PolygonBuilder::findShell assertion 'shellcount <= 1' failed [JTS fails!]

[strk]

See:

 http://lists.osgeo.org/pipermail/geos-devel/2011-February/005099.html

[strk]

r3269 adds an XML version of this test.

[strk]

Found the difference: JTS is doing this:

    PlanarGraph.linkResultDirectedEdges(nodes);

right in the PolygonBuilder::add function. GEOS has that call commented out. WHen commenting it in JTS we get the same assertion failure...

[strk]

Uhm, false alarm, the GEOS code is indeed inlining the equivalent of that function

[strk]

The first real difference between GEOS and JTS is at the point where a minimalEdgeRing infact dimensionally collapsed is considered as an hole by JTS and not by GEOS.

I attach a snapshot showing the two MinimalEdgeRing? geometries. You can see the blue one collapsed.

[strk]

The collapsed geometry according to GEOS, in HEXWKB form:

0102000000040000000000000000000000841D588465963540F56BFB214F0341408F26B714B2971B40F66BFB214F0341408C26B714B2971B400000000000000000841D588465963540

Haven't tested if JTS considers coordinates of this CW or CCW

[strk]

I've checked that JTS gets to the _exact_ same HEXWKB, so obviously JTS considers that thing to be CCW while GEOS does not.

[strk]

Ok, JTS's version is _not_ exactly the same. Thanks to robe for finding out, as I was getting mad:

GEOS:
0102000000040000000000000000000000841D588465963540F56BFB214F0341408F26B714B2971B40F66BFB214F0341408C26B714B2971B400000000000000000841D588465963540
JTS:
0102000000040000000000000000000000841D588465963540F56BFB214F0341408F26B714B2971B40F66BFB214F0341408E26B714B2971B400000000000000000841D588465963540
---------------------------------------------------------------------------------------------------^ difference here

Now I'm even more curious about whether JTS considers the former clockwise too...

[strk]

in any case, turning the assertion failure in a robustness exception triggers engaging of snapping policy which results in an output with the collapsed "ring" removed, which passes the expectances of the tester and seems better for me too.

[strk]

Now this is funny. I've modified the testcase to expect GEOS output (the one with the collapsed ring removed) and JTS reacts to that new testcase with the shellcount <= 1 assertion.

It's funny because the input did not change, only the expected output changed.

I attach the new testcase here, in case Martin wants to take a look.

Will commit the new testcase and the exception solution, which I belive makes sense, being this clearly a robusness issue.

[strk]

the new test, same input and op as old, different expected result, failing JTS

[strk]

Fixed in r3286

[strk]

For the record: JTS agrees with GEOS about the ring computed by GEOS being CW:  http://sourceforge.net/mailarchive/message.php?msg_id=27379769

Our r3288 puts both rings under regression testing.

[warmerdam]

With the error report now being an exception it is necessary for the related logic to no longer be run only when NDEBUG is not defined. Fixed in trunk (r3373).