Find a suitable replacement for Triangle

Question

Find a suitable replacement for Triangle

kbevers opened this issue 7 years ago · comments

Triangle is a very nice library but it in terms of licensing it is horrible. A new gridding library would allow us to simplify the install-script and let us distribute a pre-build package.

Alternatives to Triangle includes:

Qhull, python bindings
MeshPy
FLANN (we would have to change gridding strategy completely, since it is inverse distance weighted interpolation)

We need to figure out if any of the alternatives are usable. Are they fast enough? Producing the desired results? How big a change to the existing code-base is needed? Any license issues?

Peter Limkilde Svendsen · Answer 1 · Mon Jan 29 2018 17:50:42 GMT+0800 (China Standard Time)

Qhull is an easy option, being used behind the scenes in the Delaunay and griddata functions in SciPy. Used in this way, it can triangulate 1 million random points in about 10 seconds.

Another interesting option could be porting the delaunator JavaScript library (ISC-licensed), which can apparently triangulate 1 million points in about 2 seconds. (Not to be confused with the Delaunator python library.)

The very fastest library I've heard of is fqDel (paper, slides), though no source seems available. The authors report being able to triangulate 1 million points in 0.2 seconds, a 4x speed improvement over Triangle.

Kristian Evers · Answer 2 · Mon Jan 29 2018 18:23:11 GMT+0800 (China Standard Time)

We have experimented with Qhull before. At the time it wasn't as robust as Triangle and also slower. Things might have changed since, but I don't think so.

Another thing to this issue is that it would be preferable to use an existing Python module, so we limit the amount of code we have to maintain and also avoid compiling binaries as part of the installation Just a thing to keep in mind.

Peter Limkilde Svendsen · Answer 3 · Fri Feb 02 2018 17:55:42 GMT+0800 (China Standard Time)

I have investigated a bit more. The Geogram C++ library seems very suitable for our purposes:

Small, standalone Delaunay module with no mandatory dependencies outside standard library
BSD (3-clause) license
Fast and robust (I have successfully triangulated 100M random points in ~3 CPU minutes), time seems to scale roughly linearly with number of points

We'll just need to generate Python bindings for it.

CGAL (under GPLv3) can also handle at least 75M points (tested on an Ubuntu machine), but it is a fairly heavy library using lots of C++ templates, and does not seem to be anywhere as fast as Geogram without a bit of optimization.

Qhull is indeed not as fast, and complains when fed more than 16.7M input points.