Repaired non-robust Delaunay triangulator. #1699

jpolchlo · 2016-10-25T13:45:07Z

This PR replaces the custom, non-robust Delaunay triangulator with wrapped version of the JTS triangulator. The old triangulator did not implement robust predicates for collinearity/cocircularity which could cause failures in some real-world data sets.

The JTS triangulator was wrapped to allow a more idiomatic Scala usage. Namely, allowing for PointFeatures to be used more easily, which would facilitate the use of a triangulation for the purposes of, say, interpolating a field of numbers defined over a sample of points. This wrapped version also replaces the quad-edge structure used by JTS for a lighter-weight half-edge structure. Documentation for that structure can be found in docs/vector/voronoi.md.

This PR closes issue #1683.

…n of the JTS triangulator

lossyrob · 2016-10-25T15:08:20Z

@jpolchlo you made benchmarks for the old voronio, right? Is it possible to rerun those benchmarks, against the two different versions, to see if we're losing a lot of speed?

jpolchlo · 2016-10-25T19:02:32Z

You and I sat down to look at benchmarks, but I didn't write them myself,
so I'm not 100% on how to proceed. With some pointers, I'm sure I could do
it. But I wonder what you have in mind here, since the version that got
replaced will fail on a certain class of inputs, and surely, if I were to
make my original code robust, it will necessarily slow down based on the
additional operations needed to ensure correctness.

On Tue, Oct 25, 2016 at 11:08 AM, Rob Emanuele notifications@github.com
wrote:

@jpolchlo https://github.com/jpolchlo you made benchmarks for the old
voronio, right? Is it possible to rerun those benchmarks, against the two
different versions, to see if we're losing a lot of speed?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1699 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/ARjt0aoZ-AZkd9adAEvp0GmdZEsPb57Dks5q3htngaJpZM4KgAI8
.

lossyrob · 2016-10-26T02:47:29Z

@jpolchlo the idea is to have a sense of how much things have slowed down. And to have an idea of how efficient this operation is...I remember that you were calculating triangulations on the nickel dataset in about 200ms. If this turns out to be a 2x vs a 10x slowdown, that would be good to understand. If it's a serious slowdown, I'd want to either try to optimize before release or at least put in an issue for future optimization.

jpolchlo · 2016-10-26T15:46:59Z

Roger, boss. I'll see what I can figure out.

On Tue, Oct 25, 2016 at 10:47 PM, Rob Emanuele notifications@github.com
wrote:

@jpolchlo https://github.com/jpolchlo the idea is to have a sense of
how much things have slowed down. And to have an idea of how efficient this
operation is...I remember that you were calculating complex triangulations
on the nickel dataset in about 200ms. If this turns out to be a 2x vs a 10x
slowdown, that would be good to understand. If it's a serious slowdown, I'd
want to either try to optimize before release or at least put in an issue
for future optimization.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1699 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/ARjt0WyJPoijxi0-UCG5rQR9NBuSpUBeks5q3r9EgaJpZM4KgAI8
.

lossyrob · 2016-10-30T17:57:01Z

vector/src/main/scala/geotrellis/vector/voronoi/Delaunay.scala

+  val cm = ColorMap(scala.collection.immutable.Map(1 -> 0x000000ff, 255 -> 0xffffffff))
+  tile.renderPng(cm).write("delaunay.png")
+}
+*/


What's going on with all the commented code here?

lossyrob · 2016-10-30T17:59:02Z

vector-test/src/test/scala/spec/geotrellis/vector/voronoi/DelaunaySpec.scala


      (dt.triangles.forall{ case ((ai,bi,ci),_) =>
        val otherPts = (0 until numpts).filter{ i: Int => i != ai && i != bi && i != ci }
        otherPts.forall{ i => ! localInCircle((ai,bi,ci),i) }
      }) should be (true)
    }
+
+    it("should work on a real dataset with many collinear points") {
+      val allcities = ResourceReader.asString("populous_cities.txt")


We don't normally read test data out of resources packed in the JAR - is there an advantage you see of doing it this way, other then adding to test files similar to other tests? We would put it here https://github.com/geotrellis/geotrellis/tree/master/vector-test/data

lossyrob · 2016-10-30T17:59:32Z

util/src/main/scala/geotrellis/util/ResourceReader.scala

+import java.io._
+
+object ResourceReader {
+  def asString(filename: String): String = {


This is for test functionality, not sure it's worth putting in the core utils JAR

Had talked to Eugene about this. He suggested putting it in util, but it's not super valuable, so I'll nuke it.

…diagram (JTS requires no testing)

lossyrob · 2016-11-01T03:49:17Z

vector/src/main/scala/geotrellis/vector/voronoi/Delaunay.scala

-   * boundary(.next)*.face == None.
-   */
-  val boundary = triangulate(0, vIx.length-1)
+  val triangles = _triangles.map{ poly => Polygon.jtsToPolygon(poly.asInstanceOf[JTSPolygon]) }


can just use Polygon(poly.asInstanceOf[JTSPolygon]

lossyrob · 2016-11-01T03:49:23Z

vector-test/src/test/scala/spec/geotrellis/vector/voronoi/DelaunaySpec.scala

-
-import org.scalatest.{FunSpec, Matchers}
-
-class DelaunaySpec extends FunSpec with Matchers {


Should we really scrap all these tests? I think still testing our JTS backed Delaunay class makes sense.

As much as I would like to provide some unit tests here, I'm just at a loss for what should be tested. It seems that most anything that I could test for tests JTS's feature set, and not the features of this package. I'm not sure if that's appropriate. As it is, there is an indirect test of Delaunay and Voronoi in the EuclideanDistanceTileSpec. I'd be happy for some suggestions about what should be tested for in these classes!

After some discussion, we've opted to re-insert some tests essentially as regression tests.

lossyrob · 2016-11-01T04:01:15Z

vector/src/main/scala/geotrellis/vector/voronoi/Delaunay.scala

-    }
+  private val _triangles = {
+    val tris = subd.getTriangles(gf)
+    for ( i <- 0 until tris.getNumGeometries ) yield tris.getGeometryN(i)


use a cfor loop for performance...also do we need the private val?

import spire.syntax.cfor._ val triangles: Seq[Polygon] = { val arr = Array.ofDim[Polygon](tris.getNumGeometries) val len = arr.length cfor(0)(_ < len, _ + 1) { i => arr(i) = Polygon(tris.getGeometryN(i).asInstanceOf[JTSPolygon]) } arr }

lossyrob · 2016-11-02T21:06:47Z

@jpolchlo "This PR closes issue #1683." doesn't kick off the auto close. Has to be "Fixes #1683".

Replaced custom, non-robust Delaunay triangulator with wrapped versio…

d8f091e

…n of the JTS triangulator

lossyrob added this to the 1.0 milestone Oct 26, 2016

jbouffard force-pushed the master branch from 64b2f0c to f1c58f9 Compare October 28, 2016 17:08

lossyrob suggested changes Oct 30, 2016

View reviewed changes

jpolchlo added 3 commits October 31, 2016 10:09

Switched to JTS Delaunay triangulator

57ffde1

Removed test suite for superceded Delaunay triangulation and Voronoi …

85bc94b

…diagram (JTS requires no testing)

Remove ResourceReader from util

05a341a

lossyrob suggested changes Nov 1, 2016

View reviewed changes

jpolchlo added 2 commits November 1, 2016 14:15

Added tests and documentation

149cf2c

Improved unit test of Voronoi

ed51dd7

lossyrob approved these changes Nov 2, 2016

View reviewed changes

lossyrob merged commit 5ca4a8b into locationtech:master Nov 2, 2016

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repaired non-robust Delaunay triangulator. #1699

Repaired non-robust Delaunay triangulator. #1699

jpolchlo commented Oct 25, 2016 •

edited

Loading

lossyrob commented Oct 25, 2016

jpolchlo commented Oct 25, 2016

lossyrob commented Oct 26, 2016 •

edited

Loading

jpolchlo commented Oct 26, 2016

lossyrob Oct 30, 2016

lossyrob Oct 30, 2016

lossyrob Oct 30, 2016

jpolchlo Oct 31, 2016

lossyrob Nov 1, 2016

lossyrob Nov 1, 2016

jpolchlo Nov 1, 2016

jpolchlo Nov 1, 2016

lossyrob Nov 1, 2016

lossyrob commented Nov 2, 2016


		import org.scalatest.{FunSpec, Matchers}

		class DelaunaySpec extends FunSpec with Matchers {

Repaired non-robust Delaunay triangulator. #1699

Repaired non-robust Delaunay triangulator. #1699

Conversation

jpolchlo commented Oct 25, 2016 • edited Loading

lossyrob commented Oct 25, 2016

jpolchlo commented Oct 25, 2016

lossyrob commented Oct 26, 2016 • edited Loading

jpolchlo commented Oct 26, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

lossyrob commented Nov 2, 2016

jpolchlo commented Oct 25, 2016 •

edited

Loading

lossyrob commented Oct 26, 2016 •

edited

Loading