Dear all,
Ullrich Koethe wrote:
>
> Roland Richter wrote:
>
> > Simple example:
> > ===============
> >
> > Given an image in pixel x pixel coordinates, provide a view in "realworld
> > coordinates", for instance in millimeters x millimeters or in polar
> > coordinates, i.e. millimeters x degree.
> >
> > [...]
>
> This might be more difficult than it looks at first. If you provide an
> index that rescales the coordinates, data requests will in general *not*
> fall exactly on pixel locations. Therefore, you must interpolate gray
> levels at the requested locations. This at least changes the return type
> of operator[] to a (non-writable) value parameter. As long as you
> perform only linear interpolation, views and iterator adapters are
> equally suitable, but cubic interpolation is easier realized with a view
> plus an index object.
Yes, you are completely right on that. Sorry, I failed to tell the "whole
story". Here it is:
What we had to implement was:
1) Provide an interpolated view (bi-linear) of the image.
2) Provide a view of the interpolated image which is in "real world"
coordinates, i.e. mm x mm. Due to distortions in the camera picture,
the transformation pixel x pixel --> mm x mm was non-linear.
3) Apply rotation and translation to this image (parameters given in
"real world" coordinates).
4) Convert the result back to pixel x pixel for a region of interest.
To maintain the original code, written in C, was simply a nightmare;
refactoring it using the data/view concept helped a *lot* (and caused
a significant performance increase).
Since I was required to produce MSVC++ 6.0 compileable code, I wrote
a) a view doing the interpolation
b) a view doing index transformations (as described in my last mail)
c) two classes representing translation and rotation
in a rather simple, straight-forward manner, and it worked.
Interpolation, of course, is a rather special need.
However, I think this might be generalized. An interpolation view
- does index transformations (for instance, rounding doubles to ints)
- acts on more than one element, typically on elements in a "neighbourhood"
- combines these elements to give the result,
I guess concepts like "neighbourhood" or views of "local data"
are not limited to image processing (examples?)
All that might be included within a rather general view, perhaps
as an extension of the current transform_view. (I still don't know a
proper name for that -- suggestions?)
Question is, how to describe a "neighbourhood"?
1) as a sliding window, similar to a window_view in VTL
2) invloving expression templates
1) is simpler, but a window only describes a linear range, not
an arbitrary, multi-dimensional neighbourhood
2) seems to be much more complicated
I guess I'll have a further look at Blitz++. There, "stencils"
seems to be something similar as neighbourhoods, and the concept
of "index placeholder" might be useable, too ...
- Roland
--------------------- Object Oriented Numerics List --------------------------
* To subscribe/unsubscribe: use the handy web form at
http://oonumerics.org/oon/
* If this doesn't work, please send a note to owner-oon-list@oonumerics.org
This archive was generated by hypermail 2b29 : Wed Feb 20 2002 - 03:20:15 EST