Re: GSoC 2017 - Implement boolean operations on polygons

[Philip Nienhuis]

John Swensen wrote:
> > On Mar 17, 2017, at 1:53 PM, Philip Nienhuis <address@hidden
> > <mailto:address@hidden>> wrote:
> >
> > John,
> >
> > John Swensen wrote:
> > > > On Mar 17, 2017, at 2:56 AM, Juan Pablo Carbajal
> > > > <address@hidden <mailto:address@hidden>
> > > > <mailto:address@hidden>> wrote:
> > > >
> > > > On Fri, Mar 17, 2017 at 9:41 AM, PhilipNienhuis
> > > > <address@hidden <mailto:address@hidden>
> > > > <mailto:address@hidden>> wrote:
> > > > > Juan Pablo Carbajal-2 wrote
> > > > > > On Thu, Mar 16, 2017 at 9:56 PM, piyushjain &lt;
> > > > >
> > > > > > piyushjain1sa@
> > > > >
> > > > > > &gt; wrote:
> > > > > > > Sir, I went through the current geometry package development and also
> > > > > > > read
> > > > > > > about mkoctfile command, and how to use oct create and use oct-files.
> > > > > > >
> > > > > > > So, can you tell me what are the milestones of the project? I
> > > > > > > mean,going
> > > > > > > step-wise, what do I need to implement?
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > --
> > > > > > > View this message in context:
> > > > > > > http://octave.1599824.n4.nabble.com/GSoC-2017-Implement-boolean-operations-on-polygons-tp4682161p4682435.html
> > > > > > > Sent from the Octave - Maintainers mailing list archive at
> > > > > > > Nabble.com.
> > > > > > So you know what you want to pursue already?
> > > > > > 1. Cleanning up the boost interface
> > > > > > 2. Adding clipper interface, read this
> > > > > > https://savannah.gnu.org/patch/?func=detailitem&item_id=9000#options
> > > > >
> > > > > From my memory, the Clipper interface should be about done.
> > > > > I'm still busy with making it clip polylines (works now) and
> > > > > interpolating Z
> > > > > values (got stuck there); but for "ordinary" 2D polygons I think it
> > > > > just
> > > > > works as it is.
> > > > >
> > > > > Philip
> > > > >
> > > > >
> > > > >
> > > > > --
> > > >
> > > > In that case I suggest you provide patches to integrate Philip work
> > > > into geomtry, following the guidelines described in that savanah
> > > > thread.
> > > > I was planning to do this in Octconf so I will give you quick feedback
> > > > if you provide something beginning of next week.
> > >
> > > One final comment. I know I pushed the Boost solution because at the
> > > time I thought it had the most robust solution for self-intersecting
> > > polygons, polygons with parallel lines, etc. I now feel a little bad
> > > that we spent all that time and it hasn’t been incorporated into the
> > > package. It is quite cumbersome to get Boost set up to compile with the
> > > version that supports self-intersection correction. I think that more
> > > work could be done to make it a much better solution by making a “Point
> > > concept abstraction” that would allow it to operate on the Octave
> > > variable data in-place without having to copy. For larger datasets, this
> > > could be a huge memory (and potentially performance) improvement over
> > > the other solutions where you have to transform the Octave data into the
> > > format required by the library.
> >
> > I have been trying the Boost version by Amr for some time, using Amr's
> > unofficial geometry-3.0.0 version from github and the mapping package
> > (also an unofficial version, but I'm the mapping pkg maintainer :-) ).
> > AFAICS it does work fine, reliably and fast.
> > Also the poly2fv function is very useful for graphical output.
> >
> > We (package maintainers JuanPi and me) didn't know that Boost is so
> > big. Later on Amr mentioned he has weeded it out but AFAICS that info
> > is nowhere.
> > FYI Boost (an enormous collection of include files) alone is > 10 %
> > disk space of an entire mxe-octave windows installer. That wouldn't be
> > bad in itself, but the fact that it's used for just a few specialized
> > functions puts that into another perspective.
> >
> > Where you write "boost with self-intersection correction", I suppose
> > that's version 1.61 and up, and it would be desirable to be able to
> > have polybool be based on that version as well, depending on compile
> > time availability through configure.
> > So a GSOC project could be bringing the polybool/boost solution
> > further so that it can be included in OF-geometry.
> > I would really like that to happen.
> >
> > > Has anyone ever explored the license for GPC
> > > (http://www.cs.man.ac.uk/~toby/gpc/
> > > <http://www.cs.man.ac.uk/%7Etoby/gpc/>
> > > <http://www.cs.man.ac.uk/%7Etoby/gpc/>)? It says it is free for
> > > private/hobbyist/education purposes, but I don’t know whether the
> > > license is compatible with the GPL and Octave. From my experience, GPC
> > > is one of the fastest and most robust polygon clipper out there. Maybe
> > > we overlooked it because of its split license and the fact that the
> > > license isn’t very well defined?
> >
> > The GPL explicitly does not limit use of free SW. "Free for just
> > education purposes" is incompatible with the GPL.
> >
> > There are several benchmarks out there, some or most of them outdated
> > now, that do not quite put GPC at #1. But I think many of those
> > benchmarks are a bit "colored" as they usually relate to a specific
> > polygon SW solution.
> > At my work we've been using it once (with matlab), my impression of
> > that one -probably atypical- occasion was that GPC isn't so very fast.
> > Therefore I'm glad you share with us your experience with several
> > polygon clipping libraries, I think that is very valuable, thanks.
> >
> > Maybe we could find out what GIS SW like QGIS uses internally?
> >
> > Philip
>
> I think QGIS uses GEOS [1]. I had never really even investigated that
> one because this website [2] said that is was significantly slow for
> larger datasets.
>
> One of the problems with the Boost self-intersection correction is that
> it is broken again as of 1.63. I had been under the impression they were
> going to try and make it “functional but unreleased” from 1.61 and
> onward. That appears to not be the case.

So boost is a little unstable as well...
Anyway such issues related to specific versions can be catched in a 
configure script.

> I have been meaning to dig into the Boost::Geometry examples at [3] to
> see if I can implement an iterator facade for a custom polygon object
> that allows us to operate directly on the octave data input without
> having to copy all the points into the Boost::Geometry model for points,
> rings, polygons, and multi-polygons. I think it should be possible. It
> may speed up execution because we aren’t copying a bunch of double
> values, but it should most certainly reduce memory usage a ton. If we
> could push the performance/resources even better than now, it might be a
> compelling reason to keep Boost.

(Maybe I do not fully understand your aim.)
For any polygon library the input data will have to be morphed somehow 
from the application/use case at hand into a format that the polygon 
library expects; vice versa for output.
I'm unsure if investing a lot of work in specialized data access in some 
.oct file is better than clever and vectorized preprocessing at the 
Octave side. The latter is much more flexible and adaptable to the use 
case at hand.

The Clipper library currently comprises a .mex wrapper file (in the 
Third party subdir, in recent Clipper releases pointing to Matlab 
Central).
While working on polyline clipping I noted there's quite some code 
required to copy info from/to the structs that constitute the IMO rigid 
MEX data I/O structures.  I suppose some of that code can be avoided, or 
reduced, by rewriting the .mex wrapper into a proper .oct file using 
octave_value objects.  That'll give more flexibility as well. I suppose 
even the double<->int64 preprocessing can be absorbed in the .oct file 
(currently I've put that in an .m file wrapper).

> One other option is to use the Boost-provided bcp utility to create the
> subset of Boost headers that is are needed to compile Boost::Geometry
> and include it in the package release. This is still 30MB uncompressed,
> but it is much smaller than the whole Boost. this would also allow us to
> distribute a version that has the self-intersection correction.

IIRC Amr mentioned he had brought down the boost stuff to around 11 or 
12 MB.
I think the Boost include files shouldn't be in the geometry package 
directory; they'd rather live in /usr/include.

Philip