11/8/2022

INF 574
Geometric Modeling: Digital Representation
and Analysis of Shapes

Last Time: Polygonal Representations

• Why mesh representations are so ubiquitous

• Mesh terminology

 from abstract complex to geometric realization

 regularity (structure), isotropy/anisotropy, gradation, etc.

• Typical mesh data structures

• Two mesh processing methods

 simplification

 mesh compression (just an intro)

1
 11/8/2022

Today: Surface Reconstruction

• How do you turn scanned data into usable meshes?

• Two main family of methods

 using Computational Geometry

• reliance on Voronoi/Delaunay

 via implicit function fitting

• Poisson reconstruction

• and noise-robust variants

Reconstruction
inferring a surface from points

2
 11/8/2022

Motivation
Scanners provide highly detailed pointsets of 3D shapes
http://www.jhu.edu/digitalhammurabi/ http://graphics.stanford.edu/projects/mich/

Cuneiform tablets

Motivation
Scanners provide highly detailed pointsets of 3D shapes

We want a 3D mesh for flexibility!

 allows for
• parameterization from P. Alliez

• computer-aided analysis
• rapid prototyping
• rendering
 and so much more…

3
 11/8/2022

Surface Reconstruction
How to generate a mesh from a set of surface samples?

Surface Reconstruction
How to generate a mesh from a set of surface samples?
Three general approaches:
 Computational Geometry (often “interpolating” reconstruction)
Boissonnat, 1984 Edelsbrunner, 1984
Amenta et al., 1998 Dey et al., 2003
 variational surface fitting (for a given topology and smoothness)
Terzopoulos et al., 1991 Chen et al., 1995
 Implicit function fitting
Hoppe et al., 1992 Curless et al., 1996
Whitaker, 1998 Carr et al., 2001
Davis et al., 2002 Ohtake et al., 2004
Turk et al., 2004 Shen et al., 2004
Kazhdan, 2005 and many more!

4
 11/8/2022

Reconstruction
using Computational Geometry

Early Approaches
How to make sense of a pointset?
Idea 1: ball rolling [Edelsbrunner et al.]
 alpha-shape: created by a radius-α ball rolled along the pointset
• space delineated by point pairs touched by an empty disc of radius α
 interesting for topology analysis too
 but finding a fixed alpha is “mission: impossible”

10

5
 11/8/2022

Early Approaches
How to make sense of a pointset?
Idea 1: ball rolling [Edelsbrunner et al.]
 alpha-shape: created by a radius-α ball rolled along the pointset
• space delineated by point pairs touched by an empty disc of radius α
 interesting for topology analysis too
 but finding a fixed alpha is “mission: impossible”

Idea 2: graph-based approaches [Mencl]

 calculate the Euclidean minimum spanning tree (EMST) of points
• acyclic connected graph so that the sum of its edge lengths is minimized
 augment with a surface description graph (SDG) in successive stages
• complete wireframe based on tangent-plane heuristics
 Fill final SDG with triangles
• and you may get (very) lucky if you started from a decent sampling….

11

Voronoi Diagram of Point Sets

The Voronoi Diagram of a point set S={pi}i in ℝn
 partition of space into regions V(pi) …
 such that all points in V(pi) are closer to pi than any other point in S

12

6
 11/8/2022

Voronoi Diagram of Point Sets

A point can belong to one of the following
 Voronoi region ( is closer to than to any other )

 Voronoi edge ( is equidistant from and )

 Voronoi vertex (equidistant from )

Voronoi region Voronoi edge Voronoi vertex

13

Voronoi Diagrams in Nature

V
o
r
o
n
o
i

Fish territories at the bottom of the sea

14

7
 11/8/2022

Beautiful Duality
Voronoi diagrams are only half the picture
 connectivity of Voronoi regions is nice… Georgy Voronoï Boris Delaunay
(1868-1908) (1890-1980)
 introducing Delaunay triangulation!
• dual of Voronoi diagram
• Voronoi vs. Delaunay edges are orthogonal!
 canonical construction, just from points!

Delaunay triangulation
Voronoi diagram

15

Medial Axis (a.k.a. Skeleton)

Locus of sphere centers touching the boundary in two+ points
 wave-front collision/grass fire analogy
 Media Axis Transform: skeleton together with the associated radii
• shape descriptor; sort of “dual’’ of a shape

16

8
 11/8/2022

Medial Axis (a.k.a. Skeleton)

Locus of sphere centers touching the boundary in two+ points
 wave-front collision/grass fire analogy
 Media Axis Transform: skeleton together with the associated radii
• shape descriptor; sort of “dual’’ of a shape

Voronoi diagrams are discrete medial axes!

 Voronoi balls are like medial balls
 in 2D, Voronoi vertices form the
medial axis as sampling density grows
 rather easy to compute
• can even subsample if too many points

17

Reconstruction from Union of Balls

Voronoi balls nicely approximate shape inside and outside
Exploitable for reconstruction! Crust algorithm [Amenta ‘98]
 a “shape” edge between two points is far from all Voronoi vertices
 so construct Delaunay triangulation of points & their Voronoi vertices
• any edge present between two samples
is part of the shape reconstruction
 Perfect! Right??
• guaranteed in 2D for dense enough input
• but fails miserably in 3D…

18

9
 11/8/2022

3D Case
In 3D, some Voronoi vertices are not near medial axis…
 spurious Voronoi vertices create issues!
 luckily, we can “filter them out…
• for each Voronoi cell, pick the two extremes
• called “poles”
• this subset of Voronoi vertices
are good approximation of skeleton

Spurious Voronoi vertex

19

3D Surface Reconstruction
Compute Delaunay triangulations of input points and poles…
and keep only triangle on each triple of sample points
 extensions: cocone and power crust
• to reduce complexity
 issues with this family of approaches:
• can break down in case of non-smooth surfaces
• sensitive to noise, outliers, and sparsity of data

20

10
 11/8/2022

Reconstruction
via Implicit Functions

21

Implicit Function Approach

A general, two-step approach
 Infer a function with:
• negative value outside
• positive values inside
• (or reverse)
 extract the zero-set

implicit function reconstruction <0 0 >0

22

11
 11/8/2022

Implicit Function Approach

A general, two-step approach
 Infer a function with:
• negative value outside
[Carr et al. 2001] FastRBF
• positive values inside
• (or reverse)
 extract the zero-set

Finding this implicit function is key

 first ways a bit brittle…
• needed EMST + normal orientation [Hoppe et al.]
• VRIP using range distances + carving [Curless-Levoy]
 or expensive kernel approach
• solve dense linear system (RBF, MPU)
• can’t really control away from samples

23

Indicator Function Approach

We may want to reconstruct the surface by solving for a
sort of “indicator” function of the model…
1 if p  M
M  p  
0 if p  M 0
0
 but how? 1
 imagine a smoothed indicator function… 1
0
0
1
0
Indicator function
M

24

12
 11/8/2022

Indicator Function Approach

We may want to reconstruct the surface by solving for a
sort of “indicator” function of the model…
1 if p  M
M  p  
0 if p  M 0 0

 but how? 0
0
 imagine a smoothed indicator function…
• its gradient looks suspiciously like our input! 0 0

Indicator gradient
M
Oriented points
(points+normals)

25

Poisson Reconstruction [Kazhdan et al.]


Represent input data by a vector field V
 ‘splat’ the normals a bit to avoid Dirac functions
 summing those splats creates local vector field near samples

Find the function  whose gradient best approximates V :
 2
min  (  V )
Turns out to be a Poisson problem

26

13
 11/8/2022

Simple Implementation
Given the points:
 set regular grid, or octree
 compute vector field
 compute indicator function
 extract isosurface

27

Implementation: Adapted Octree

Given the points:
 set regular grid, or octree
 compute vector field
 compute indicator function
 extract isosurface

28

14
 11/8/2022

Implementation: Adapted Octree

Given the points:
 set regular grid, or octree
 compute vector field
• splat the samples
 compute indicator function
 extract iso-surface

29

Implementation: Adapted Octree

Given the points:
 set regular grid, or octree
 compute vector field
• splat the samples
 compute indicator function
 extract iso-surface

30

15
 11/8/2022

Implementation: Adapted Octree

Given the points:
 set regular grid, or octree
 compute vector field
• splat the samples
 compute indicator function
 extract iso-surface

31

Implementation: Adapted Octree

Given the points:
 set regular grid, or octree
 compute vector field
• splat the samples
 compute indicator function
 extract iso-surface

32

16
 11/8/2022

Implementation: Adapted Octree

Given the points:
 set regular grid, or octree
 compute vector field
 compute indicator function
• compute divergence of normals
• solve Poisson equation
 extract iso-surface

Typically, finite differences

33

Implementation: Indicator Function

Given the points:
 set regular grid, or octree
 compute vector field
 compute indicator function
• compute divergence of normals
• solve Poisson equation
 extract iso-surface

34

17
 11/8/2022

Implementation: Extraction
Given the points:
 set regular grid, or octree
 compute vector field
 compute indicator function
 extract iso-surface

35

Michelangelo’s David

• 215 million data points

 from 1000 scans
• 22 million triangle after MC
• Compute time: 2.1 hours
• Peak memory: 6.6GB

36

18
 11/8/2022

David – Chisel marks

37

David – Drill Marks

38

19
 11/8/2022

David – Eye

39

Test: Stanford Bunny

Power Crust FastRBF MPU

VRIP Poisson

40

20
 11/8/2022

Medical Data
Arbitrary topology

41

More Difficult Inputs

Poisson reconstruction widely used, but still brittle
 if normal estimation/orientation fails, it becomes useless
nice sampling
bad sampling

42

21
 11/8/2022

First Idea: Signing the Unsigned [Mullen et al.]

Noise- and outlier-robust unsigned distance
 simple evaluation based on k-nearest neighbors [Chazal et al. 09]

 roots in optimal transport (Wasserstein-2) distance

no noise in data with noise in data

43

First Idea: Signing the Unsigned [Mullen et al.]

Noise- and outlier-robust unsinged distance
 simple evaluation based on k-nearest neighbors [Chazal et al. 09]
Guess sign based on stochastic ray crossing
 from a point, shoot a few rays and guess/vote if inside or outside
 results in sign guess and uncertainty

44

22
 11/8/2022

First Idea: Signing the Unsigned [Mullen et al.]

Noise- and outlier-robust unsinged distance
 simple evaluation based on k-nearest neighbors [Chazal et al. 09]
Guess sign based on stochastic ray crossing
 from a point, shoot a few rays and guess/vote if inside or outside
Decide spatially consistent sign to get signed distance

45

First Idea: Signing the Unsigned [Mullen et al.]

Noise- and outlier-robust unsinged distance
 simple evaluation based on k-nearest neighbors [Chazal et al. 09]
Guess sign based on stochastic ray crossing
 from a point, shoot a few rays and guess/vote if inside or outside
Decide spatially consistent sign to get signed distance
 quadratic equation to minimize once sign is guessed

Favor smoothness confidence sign guess

46

23
 11/8/2022

Result I

Poisson reconstruction

47

Result II

Poisson reconstruction

Red parts: where the

normal orienter fails

48

24
 11/8/2022

Second Idea: Hybrid Approach

Mixing Computational Geometry & linear algebra
 use CGeo techniques to find unoriented normals (and confidence)
• use anisotropy of Voronoi cells
 orient normals and reconstruct surface
• using an eigenvalue approach

tensor estimation implicit function + contouring

49

Voronoi-based Eigen-Reconstruction
Use covariance matrix of Voronoi cells

no noise some noise noise

50

25
 11/8/2022

Voronoi-based Eigen-Reconstruction
Use covariance matrix of Voronoi cells [Alliez et al., Merigot et al.]
 in case of noise, using union of k cells instead (automated selection)

anisotropy

51

Voronoi-based Eigen-Reconstruction
Use covariance matrix of Voronoi cells [Alliez et al., Merigot et al.]
 in case of noise, using union of k cells instead (automated selection)
 deduce normal tensor C (direction and anisotropy to measure uncertainty)
Variational formulation to orient normal and reconstruct mesh
 Find f whose gradient best aligns to principal component of tensors Ci

Biharmonic energy
Anisotropic Dirichlet energy Measures smoothness of f
Measures alignment with tensors

for well-posedness
(regularization)

52

26
 11/8/2022

Voronoi-based Eigen-Reconstruction
Use covariance matrix of Voronoi cells [Alliez et al., Merigot et al.]
 in case of noise, using union of k cells instead (automated selection)
 deduce normal tensor C (direction and anisotropy to measure uncertainty)
Variational formulation to orient normal and reconstruct mesh
 Find f whose gradient best aligns to principal component of tensors Ci

• for regions with

 anisotropic tensors: favor alignment
 isotropic tensors: favor smoothness
• consistent orientation of f

53

Results

Poisson Generalized
reconstruction eigenvalue

54

27
 11/8/2022

Questions?

55

28