[vtk-developers] [EXTERNAL] Software vs hardware interpolation

[Kilgus, Thomas]

Thank you guys very much for the feedback! 

@David: As you guessed, we use a vtkPolyDataMapper (including vtkActor and a vtkTexture) inside our application to render the slices in 2D and 3D (due to historical reasons). In fact, I tried to employ your new image mappers last year, but run into some issues with our own interaction and further requirements related to our own applications. Hopefully, we will have resources to investigate more in adapting our parts for your mappers in the future. I would really like to use them, but unfortunately right now it's not possible for us.

@Kenneth: Adding another validation image could be an interesting option. Where exactly can I find the mentioned examples?

Regards
Thomas


-----Ursprüngliche Nachricht-----
Von: David Gobbi [mailto:david.gobbi at gmail.com] 
Gesendet: Mittwoch, 21. März 2012 16:22
An: Kilgus, Thomas
Cc: vtk-developers at vtk.org
Betreff: Re: [vtk-developers] [EXTERNAL] Software vs hardware interpolation

Hi Thomas,

What mapper are you using?  The vtkImageMapper?  A vtkActor with a
vtkTexture?  Or a vtkImageActor?  All three use different OpenGL
paths.

>From your images, I'm going to guess, like Kenneth, that you are using
a vtkTexture.  If OpenGL only supports power-of-two textures, the
vtkTexture code will resample its input to the nearest power of two,
i.e. a 5x5 texture is resampled (in software) to 8x8.  If OpenGL
supports non-power-of-two textures, the 5x5 texture is displayed
directly.

Back when I used vtkTexture to display images (eight years ago or so),
I always padded my images up to a power of two and remapped the
texture coordinates, specifically to avoid this unwanted resampling
step in vtkTexture.  Afterwards, I switched to vtkImageActor which
handles non-power-of-two images just fine.  Recently, I wrote a whole
new set of image mappers which are included in VTK 5.10.

 - David


On Wed, Mar 21, 2012 at 8:21 AM, Moreland, Kenneth <kmorel at sandia.gov> wrote:
> I don't claim to be an expert here, but the interpolation method (nearest
> vs. linear) is controlled via an OpenGL flag to glTexParameter.  If VTK is
> not properly setting this flag, this could be a bug, but a look at
> vtkOpenGLTexture looks as if the flag should be set correctly.
>
> Traditionally (circa 15 years ago) the nearest lookup was considered faster
> and the linear lookup was considered more accurate.  Nowadays, graphics
> cards have special interpolating hardware in their texture units that allows
> them to do the linear interpolation with no extra time cost.  It could be
> that the driver for this particular hardware is doing the linear
> interpolation anyway since there is no added time cost.
>
> VTK has been dealing with rendering difference between software and hardware
> rendering as well as differences between different hardware architectures
> for, like, ever.  Thus, the VTK testing framework accepts multiple valid
> images for any test.  You just add image files with the same name but with
> _1, _2, etc. added to the filename.  (There are lots of examples in the
> existing baselines.)
>
> So, the easy solution is just to check in a new image into the baseline
> image repository.  The real question is, how much do you really care that
> you are getting linear instead of nearest interpolation?  Do you have a use
> case other than this pathological test?
>
> -Ken
>
> From: "Kilgus, Thomas" <t.kilgus at Dkfz-Heidelberg.de>
> Date: Wed, 21 Mar 2012 14:24:39 +0100
> To: "vtk-developers at vtk.org" <vtk-developers at vtk.org>
> Subject: [EXTERNAL] [vtk-developers] Software vs hardware interpolation
>
> Hi,
>
>
>
> as mentioned in my previous post
> (http://vtk.1045678.n5.nabble.com/software-rendering-td5557078.html ), I
> observed different interpolation behaviors in software and hardware
> rendering of images in VTK.
>
>
>
> Here you can see an example:
> http://cdash.mitk.org/testDetails.php?test=2749290&build=107877
>
>
>
> The test shows the result of an image rendered with software rendering. The
> input 5x5 pixel image is scaled (interpolated) to 8x8. If you render the
> same image with hardware support, a nearest neighbor interpolation will be
> applied and the result will look like the ValidImage on the bottom of the
> page. I don't think that any interpolation can be considered wrong or right,
> but I think both, software and hardware rendering, should have a common
> default behavior.
>
>
>
> Please can anyone comment on this? Is there any possibility to enforce the
> same interpolation in software rendering? For reliable testing in different
> environments it is necessary both rendering methods act with the same
> default behavior, since it is not deterministic whether your machine is able
> to render with hardware support or with software rendering. (For instance a
> VPN connection established to a windows machine blocks the GPU for some
> reason, Virtual Machines show different behavior etc.)
>
>
>
> Regards
>
> Thomas
>
>
>
> _______________________________________________ Powered by www.kitware.com
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html Follow this link to
> subscribe/unsubscribe: http://www.vtk.org/mailman/listinfo/vtk-developers
>
>
> _______________________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html
>
> Follow this link to subscribe/unsubscribe:
> http://www.vtk.org/mailman/listinfo/vtk-developers
>
>