Invalid (?) buffer output [JTS works]

[yabo]

Hi,

I'm facing a case where a buffer returns an empty polygon while I believe it shouldn't.

I attached a C++ program demonstrating the problem. I tried it against geos 3.2.2 and geos-svn r2999.

The problem is in the negative buffer of a polygon. For a given value it works perfectly (image attached) but if in increase this value by one it returns an empty polygon which, given the picture, it shouldn't (or I believe so at least).

[yabo]

Minimal testcase including the WKB of the polygon.

[yabo]

In blue the original polygon, in pink the last working negative buffer.

[strk]

How big is "one" ? Ie: which distance value for the working buffer ?

[yabo]

I'm not sure to understand the question. I'm using no SRID (ie, -1), so it's "one" as in the length of 'LINESTRING(0 0, 0 1)', or in this case the difference between -9501 and -9502 for the buffer. Compared to the size of the geometry, 'one' is very small, almost un-noticeable (area of the geometry: 72083121486).

[strk]

Bug confirmed. JTS 1.12 doesn't have this problem.

[strk]

OffsetCurveSetBuilder::addPolygon thinks the polygon shell would be completely eroded by the buffer distance.

[strk]

It turns out GEOS, contrary to JTS, was using an heuristic based on MinimumDiameter to figure if a buffer would completely erode a ring or not. Disabling that heuristic (as done in JTS) fixes this case. But unfortunately it breaks 1 over 100 buffer tests we have in TestBufferExternal2.xml (JTS passes that as well).

.. I'm going on with debugging it ..

[strk]

r3291 fixes this. The other test failing was another missing heuristic based on envelope dimensions. GEOS regressions testsuite is excellent ! :)