Re: On the nature of irregular sets

Hello Bill,

Bill Hibbard wrote:

Hi Ugo,

Nice to hear from you.

For your Exceptions in DelaunayWatson and DelaunayFast,
all anyone could do would be to run under jdb or another
debugger and try to track down what is going on. I don't
have any specific idea about these, and doubt anyone else
does either. Probably best to stick with DelaunayClarkson.

Thanks for the image for your 'curtain-like artefacts'. You
are correct that they occur because the outline of your 2-D
Set of points is concave, not convex, but the Delaunay
algorithm fills in long thin edge triangles to make it convex.
My workaround for the moment was to create exterior points and set their range value (elevation) to NaN. I'm still trying things out. We actually need none of the visualization features (at least not yet; we plan in the near future to combine the former elevation model viewer with our web services). In the old app we got the points already triangulated. We're now combining datasets with different resolutions and need to do the triangulation on the fly. That is, I need to extract the triangles from the delaunay.

The only thing I can think of is an ad hoc algorithm to detect
and remove triangles on the edge (i.e., with an edge not shared
with another triangle: look for -1 values in the Delaunay.Walk
array, I think) and such that that outer, unshared edge, is
very long compared with the triangle's shortest edge.
OK, thanks, I'll keep that in mind.

Such an algorithm would be generally useful, so if you wanted
to contribute it back as a (static?) method of
that would be great. But I understand that you are in business
and may want to keep it proprietary.
True, we're in business, but firm philosophy is one of open source software. Thus, if we come to implementing that, most likely we'll contribute it back to the community.

Thanks for the support.




On Tue, 23 Nov 2004, Ugo Taddei wrote:


I couldn't put the accompanying image on our server yesterday; I'm
attaching it (and hoping this is allowed in this list!).

One sees on the left the "curtains" I mentioned. These are planes which
are very close together and on the edge of the dataset. Intersting is
really one one rotates the image, the planes sort of flicker.

On the right it's the same effect. This is not really a bug, I know (or
think); but it's created when the algorithm joins exterior points
(points on the boundary of the dataset), and which are concave. Is there
a way to avoid that?



Ugo Taddei wrote:


l a t / l o n  GmbH
Meckenheimer Allee 176          53115 Bonn, Germany
phone ++49 +228 732838          fax ++49 +228 732153