Comments on: Fast Convex Hull Algorithm

By: Luigi Giaccari

Luigi Giaccari — Tue, 15 Dec 2009 19:55:54 +0000

Oh you seem pretty expert!

can you please send me the paper, I would like to give it a look!

Thanks

By: JackM

JackM — Tue, 15 Dec 2009 18:51:05 +0000

Yes, this is nothing new but just a variation of Graham’s scan (FYI, you don’t need to calculate angles in Graham’s scan). If you are looking for fast convex hull methods, there is for example a paper “Optimal Output-Sensitive Convex Hull Algorithms in Two and Three Dimensions” by Timothy Chan, which explains a method which run in O(n*log(h)) time.

By: Luigi Giaccari

Luigi Giaccari — Sat, 26 Sep 2009 11:10:50 +0000

My routine is faster if all points are already part of the convex hull.
Run the following test:

N=1000000;
teta=linspace(0,2*pi,N);
x=cos(teta);
y=sin(teta);
 
tic
chull1=ConvHull2D(x,y);
fprintf('ConvHull2D took : %4.4f s\n', toc);
 
tic
chull2=convhull(x,y);
fprintf('Matlab Routine: %4.4f s\n', toc);

it leads to:

ConvHull2D took : 0.6692 s

Matlab Routine: 5.5856 s

The built-in uses the quick hull algorithm this point cofiguration is patological.

My routine suffers logspace data where generally the convhull is formed by only a few points. On this data the filtering step erases no points. Anywa it should be faster in this case too, since ConvHull2D doesn’t call the unique fucntion which is the main bottlenech for the Matlab convhull.

We also have to consider that Convhull2D, apart from the sort function, is totally m-coded, no precompiled files. Convhull instead is a MEX routine.

If you need more infos just ask.

Luigi

By: Taywin

Taywin — Fri, 25 Sep 2009 19:03:15 +0000

I’m trying again… Last time it submitted an empty text…

Have you every tried to compare for the case when all points are vertices of convex hull as well? I mean all 1 million points are actually vertices of a convex hull. Because your algorithm speed seems to rely heavily filtering. If there is a case that your filtering does not benefit the speed at all, I just want to know whether the built-in function is still slower than yours or how much faster it is compared to yours. Of course, this is an extreme case, but I just want to see how it goes. Thank you.

By: Luigi Giaccari

Luigi Giaccari — Fri, 18 Sep 2009 21:35:55 +0000

Umh,

You should really study
In Graham scan there is a sorting operation based on the angle others points forms with the first one. Here sort is performed on y (or x ) coordinate, plus there is a filtering preprocessor step.
I actually thing this algorithm is not something new since it seems just too easy. Anyway it works, it is very fast.

Please report comments on http://www.advancedmcode.org/
I don’t usually go on Youtube.

Last question, no I don’t have no code about incremental convex hull, this one was coded for fun, I don’t work in the convex hull field…..Sorry

For any more question just ask,

Bye

Luigi

By: r0l0kcs

r0l0kcs — Fri, 18 Sep 2009 21:28:42 +0000

This looks like Graham Scan to me.
May I ask: Do you have any sources on incremental convex hull algorithms “for dummies” like me? That would be great. Or is Graham Scan the same as incremental convex hull?