2d Delaunay triangulation with mesh refinement for Unity with Burst compiler
A single-file package which provides simple Delaunay triangulation of the given set of points (float2
) with mesh refinement.
Implemented classic Delaunay triangulation is based on
delaunator
and delaunator-sharp
.
Refinement algorithm is based on Ruppert's algorithm[^ruppert.1995] with Bowyer–Watson algorithm[^bowyer.1981] [^watson.1981] point insertion.
Refinement procedure is inspired by Shewchuk's terminator algorithm[^shewchuk.2002].
The package provides also constrained triangulation (with mesh refinement) which is based on Sloan's algorithm[^sloan.1993].
As an illustrative example, we present the triangulation of Lake Superior with various refinement parameters. The top-left image shows the result without any refinement.
Install the package using one of the following methods
manifest.json
manually.
Add or modify scoped registries in the manifest
"scopedRegistries": [ { "name": "OpenUPM", "url": "https://package.openupm.com/", "scopes": [ "com.andywiecko" ] } ]and in the dependencies provide selected version of the package
"dependencies": { "com.andywiecko.burst.triangulator": "2.5.0", ...See Unity docs for more details https://docs.unity3d.com/2021.1/Documentation/Manual/upm-scoped.html
git
install package.json
using Package Manager (Window/Package Manager
).
Runtime/Triangulator.cs
Runtime/Triangulator.cs
somewhere in the project to use it independently.
To view the documentation for the scripting API locally, navigate to Documentation~/docfx
, or access it online here.
The manual content in this README.md
will soon be migrated into DocFX.
Below one can find example usage of the Triangulator
with input set as four
points that form the unit square:
using var positions = new NativeArray<float2>(new float2[]
{
new(0, 0),
new(1, 0),
new(1, 1),
new(0, 1),
}, Allocator.Persistent);
using var triangulator = new Triangulator(capacity: 1024, Allocator.Persistent)
{
Input = { Positions = positions }
};
triangulator.Run();
var outputTriangles = triangulator.Output.Triangles;
var outputPositions = triangulator.Output.Positions;
[!TIP]
To run triangulation on the main thread, use thetriangulator.Run()
method. If you want to call this within a jobs pipeline, schedule a job usingtriangulator.Schedule(dependencies)
.
The result of the triangulation procedure will depend on selected settings. There are a few settings of the triangulation, shortly described below:
using var triangulator = new Triangulator(1024, Allocator.Persistent)
{
Settings =
{
RefinementThresholds = {
// Specifies the maximum area constraint for triangles in the resulting mesh refinement.
// Ensures that no triangle in the mesh has an area larger than the specified value.
Area = 1f,
// Specifies the refinement angle constraint for triangles in the resulting mesh.
// Ensures that no triangle in the mesh has an angle smaller than the specified value.
Angle = math.radians(20),
},
// If true refines mesh using Ruppert's algorithm.
RefineMesh = true,
// If true the mesh boundary is restored using Input constraint edges.
RestoreBoundary = false,
/// If set to true, the provided data will be validated before running the triangulation procedure.
/// If one of the conditions fails, the triangulation will not be calculated.
/// This can be detected as an error by inspecting OutputData.Status value (native, can be used in jobs).
ValidateInput = true,
// If set to true, caught errors with Triangulator will be logged in the Console.
Verbose = true,
// Type of preprocessing algorithm, see the section below for more details.
Preprocessor = Triangulator.Preprocessor.None,
}
};
If the triangulation algorithm fails, checking the status and handling it in the job pipeline can be considered. For example:
[BurstCompile]
private struct Job : IJob
{
NativeReference<Triangulator.Status>.ReadOnly status;
public Job(Triangulator triangulator)
{
status = triangulator.Output.Status.AsReadOnly();
}
public void Execute()
{
if(status != Triangulator.Status.OK)
{
return;
}
...
}
}
...
var dependencies = default(JobHandle);
dependencies = triangulator.Schedule(dependencies);
dependencies = new Job(triangulator).Schedule(dependencies);
...
Below, you can find the results of the triangulation for different selected options. The cool guitar was used as an input test case.
To use classic, i.e. non-constrained without refinement, Delaunay triangulation one can use the following
using var positions = new NativeArray<float2>(..., Allocator.Persistent);
using var triangulator = new Triangulator(Allocator.Persistent)
{
Input = { Positions = positions }
};
triangulator.Schedule().Complete();
var triangles = triangulator.Output.Triangles;
The result without mesh refinement (Delaunay triangulation):
To proceed with triangulation with the mesh refinement one has to set a proper refinement option
using var positions = new NativeArray<float2>(..., Allocator.Persistent);
using var triangulator = new Triangulator(Allocator.Persistent)
{
Input = { Positions = positions },
Settings = {
RefineMesh = true,
RefinementThresholds = {
Area = 1f,
Angle = math.radians(20f)
},
}
};
triangulator.Schedule().Complete();
var triangles = triangulator.Output.Triangles;
The result with mesh refinement:
The refinement process is controlled by two threshold parameters:
Area
: denoted as $C_\triangle$Angle
: denoted as $C_\theta$These parameters allow fine-tuning of the refinement results based on specific criteria. Below, you can observe a set of results obtained by applying the refinement process to input data from Lake Superior (open image in a new tab to see the details).
It is not guaranteed that the boundary of the input will be present in the classic Delaunay triangulation result. One needs to specify the constraints to resolve this issue. To specify the edges which should be present in the final triangulation provide the additional input data
// Provided input of constraint edges
// (a0, a1), (b0, b1), (c0, c1), ...
// should be in the following form
// constraintEdges elements:
// [0]: a0, [1]: a1, [2]: b0, [3]: b1, ...
using var constraintEdges = new NativeArray<int>(64, Allocator.Persistent);
using var positions = new NativeArray<float2>(..., Allocator.Persistent);
using var triangulator = new Triangulator(Allocator.Persistent)
{
Input = {
Positions = positions,
ConstraintEdges = constraintEdges,
},
};
triangulator.Schedule().Complete();
var triangles = triangulator.Output.Triangles;
After providing the corresponding input for the constraints, the result of the constrained triangulation fully covers all specified edges by the user
Constrained triangulation can be also refined in the same manner as non-constrained one, by enabling corresponding options in triangulation settings:
triangulator.Settings.RefineMesh = true;
After enabling the refinement and the constraint and providing the input, the result of the constrained triangulation fully covers all specified edges by the user and the mesh is refined with the given refinement conditions.
The package provides also an option for restoring the boundaries. One has to enable corresponding options and provide the constraints
settings.RestoreBoundary = true;
The package provides also an option for creating holes.
Except for setting the ConstraintEdges
, a user needs to provide positions of the holes in the same space as the Input.Positions
.
Enabling RestoringBoundary
option is not mandatory, holes could be introduced independently of preserving the boundaries
settings.RestoreBoundary = true; // optional
using var holes = new NativeArray<float2>(new[]{ math.float2(0.5f, 0.5f) }, Allocator.Persistent);
input.HoleSeeds = holes;
Below one can find the comparison of the results of all possible settings which are available in the package.
If Triangulator.Settings.ValidateInput
is set to true, the provided data will be validated before running the triangulation procedure.
Input positions, as well as input constraints, have a few restrictions:
If one of the conditions fails, then triangulation will not be calculated.
One could catch this as an error by using triangulator.Output.Status
(native, can be used in jobs).
using var triangulator = new Triangulator(Allocator.Persistent)
{
Input = { ... },
Settings = {
ValidateInput = true
},
};
triangulator.Run();
var status = triangulator.Output.Status.Value;
BurstTriangulation
input can be generated with job pipeline. One has to use DeferredJobArrays
, see the example snippet:
using var positions = new NativeList<float2>(64, Allocator.Persistent);
using var constraints = new NativeList<int>(64, Allocator.Persistent);
using var holes = new NativeList<float2>(64, Allocator.Persistent);
using var triangulator = new Triangulator(64, Allocator.Persistent)
{
Input =
{
Positions = positions.AsDeferredJobArray(),
ConstraintEdges = constraints.AsDeferredJobArray(),
HoleSeeds = holes.AsDeferredJobArray()
}
}
var dependencies = new JobHandle();
dependencies = new GenerateInputJob(positions, constraints, holes).Schedule(dependencies); // Lists are fed here.
dependencies = triangulator.Schedule(dependencies);
dependencies.Complete();
Triangulation for non-uniform data can be demanding, and a few algorithm steps may get stuck if the data is not preprocessed properly. It is highly recommended that the user should prepare the input data on his own, however, this project provides a few built-in methods.
Preprocessor | Description |
---|---|
None | Default, no effect. |
COM | Transforms input into normalized local space, i.e. [-1, 1] box. |
PCA | Transforms input into normalized coordinate systems obtained with principal component analysis. |
To use one of the following preprocessors use corresponding settings
triangulator.Settings.Preprocessor = Triangulator.Preprocessor.COM;
The algorithm usually can help in situations when the Sloan algorithm gets stuck. The transformation can be applied using the following steps:
To summarize the transformation is given by:
$$ \boxed{x_i \to s[U^{\mathsf T}(x_i - \mu) - c]} $$
and inverse transformation
$$ \boxed{x_i \to U(x_i / s + c) + \mu}. $$
[!NOTE]
The PCA transformation does not preserve theSettings.MinimumAngle
used for refinement. As a result, triangles can be classified as bad in the PCA local space.
The package utilizes the Burst
compiler, which generates highly optimized native code using LLVM.
Below, you'll find a performance comparison (with Burst enabled) between v2.0.0
and v2.1.0
, as well as a comparison with delaunator-sharp
for classic Delaunay triangulation (without refinement or constraints).
Below, you can find a benchmark for constrained triangulation for both v2.1
and v2.2
. The test specimen consists of a 100×100 grid with additional #constraints
-points distributed in a circle at the center of the grid. In some cases of v2.1
, the algorithm gets stuck. Reference timings for non-constrained triangulation are marked with a gray line.
In the figure below, you can also see example test cases: red represents resulting triangles, and blue represents constrained edges.
Furthermore, we present a performance comparison (with Burst enabled) between v1.0
, v2.0
, v2.3
, and v2.4
for the refinement task.
[!NOTE]
Since v2.4, the triangulation refinement algorithm has been updated, resulting in improved mesh quality.
#103: Leak Detected Warning in the Console.
In the Unity Editor, you may encounter the following log message:
Leak Detected : Persistent allocates 257 individual allocations. To find out more please enable 'Jobs/LeakDetection/Full StackTraces' and reproduce the leak again.
Not to worry, this issue is likely related to an internal bug in the Unity.Collections
or Unity.Burst
package (related to NativeQueue<>
allocation).
#105, #106: Incorrect triangulations for complicated input.
Due to floating-point precision, triangulation may fail for some input. This is often related to single-point precision. Changing coordinates from float2
to double2
solves the issue. This will be addressed in the upcoming release. If you want to try it now, there is an experimental branch available here.
halfedges
approach.halfedges
approach.halfedges
approach.ConstrainEdges
as obsolete.constrainedHalfedges
.SplitPermitted
for terminator.[^bowyer.1981]: A. Bowyer. "Computing Dirichlet tessellations". Comput. J. 24 (2): 162–166 (1981). [^watson.1981]: D. F. Watson. "Computing the n-dimensional Delaunay tessellation with application to Voronoi polytopes". Comput. J. 24 (2): 167–172 (1981). [^sloan.1993]:S. W. Sloan. "A fast algorithm for generating constrained Delaunay triangulations." Comput. Struct. 47.3:441-450 (1993). [^ruppert.1995]:J. Ruppert. "A Delaunay Refinement Algorithm for Quality 2-Dimensional Mesh Generation". J. Algorithms 18(3):548-585 (1995). [^shewchuk.2002]:J. R. Shewchuk. "Delaunay refinement algorithms for triangular mesh generation." Comput. Geom. 22.1-3 (2002).