Hi,
Recently I've run into an issue that is related for instance with
the R-tree. It's caused by the fact that the if the floating-point
coordinates are used, spatial relations of OGC-Geometries like Point
or Segment are checked WRT machine epsilon but in the case of Boxes
the coordinates are compared without taking epsilon into account.
This of course makes the Boxes checks faster. However it also means
that if a Box is a "minimum bounding box/rectangle", i.e. min/max
values are exactly the min/max coordinates of the corresponding
bounded geometries, the spatial relation may give unexpected
results. For instance:
point p(0, 0);
point p_eps(-eps/2, -eps/2);
box b(point(0, 0), point(1, 1));
bg::equals(p, p_eps) == true;
bg::intersects(b, p) == true;
// but
bg::intersects(b, p_eps) == false;
It looks more or less like this:
And this is true for any non-Box geometry, e.g. for two Polygons and
their bounding boxes the results may be similar to the ones showed
above. It could be observed in the R-tree or any other space
partitioning data structure or algorithm using Boxes. In particular,
it means that in some very specific cases R-tree spatial queries may
give different results than spatial relations checks done without
the spatial index.
Have anyone noticed this problem?
Do you think that it is a problem that needs to be fixed?
Now that I've described this problem, do you find some specific case
in your own code that could be affected by it?
I'm asking because it may be acceptible to prune slightly more
nodes/values (so not interesting space) even if the result wouldn't
be prefectly "correct".
To improve/fix it we could check the spatial relations for Boxes WRT
epsilon. However this slows down many algorithms in the library,
e.g. buffer() by 25% and the rtree spatial queries 2x or more which
I think is unacceptable.
Another thing that could be done is to enlarge the bounding boxes
WRT epsilon, in a way consistent with the spatial equality checks
used in the library. This solves the problem and there is no direct
performance loss. I'm saying direct because this would cause that
slightly greater number of nodes could be traversed and values
returned simply because this would be the correct behavior. Though
the enlargement is so small that this shouldn't be noticeable, at
least I haven't noticed any change. So it seems that this is the way
to go. The catch is that the rtree bounds (bounding box) would no
longer have the min/max coordinates exactly the same as the min/max
coordinates of represented Points or Segments stored in the rtree.
Btw, there is no need to enlarge the nodes' Boxes if value Boxes are
stored because the spatial relations are checked the same way for
both. If the value Boxes represented some non-Box Geometries like
Polygons they could probably be enlarged before storing them in the
index. Then the R-tree would simply use those enlarged Boxes.
Do you think that the slightly enlarged R-tree bounds for Point and
Segments could cause some other problems?
Would you prefer if it was possible to enable or disable this
feature?
Would you prefer to have it enabled or disabled by default?
Regards,
Adam