
> application level though too, I was more thinking out loud and wondering if
> other people had explored this or solved it in other ways. I get the
> historical perspective, and possibly that what I want may not be needed by
> the wider VTK community.
>
> I would like to leave as much of VTK alone as possible, and so let most
> VTK stuff continue as it always has, but be able to pass a vtkImageData to
> a 3D NumPy array as C ordered, operate on it, and then render the result.
> We currently go to great pains to ensure if comes back as Fortran ordered,
> and there is no way to set a default. I also find that most stuff assumes
> the C ordering, so it can require some reshuffling of the data which can
> involve large volumes.
>

It's already possible to add extra flags to any VTK data set, without
making any changes to VTK itself.

The simplest way is to add an array to the vtkFieldData of the data set.
You could add a field data array called "ImageOrdering" that would hold the
flag.

Another way (a little trickier) is to add your own information key.  For
example, for my VTK DICOM pipelines I've defined a key
called PATIENT_MATRIX that stores a 4x4 matrix to define the image
orientation.

 - David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/vtk-developers/attachments/20170719/e0b054e2/attachment.html>