Instead of cleaning polys after clipping, to get good triangulation, try to
clean them before clipping, snap the verticies to 1 1cm x 1cm grid, and
merge any resultant slivers back into adjacent polys.
This is very similar to how fgfs-construct worked.
Although the previous methos only failed on 3 airports, the results weren't
very good. There were lots of cracks, and mmissing triangles, as the polys were
cleaned.
New methos freezes on two airports during base triangulation, and crashes on four.
I think the quality of the generated airports overcomes the slight regression.
I also added some debug switches so I don't have to keep rebuilding from source
to debug crashes. --dump-rwy=xxx --dump-pvmt=xxx, --dump-feat=xxx and --dum-base=1
can be used to create shapefiles to debug the particular polys before clipping.
Currently, there needs to be a cp_debug directory to place the shapefiles into.
Probably need better debug infrastructure going forward...
NOTE: I turned off runway shoulder generation in this build. Polygon snapping for runways,
helipads, and shoulders not implemented yet. This may help on the failing airports.
This is next on my todo list.
Just three airports crashed, and no hangs. One of the crashes is still NZSP
I don't know if I'll fix this soon. It's the south pole problem with stuff
crossing the IDL.
- Added dome debug support for generating shapefiles from TGPolygon.
This allows directly creating shapefiles instead of chopping a poly, then
having to run poly2ogr on it. The API is flexible, allowing multiple
datafiles, and layers open simutaneously, and the features in the shapefiles
have a description field that can be set.
- Added a restart-id= which allows a bash script to check the exit value of genapts,
and automatically restart on the airport after the one that crashed.
I've included an example script I have used.
- Stability fixes:
1) actually, this is both stability, and graphical fix. Some airport linear
features and polygon contours were using bezier nodes for straight lines.
When the control point lies just outside the line, the parser would attempt to
draw a 'hook' unfortunately, some of my calcs return NAN, and sometimes you just
get crazy looking polys to attempt to clip - not good.
Fix is to check cubic and quadratic nodes to see if they should really be linear.
Substitute linear segments where apropriate.
2) Along the same line as 1, There are some really short bezier segments, that were
getting subdivided into the hard coded 8 segments. Now, if the approximately linear
distance of the curve is less than 4 meters, the curve is broken into 1/2 meter segments.
Likewise, if the curve is greater than 800 meters, the curve is broken into 100 meter
segments.
3) Sometimes, during clipping, degenerate ontours are created. Although these are cleaned
before triangulation, there were still issues with generating the accumulation poly with
the union operator, causing a crash. I now simplify, and reduce degeneracy on all polys
before clipping against the accumulator.
- Extra special bonus - successful completeion of an airport is printed in green :)
- ran through the last commit, and only 10 airports crashed / locked up.
- EICL, SSPK, and LFBT fix was to simplify the base polygon during cleaning
reduce_degeneracy() could also be used to fix the issue
2 polygons clipped against each other shared a vertex. This causes the
resulting contour to have a duplicate point with different line segments.
Luckily, clipper's polygon simplification cleans this up by adding a
hole contour. In the above airports, triangulation suceeded, but I'm
not 100% convinced it will work for all cases. reduce_degeneracy() will,
at the expense of modified geometry (very noticible)
- EGLW and SWCH (I think - can't read my own writing) were fixed by expanding
the user defined boundary by 2 meters for the base, and 5 meters for the
clearing. These 2 airports had the user defined boundary touching the exact
airport layout. When the base was diffed with the accumulation buffer, the
resulting polygon was NIL.
Starting a whole new stability test from the beginning to check for
regressions. Down to 5 airports with isses left.
- created a new function that will take the higher elevated dup
(as most are ar -9999.0 at initialization. Use the one with real elevation data)
Fixes UMNB
- found a bug in the polygon library: remove_small_cycles.
If the cycle wrapped around the end of the point list, it went undetected.
I created a new detection method, that is slower, but handles this case
FIXES DNAK
- Added a new polygon cleaner function remove_spikes
triangulation library doesn't like when segments for close to 0 degree
angles. When such an angle is seen (<.1 degree), the point is removed
FIXES LFKJ (linear features) and YBAF (pavements and base polys)
- I also modified the runway shoulder code to use 11+ meter width.
I was having problens with polys very close to this leaving a gap.
The code needs revisiting, since I now see z-fighting betwen the shoulder
and grass border.
From this point on, all my checkins need to get past the previous checkins
airport. I am now parsing past LFSB
- When adding intermediate nodes, keep linear feature nodes list seperate from the rest, as they are
seperate accumulation buffers.
- Revert expand poly by 5cm. It broke more airports than it fixed :(
- Going back to GPC as well.
Feature Fixes
- finally got around to the last segment in closed linear features.
I think I am feature complete now.
Remaining issues.
- I still see terrain holes between the base and clearing polys when an airport traverses a tile
boundary. I'm not sure where to start here, as genapt doesn't really know about the tile boundary.
(Is the clearing part of the tile construction? - need to look into this)
)
- used to expand pavement polygons so they overlap slightly (5cm)
This prevents slivers being formed between pavements when designers
add their own taxiway shoulders, and try to line them up instead of
placing the boundary 'under' the taxiway.
- also useed to create the default airport base. This allows the base
to follow the taxiways better - allowing concave boundaries, instead
of the previous convex hull.
Memory leaks squashed. Fixed several memory leaks reported by valgrind.
- GenApt850 can now run forever (or at least until it segfaults)
Very long linear segments in linear features are now broken up to
avoid triangulation errors
Added back some clipper boolean operations (diff and union). These can
be used instead of the GPC functions by using tgPolyGonDiffClipper and
tgPolygonUnionClipper. GPC had crashed in a few airports when clipping
linear features. Using the clipper routines don't crash. More investigation
is needed. It may be something about the polygons being clipped more that
the clipping action itself. In these problem airports, there are some
linear features extending far from the airport.
( I had thought those issues were fixed :( )
More warning cleanups.
success / failure. I have some local changes to the lib to support it, but
they aren't verified. Checking in the support, that always returns success.
- allow clipper and GPC libraries to be included together.
I need clipper for expanding polys, but I'm still using GPC for clipping.
- break up and expand shoulder polys, so they overlap runways.
This keeps cracks from forming between the shoulder and runway
do to roundoff error
- Add verbose option to polygon_tesselate functions to aid in debugging
can set at runtime to produce triangulation output for just 1 poly.
- add remove_small_cycles to poly cleanup routines before triangulation
- more warnings cleanup in ClosePoly.cxx
- add more output in normal execution path, as some airports can go for
hours without printing any status. Hard to tell if it's locked up
- use ogr-decode with new option --texture-lines to generate polygons
with texture info
- possible fix for triangulation errors when triangulating each poly by
itself.
(It's worked for me, but more testing needed)
- Still using GPC for clipping. CLIPPER adds time, and looks to have
more errors. More testing needed here as well.
Simply edit the #define in polygon.h to choose clipping library.
Major TODOs:
1) There is 1 tile (with the most heavy OSM data) that blows up, and I get huge
road textured polys all over the place. Must be an overrun somewhere.
2) Lake / Ocean smoothing. The current algorithm handles all nodes at once. I'm
going to need to modify for the new structure
3) neighbor tile matching. This actually doesn't look so bad without it implemented.
The old method would add points in the triangulator. If a point was added on the
tile boundary, a t-junction would be created. Points are no longer added in triangulation,
so the polygon splitting in ogr-decode should make the points equal on each side of
the tile boundary. (I haven't seen any t-junctions, at least)
4) cleanup - this is a bit of a mess. I really want to refactor to make the code more understandable.
- moved poly_extra from genapts to polygon library (add intermediate nodes)
- change TGPolyList to arrays of superpolys and texparams in preperation for
generating tex coordinates for line data
This was harder than expected as the shoulders were often partially
covered by runway polygons. This was caused by inaccuracies introduced
by the calculation of dimensions from the runway center.
Now we create shoulders if set in apt.dat. If no shoulders are set,
a 1m broad stripe is added to concrete and asphalt runways to make
the blending runway-airport surface more visually appealing.
This function uses the 2 runway end points as supplied by apt.dat,
which eliminates some inaccuracies that occur when calculating with
only one center point.
- experimental : I haven't seen that this change has any detrimental effectt, and it avoids a
crash in triangle library if I don't split long edges before diffing the accumulator to
generate base_poly. Side effect could be t-junctions, but preliminary look at KATL seemed to give
the same (poor) effect - sometimes there are vertical gaps between the airport base and the terrain.
- Bad data on a few airports - Linear feature definition followed immediately by a
termination node - can't have a single point for linear feature definition, so I discard it.
clipping library format, so it doesn't have to be converted / reconverted
between clipping and TGPolygon types for each operation - shaves about
12 seconds from KATL build.
This does now support literal runway names like "07" instead of "7".
Also support upcoming "XX" designation which marks a runway as
non-numbered. Testing needed.
The width calculation is corrected now and should fix distorted
runway textures.
Hopefully we can get rid of this width mumbo-jumbo soon, too.
It has to stay for now as I found no solution.
