Le 26/08/2011 20:41, Curtis Jensen a écrit :
Using math 3.0, I have two polygons with many points. One is
completely contained within the other. When I do a difference on the
two, I expected to get a polygon with a hole in it. However, I get 86
polygons, that roughly make up a polygon with a hole in it. If I
scale the points by a factor of 0.1, I get 7 polygons, and if I scale
it differently in the two directions, I get a different number of
polygons. Sometimes the resultant polygons don't seem to make a shape
resembling a polygon with a hole in it.
How should I interpret the results of the difference method? i.e. How
do I process the 86 or 7 or however many polygons so that it resembles
1 polygon with 1 hole in it?
I finally had a look at this.
It seems when the outer polygon is built, it is already split in many
sub-polygons. There are many tiny artefacts throughout the plane. This
can be seen by reading the polygon and immediately retrieving its
boundary. The retrieved boundary is not equal to the initial one!
As an example, you can zoom in the rectangle ranging from -10.4346 to
-10.4344 along x and from 10.6585 to 10.6587 along y. In the original
boundary from src_ccw.csv file, the boundary goes roughly from upper
right to lower left of this small rectangle, with an almost rectangular
dent in the middle. The retrieved boundary dos not have this dent but
instead has a long boundary from upper right to lower left, and about
four almost infinitely thin separate polygons at the dent bottom.
So their is at least a bug in the building routine. Given the very
complex nature of this boundary, I don't know how to identify the bug
and solve it. Here is what I think now:
We have a boundary defined by many points, most of them being aligned
with their neighbors as depicted in the ASCII art below, where thew
characters represent vertices and '-' and '|' represent edges between
vertices.
x-x-x-x-x-x-x-x-x
|
x
|
x x-x-x-x-x-x-x-x-x
| |
x x
| |
x-x-x-x-x-x-x-x-x
We use each edge to define a line, and build the BSP tree by computing
the intersection of these lines. Here, the intersections of the top
horizontal edges are *really* difficult to compute as they are parallel.
This is a very difficult problem because the BSP representation is
really different from the boundary representation we start from. We
don't ask about the connectivity of the edges (mainly because we don't
know how to handle it ...), and in this case this lead the algorithm to
fail miserably :-(
Could you open a Jira issue with a reduced points set reproducing a
similar behavior (say a single polygon with less than 50 points, which
could simply be a chopped sub-region of this one) ?
Do you have some clever idea to use connectivity information to help
build the set (perhaps by trying to eliminate redundant intermediate
points) ?
I guess this bug will be difficult to solve.
Luc
Thanks,
Curtis
Attached are two csv files with the points in CCW order. Also
attached is a plot of the points in the two files. Below is code I
added to the org.apache.commons.math.geometry.euclidean.twod.PolygonsSetTest
class to test with (It uses the Apache Common FileUtils too)
@Test
public void testDifferenceManyPoints() throws IOException {
PolygonsSet set1 = csv2set(new File("src_ccw.csv"));
PolygonsSet set2 = csv2set(new File("inner_ccw.csv"));
PolygonsSet set = (PolygonsSet) new
RegionFactory<Euclidean2D>().difference(set1.copySelf(),
set2.copySelf());
Vector2D[][] verts = set.getVertices();
System.out.println(verts.length);
}
private PolygonsSet csv2set(File file) throws IOException {
List linesObj = FileUtils.readLines(file);
Vector2D[][] verts = new Vector2D[1][linesObj.size()];
for (int i = 0; i< linesObj.size(); i++) {
String line = (String)linesObj.get(i);
String[] tokens = line.split(",");
double x = Double.valueOf(tokens[0]);
double y = Double.valueOf(tokens[1]);
verts[0][i] = new Vector2D(x, y);
}
return buildSet(verts);
}
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
