Mesh Generation

(1)

Mesh Generation

Jean-Daniel Boissonnat Geometrica, INRIA

http://www-sop.inria.fr/geometrica

Winter School, University of Nice Sophia Antipolis January 26-30, 2015

(2)

Meshing surfaces and 3D domains

visualization and graphics applications CAD and reverse engineering

geometric modelling in medecine, geology, biology etc.

autonomous exploration and mapping (SLAM) scientific computing : meshes for FEM

P d

dO

Ω

O o

Winter School 4 Mesh generation Sophia Antipolis 2 / 23

(3)

Mesh generation : from art to science

Grid methods

Lorensen & Cline [87] : marching cube Lopez & Brodlie [03] : topological consistency

Plantiga & Vegter [04] : certified topology using interval arithmetic

Morse theory

Stander & Hart [97]

B., Cohen-Steiner & Vegter [04] : certified topology

Delaunay refinement

Hermeline [84]

Ruppert [95]

Shewchuk [02]

Chew [93]

B. & Oudot [03,04]

Cheng et al. [04]

(4)

Main issues

Sampling

How do we choose points in the domain ?

What information do we need to know/measure about the domain ?

Meshing

How do we connect the points ?

Under what sampling conditions can we compute a good approximation of the domain ?

(5)

Restricted Delaunay triangulation

^{[Chew 93]}

Definition

Therestricted Delaunay triangulationDelX(P)to X⊂R^d is the nerve of Vor(P)∩X

If P is anε-sample, any ball centered onXthat circumscribes a facet ofDelX(P) has a radius≤εrch(M)

(6)

Restricted Delaunay triangulation

^{[Chew 93]}

Definition

(7)

Restricted Delaunay triangulation

^{[Chew 93]}

Definition

(8)

Restricted Delaunay triangulation

^{[Chew 93]}

Definition

(9)

A variant of the nerve theorem

[Edelsbrunner & Shah 1997]

LetMbe a compact manifold without boundary. If, for any face f ∈Vor(P) s.t. f ∩M6=∅,

1 f intersectsMtransversally

2 f ∩M=∅or is a topological ball thenDel_M(P)≈M

(10)

A variant of the nerve theorem

[Edelsbrunner & Shah 1997]

LetMbe a compact manifold without boundary. If, for any face f ∈Vor(P) s.t. f ∩M6=∅,

1 f intersectsMtransversally

2 f ∩M=∅or is a topological ball thenDel_M(P)≈M

(11)

Proof of the closed ball property

Barycentric subdivision

ofVor(P)∩M ofDel_M(P)

(12)

Good sampling, scale and dimension

(13)

Sampling conditions

[Federer 1958], [Amenta & Bern 1998]

Medial axis ofM: axis(M)

set of points with at least two closest points onM

Reach

∀x∈M, rch(x)=infimum of the radii of the medial balls tangent toMatx rch(M)=infx∈Mrch(x)

(, η)-net ofM

1.P ⊂M,∀x∈M: d(x,P)≤rch(x)

(14)

Sampling conditions

[Federer 1958], [Amenta & Bern 1998]

Medial axis ofM: axis(M)

set of points with at least two closest points onM

Reach

∀x∈M, rch(x)=infimum of the radii of the medial balls tangent toMatx rch(M)=infx∈Mrch(x)

(, η)-net ofM

1.P ⊂M,∀x∈M: d(x,P)≤rch(x)

2.∀p,q∈ P, kp−qk ≥η min(rch(p),rch(q))

(15)

Restricted Delaunay triang. of ( ε, η )-nets

[Amenta et al. 1998-], [B. & Oudot 2005]

If

S ⊂R³ is a compact surface of positive reach without boundary

P is an(ε, η)-net withε/η≤ξ₀andε small enough

then

Del|S(S)provides good estimates of normals

There exists ahomeomorphism φ:Del_|S(P)→ S

sup_x(kφ(x)−xk) =O(ε²)

(16)

Surface mesh generation by Delaunay refinement

[Chew 1993, B. & Oudot 2003]

φ:S→R= Lipschitz function

∀x∈S, 0< φmin≤φ(x)< εrch(x)

ORACLE: For a facetf ofDel|S(P), returncf,rf andφ(cf) A facetf isbadifrf > φ(cf)

Context and Motivation

Delaunay refinement meshing engine

The algorithms refines:

Bad facets: f ∈Del_|S(P) – oversized (sizing field)

– badly shaped (min angle bound) – inaccurate (distance bound) Bad Tetrahedra : t∈Del_|O(P) – oversized (sizing field)

– badly-shaped (radius-egde ratio) Required oracle on domain to be meshed

• point location in domain and subdomains

• intersection detection/computation between boundary surfaces and segments (Delaunay edges)

PA-MY (INRIA Geometrica) CGALmesh ARC-ADT-AE-2010 9 / 36

Algorithm

INIT compute an initial (small) sampleP0⊂S REPEAT IFf is a bad facet

insert in Del3D(cf), updatePandDel|S(P) UNTIL all facets are good

(17)

Surface mesh generation by Delaunay refinement

[Chew 1993, B. & Oudot 2003]

φ:S→R= Lipschitz function

∀x∈S, 0< φmin≤φ(x)< εrch(x)

ORACLE: For a facetf ofDel|S(P), returncf,rf andφ(cf) A facetf isbadifrf > φ(cf)

Context and Motivation

Delaunay refinement meshing engine

The algorithms refines:

Bad facets: f ∈Del_|S(P) – oversized (sizing field)

– badly shaped (min angle bound) – inaccurate (distance bound) Bad Tetrahedra : t∈Del_|O(P) – oversized (sizing field)

– badly-shaped (radius-egde ratio) Required oracle on domain to be meshed

• point location in domain and subdomains

• intersection detection/computation between boundary surfaces and segments (Delaunay edges)

Algorithm

INIT compute an initial (small) sampleP0⊂S REPEAT IFf is a bad facet

insert in Del3D(cf), updatePandDel|S(P) UNTIL all facets are good

(18)

The meshing algorithm in action

(19)

The algorithm outputs a good sample

The output sampleP is sparse

∀p∈ P,d(p,P \ {p}) =kp−qk

≥min(φ(p), φ(q))

≥φ(p)− kp−qk

⇒ kp−qk ≥ ¹₂φ(p)≥ ¹₂φ₀ >0 the algorithm terminates

P is a looseε-sample ofS

I Each facet has aS-radiusr_f ≤φ(cf)< εrch(S)

I Del_|_S(P)has a vertex on each cc ofS (INIT)

(20)

Size of the sample = Θ R

S dx φ²(x)

Upper bound

Bp=B(p,^φ(p)₂ ),p∈ P R

S dx φ²(x) ≥P

p

R

(Bp∩S) dx

φ²(x) (theBpare disjoint)

≥⁴₉P

p

area^(Bp∩S)

φ²(p) φ(x)≤φ(p) +kp−xk

≤ ³₂ φ(p))

≥P

pC = C|P|

reach(S)

p φ(p)2

Lower bound

Use a covering instead of a packing

(21)

Size of the sample = Θ R

S dx φ²(x)

Upper bound

Bp=B(p,^φ(p)₂ ),p∈ P R

S dx φ²(x) ≥P

p

R

(Bp∩S) dx

φ²(x) (theBpare disjoint)

≥⁴₉P

p

area^(Bp∩S)

φ²(p) φ(x)≤φ(p) +kp−xk

≤ ³₂ φ(p))

≥P

pC = C|P|

reach(S)

p φ(p)2

Lower bound

Use a covering instead of a packing

(22)

The full result

The Delaunay refinement algorithm produces a good (dense and sparse) sample a triangulated surfaceˆS

I homeomorphic toS

I close toS (Hausdorff/Fr ´echet distance, approximation of normals)

(23)

Applications

Implicit surfacesf(x,y,z) =0

Isosurfaces in a 3d image (Medical images) Triangulated surfaces (Remeshing)

Point sets (Surface reconstruction)

see cgal.org, CGALmesh project

(24)

Results on smooth implicit surfaces

(25)

CGALmesh Achievements

Meshing 3D domains

Input from segmented 3D medical images

[INSERM] [SIEMENS]

(26)

Comparison with the Marching Cube algorithm

Delaunay refinement Marching cube

(27)

Meshing with sharp features

A polyhedral example

(28)

Meshing 3D multi-domains

Input from segmented 3D medical images [IRCAD]

(29)

Mesh Generation

Mesh Generation

Meshing surfaces and 3D domains

Mesh generation : from art to science

Main issues

Restricted Delaunay triangulation

Restricted Delaunay triangulation

Restricted Delaunay triangulation

Restricted Delaunay triangulation

A variant of the nerve theorem

A variant of the nerve theorem

Proof of the closed ball property

Good sampling, scale and dimension

Sampling conditions

Sampling conditions

Restricted Delaunay triang. of ( ε, η )-nets

Surface mesh generation by Delaunay refinement

Surface mesh generation by Delaunay refinement

The meshing algorithm in action

The algorithm outputs a good sample

Size of the sample = Θ R

Size of the sample = Θ R

The full result

Applications

Results on smooth implicit surfaces

Comparison with the Marching Cube algorithm

Surface reconstruction from unorganized point sets