[Insight-developers] Image-based Testing in Tcl

[Lorensen, William E (Research)]

I think we should store the regression "image" in one file if possible. I'm not sure whcih format
will store multi-dimensional data other than raw.

That will have to be multi-dimensional. I'd like to show a "representative difference" on the
dashboard. This might be the most "different" slice with one scalar pixel value. As Jim suggested, we
coild also report other difference information.


-----Original Message-----
From: Miller, James V (Research) 
Sent: Wednesday, September 04, 2002 12:07 PM
To: 'Brad King'; Lorensen, William E (Research)
Cc: Insight Developers
Subject: RE: [Insight-developers] Image-based Testing in Tcl


Do we need all that complexity in the DifferenceImage.  The pixel
shifting code in VTK's regression code was to get around the fact
that OpenGL is not pixel specific.  

In ITK, we might be able to do a straight "difference".

The other question is how do we want to measure the difference 
between two pixels.  Grayscale images are not a problem. Color images
should probably use a distance in Luv space.  Vector images ????

(Note that VTK's image difference uses a (somewhat) normalized difference
in RGB space. We can use a more precise metric for ITK since we are 
not rendering.)

Also we might want to report "how many pixel differed" as opposed
to the average "difference" between pixels.  



> -----Original Message-----
> From: Brad King [mailto:brad.king@kitware.com]
> Sent: Wednesday, September 04, 2002 11:40 AM
> To: Lorensen, William E (Research)
> Cc: Insight Developers
> Subject: RE: [Insight-developers] Image-based Testing in Tcl
> 
> 
> Bill,
> 
> > I've been thinking about this for a bit. I would like the 
> "regression
> > iamge" to store the n-d image. The difference filter might 
> report how
> > different each 2D slice is. Then somehow we would report the 2d
> > difference image.
> 
> The DifferenceImageFilter supports fully N-d difference 
> images.  It uses a
> standard iterator in the test image and a neighborhood iterator in the
> valid image to allow the test to tolerate n-pixel shifts (n 
> can be set by
> the test).  We will need to choose a few things:
> 
> How big should the standard 2-D test image be?  (256x256?)
> How big should the standard 3-D test image be?  (64x64x64?)
> How big should the standard N-D test image be?  (32^N ?)
> What file format should be used for 2-D images?  (png?)
> What file format should be used for 3-D images?  (png per slice?)
> What file format should be used for N-D images?  (png per slice?)
> How do we present difference images for 3-D and N-D failed tests?
>   (choose slices with largest difference?)
> 
> I'm sure there are many more questions, but this is a start.
> 
> Also, where should we store the input and valid images?  
> Currently we're
> using Insight/Testing/Data.  Should we use VTK's approach of a separte
> testing data tree?  The valid image for the curvature flow 
> test is about
> 20KB (256x256 8-bit greyscale compressed).  If most 
> regression images are
> this size, it would be fine (IMO) to put them in the source tree.
> However, N-d images can get pretty big...
> 
> -Brad
> 
> _______________________________________________
> Insight-developers mailing list
> Insight-developers@public.kitware.com
> http://public.kitware.com/mailman/listinfo/insight-developers
>