Re: [Boost-users] [boost.geometry] buffer distance strategies

Subject: Re: [Boost-users] [boost.geometry] buffer distance strategies
From: Barend Gehrels (barend_at_[hidden])
Date: 2014-11-24 13:39:51

Hi Grzegorz,

gchlebus wrote On 24-11-2014 11:06:
> Hallo Barend,
>
> 2014-11-13 22:07 GMT+01:00 Barend Gehrels [via Boost] <[hidden email] 
> </user/SendEmail.jtp?type=node&node=4669423&i=0>>:
>
>     Hi Grzegorz,
>
>
>     gchlebus wrote On 13-11-2014 18:39:
>>     Hi Barend,
>>
>>     2014-11-12 22:21 GMT+01:00 Grzegorz Chlebus <[hidden email]
>>     <http://user/SendEmail.jtp?type=node&node=4669068&i=0>>:
>>
>>
>>
>>         2014-11-12 22:07 GMT+01:00 Barend Gehrels [via Boost]
>>         <[hidden email]
>>         <http://user/SendEmail.jtp?type=node&node=4669068&i=1>>:
>>
>>             Hi Grzegorz,
>>
>>
>>             gchlebus wrote On 12-11-2014 21:52:
>>>             2014-11-08 11:53 GMT+01:00 Barend Gehrels [via Boost]
>>>             <[hidden email]
>>>             <http://user/SendEmail.jtp?type=node&node=4669040&i=0>>:
>>>
>>>                 Gu Grzegorz,
>>>
>>>                 gchlebus wrote On 8-11-2014 2:33:
>>>>
>>>>                 2014-11-07 19:30 GMT+01:00 Barend Gehrels [via
>>>>                 Boost] <[hidden email]
>>>>                 <http://user/SendEmail.jtp?type=node&node=4668825&i=0>>:
>>>>
>>>>
>>>>                     Hi Grzegorz,
>>>>
>>>>                     gchlebus wrote On 6-11-2014 19:08:
>>>>>
>>>>>
>>>>>                     2014-11-04 17:23 GMT+01:00 Barend Gehrels [via
>>>>>                     Boost] <[hidden email]
>>>>>                     <http://user/SendEmail.jtp?type=node&node=4668774&i=0>>:
>>>>>
>>>>>
>>>>>                         Hi Grzegorz,
>>>>>
>>>>>
>>>>>>
>>>>>>                         2014-10-29 23:18 GMT+01:00 Barend Gehrels
>>>>>>                         [via Boost] <[hidden email]
>>>>>>                         <http://user/SendEmail.jtp?type=node&node=4668617&i=0>>:
>>>>>>
>>>>>>
>>>>>>
>>>>>>                             gchlebus wrote On 24-10-2014 16:44:
>>>>>>
>>>>>>                             > Hi,
>>>>>>                             >
>>>>>>                             > I am wondering whether it would be
>>>>>>                             possible to achieve anisotropic
>>>>>>                             buffering
>>>>>>                             > (distances in neg x, pos x, neg y,
>>>>>>                             pos y can have different values) of a
>>>>>>                             > polygon using the buffer function
>>>>>>                             with custom-implemented distance
>>>>>>                             strategy.
>>>>>>                             > What I want to achieve is presented
>>>>>>                             on the figure 2-b in the following
>>>>>>                             > paper:
>>>>>>                             >
>>>>>>                             http://itcnt05.itc.nl/agile_old/Conference/mallorca2002/proceedings/posters/p_molina.pdf
>>>>>>                             >
>>>>>>                             > I would be grateful to hear from you
>>>>>>                             whether it is doable, and if positive,
>>>>>>                             > how one could implement such a
>>>>>>                             custom distance strategy.
>>>>>>                             The current distance strategy has
>>>>>>                             (currently) no means to get the angle,
>>>>>>                             or a vector of the new point to be
>>>>>>                             buffered. We can consider adding that.
>>>>>>
>>>>>>                             However, by writing custom strategies
>>>>>>                             for join, side, point (for
>>>>>>                             point-buffers) and possibly end (for
>>>>>>                             line-buffers) you should be able to
>>>>>>                             create this, because these have this
>>>>>>                             information.
>>>>>>
>>>>>>                             Attached a program doing similar
>>>>>>                             things with polygons and points (I vary
>>>>>>                             the distance based on angle - you
>>>>>>                             will have to do something with your
>>>>>>                             anistropic model).
>>>>>>
>>>>>>                             The output is also attached.
>>>>>>
>>>>>>                             The program defines three custom
>>>>>>                             strategies, all based on the same
>>>>>>                             mechanism, to create interesting output.
>>>>>>                             I did not do the end-strategy but
>>>>>>                             that would look similar, you can look
>>>>>>                             at the provided end-strategy (round)
>>>>>>                             and apply the same function.
>>>>>>
>>>>>
>>>>>
>>>>>                         gchlebus wrote On 31-10-2014 18:13:
>>>>>>                         I really appreciate your example code, it
>>>>>>                         helped me a lot. Attached you can find my
>>>>>>                         source code.
>>>>>>                         In my implementation of the anisotropic
>>>>>>                         buffering I didn't know how to make use
>>>>>>                         of the distance strategy, as it was
>>>>>>                         possible to make it work using only side
>>>>>>                         and join strategies.
>>>>>>                         I encountered strange behavior when
>>>>>>                         changing number of points describing a
>>>>>>                         full circle. Using 360 points produced a
>>>>>>                         good output, whereas 90 points caused
>>>>>>                         only the second polygon to be buffered
>>>>>>                         (see attached figures). I would be
>>>>>>                         thankful if you could help me to resolve
>>>>>>                         this issue as well as for any remarks to
>>>>>>                         my code.
>>>>>>
>>>>>
>>>>>                         I could reproduce this. Basically the
>>>>>                         join-strategy should always include points
>>>>>                         perp1 and perp2 (these are the two points
>>>>>                         perpendicular to the two sides which the
>>>>>                         join-strategy joints). Either they are
>>>>>                         re-calculated, or they can be just added
>>>>>                         to begin and end. So I did the last
>>>>>                         option, and that piece of code now looks like:
>>>>>
>>>>>                               double const angle_increment = 2.0 *
>>>>>                         M_PI / double(point_count);
>>>>>                               double alpha = angle1 - angle_increment;
>>>>>                         *range_out.push_back(perp1);**// added
>>>>>                         *      for (int i = 0; alpha >= angle2 &&
>>>>>                         i < point_count; i++, alpha -=
>>>>>                         angle_increment)
>>>>>                               {
>>>>>                                 pdd v = getPointOnEllipse(alpha);
>>>>>                                 Point p;
>>>>>                         bg::set<0>(p, bg::get<0>(vertex) + v.first);
>>>>>                         bg::set<1>(p, bg::get<1>(vertex) + v.second);
>>>>>                         range_out.push_back(p);
>>>>>                               }
>>>>>                         *range_out.push_back(perp2);**// added*
>>>>>
>>>>>                         My sample code of course also suffered
>>>>>                         from that, so I added it there too if I
>>>>>                         use it in the future.
>>>>>
>>>>>                         I tested your algorithm with various
>>>>>                         points and distances and it now seems
>>>>>                         always OK.
>>>>>
>>>>>                         You ask for remarks on your code: it looks
>>>>>                         good ;-) one thing, many terms are
>>>>>                         recalculated such as pow(xPos*tan(alpha),
>>>>>                         2)); or just tan(alpha), I usually store
>>>>>                         these into variables, to avoid expensive
>>>>>                         recalculations of the same terms, though
>>>>>                         maybe they are optimized by the compiler.
>>>>>
>>>>>                         Regards, Barend
>>>>>
>>>>>
>>>>>                         P.S. this list discourages top-postings
>>>>>
>>>>>
>>>>>                     Hallo Barend,
>>>>>
>>>>>                     I corrected the join strategy, but still the
>>>>>                     buffering doesn't work in all cases as
>>>>>                     expected. When using xPos = 1, and other
>>>>>                     values equal 0, the buffered polygon contains
>>>>>                     a hole (see xPos1.svg), whereas setting xPos
>>>>>                     to 2 produces a correct result (xPos2.svg). Do
>>>>>                     you know how to fix it? I attached also
>>>>>                     main.cpp, as I changed the code a bit and it
>>>>>                     contains the polygon for which causes the
>>>>>                     strange behavior.
>>>>>
>>>>
>>>>
>>>>                     That is most probably caused by an error in
>>>>                     some of your calculations:
>>>>
>>>>                     The line y = sqrt(yPos2 * (1 - pow(x, 2) / xNeg2));
>>>>                     causes a NAN for this input:
>>>>
>>>>                     alpha about PI
>>>>                     then xNeg2 = 0.010000000000000002
>>>>                     and x = -0.10000000000000002
>>>>                     and yPos2 = 0.010000000000000002
>>>>
>>>>                     This adds a weird line containing NAN to the
>>>>                     join, causing the buffer process fail.
>>>>                     I got this using these parameters:
>>>>                     double xPos = 1.0, xNeg = 0.1, yPos = 0.1, yNeg
>>>>                     = 0.1;
>>>>
>>>>                     and not the parameters you have (that was fine
>>>>                     for me).
>>>>
>>>>                     I think you should make the calculations
>>>>                     full-proof first...
>>>>
>>>>                     For example add a line in the join-strategy:
>>>>                     std::cout << i << " "<< angle1 << " " << angle2
>>>>                     << " " << v.first << " " << v.second << std::endl;
>>>>
>>>>
>>>>                     Regards, Barend
>>>>
>>>>
>>>>                 Thanks, I'll try to improve my calculations.
>>>>                 By the way, I was playing with different strategies
>>>>                 combinations and I found out that when using only
>>>>                 boost buffer strategies:
>>>>                 double points_per_circle = 36;
>>>>                 double distance = 130;
>>>>                 bg::strategy::buffer::distance_symmetric<double>
>>>>                 distance_strategy(distance);
>>>>                 bg::strategy::buffer::end_flat end_strategy;
>>>>                 bg::strategy::buffer::point_circle
>>>>                 point_strat(points_per_circle);
>>>>                 bg::strategy::buffer::side_straight sideStrat;
>>>>                 bg::strategy::buffer::join_round
>>>>                 joinStrat(points_per_circle);
>>>>
>>>>                 the buffer function can still fail (produce no
>>>>                 output) when the distance is higher than 128 (e.g,
>>>>                 128, 130, 150, 300, 400). But this happens up to a
>>>>                 certain value, where the buffer function starts
>>>>                 producing a correct output (e.g., distances 900,
>>>>                 1000).
>>>>
>>>
>>>                 Hmm, I see (starting at different values, but I can
>>>                 reproduce).
>>>
>>>                 I created a ticket, will be looked at. Thanks for
>>>                 reporting.
>>>                 https://svn.boost.org/trac/boost/ticket/10770
>>>
>>>                 Barend
>>>
>>>
>>>             Hi Barend,
>>>
>>>             I'm glad that I could help.
>>>             Anyway, I fixed the bug with NAN, but still when using
>>>             (e.g. xNeg = 1, other 0) the buffer produces no output.
>>>             I am really wondering, how it could work on your machine.
>>>             I printed the values used by join and side strategies
>>>             and they seem to be fine (no NANs or other strange
>>>             values) - see attached log.txt and updated main.cpp used
>>>             to produce the log file.
>>>
>>>             I've compiled my code using msvc 12.0 and gcc 4.8.
>>
>>             Which branch or version of Boost do you use?
>>
>>             In the meantime, I managed to fix the bug you reported,
>>             and it is committed today.
>>
>>             Sorry for my question, but could you try it again with
>>             the latest branch "develop" from github ?
>>
>>             Regards, Barend
>>
>>
>>
>>         Hi Barend,
>>
>>         sure, I downloaded the geometry lib from the develop branch,
>>         but still buffer returns an empty geometry (using xNeg = 1,
>>         others 0).
>>         I've ran my code also for big distances, and it works
>>         perfectly. Good job!
>>
>>         Best,
>>         Grzegorz
>>
>>
>>     I made today an interesting observation. It seems that the buffer
>>     function has a problem when the extents (xNeg, xPos, yNeg, yPos)
>>     are zero. Taking my last example, setting xNeg to 1 and others to
>>     1e-3 makes the buffer work.
>
>
>     Sure - thanks for sending me, that saves me time to find it out.
>     Buffer-distance 0 is not supported (yet). Some GIS packages use
>     that to clean a polygon, but we have designed dissolve for that.
>     We might support it later but currently, indeed, as you found out,
>     it will not work. I hope small distances are OK for your
>     application, for now.
>
>     Negative distances should work if used in the distance-policy,
>     making the polygon smaller, but I don't think that it will work
>     currently out of the box if you mix negative and positive
>     distances in one join-strategy...
>
>     Regards, Barend
>
>
> with small adjustment I made also shrinking possible, but I had to do 
> this in a not very clean way. I had to add to the join strategy 
> additional bool parameter in order to know whether shrinking or 
> expanding should be performed. Do you have any idea how one could find 
> out inside the join strategy whether the structure is being shrunk or 
> expanded? For example in side strategy one of the parameters is 
> distance strategy, which can tell me whether it's negative or positive.

Yes but these strategies are used differently. join-strategy is only 
called for convex joins. If a join is concave, internally another piece 
is created, there is no strategy for that (from user-perspective you 
don't see anything because the sides intersect and the concave piece 
disappears).

Now for a shrink/deflate operation, the polygon is walked reversily, so 
counter-clockwise (if your polygons are clockwise). What is then a 
concave configuration becomes convex, and v.v. so your join-strategy is 
then only called for concave turns, creating the join on the inside. The 
convex turn gets then nothing visible.

For your operation I believe it becomes more complex then because the 
angle is different. In the next version we will add some more info to 
the apply method of the strategy such that you can more easily detect 
this, and that probably will also solve this problem.

Besides that, I have no problem to add the distance-strategy there too, 
that is a good idea and makes the strategies more similar.

>
> Btw, I'm just wondering how complex it would be to allow zero distances?

Yes, that is complex but in the future we will work on it. But that 
won't be next version.

Regards, Barend