Discrete Differential

Geometry
for triangle meshes
JAB 2013
Consider a plane curve
and a surface
... and the Gauss Map
 Now, take a mesh
tangent plane
discontinuous at
vertices
Planar almost
everywhere

and edges
 Making sense of
curvature on a mesh
• At any point on a mesh, normal curvature in any
direction is either infinite or 0
• However, we can fit a smooth surface to the mesh
• We can replace edges and corners with blends.
• We want curvature integrated over a region
 Seeing Curvature
• Actually ... use reflection lines

More smooth
 Hathing

Why compute
Hatching along lines of curvature [Praun et al.]

curvature?
• Analysis Parameterization and Remeshing
Result of the QuadCover algorithm [Klberer, Nie

• non-photorealistic rendering
Polthier]

• remeshing
• improving geometry
Computing the
The Angle Weighted Normal
Normal

The angle weighted normal,

P
i ↵ i ni
n↵ = P
|| i ↵i ni ||
 Computing Normal
Curvature
• We need the shape
operator:

• Maps tangent plane

directions to normal
derivative

• Derivative of the
Gauss map

• We need: Smooth
quadratic patch fitted
to surface
Estimate a smooth function whence curvature is readily
computed:
Fitting a Patch
1 2 2
f (u, v ) = (au + 2buv + cv ) (5)
2
• For each
For each vertex vertex

• Compute frame
I Compute the normal
(N,T ,T )and a frame T
1 2

I Project one-ring neighbours using

•(uProject
vj)=T
j , ,v
(u j j T
1-ring
j
T
i
vertices:
) =(p T−p )(pj pi )
I Get height values using
•hj Get height values: hj =N·(pj −pi)
= N · (pj pi )
I •FitNow
f asfit shown
(on next in
slide)
the following
 The Shape Operator (osculating pa
Fitting a patch
Coefficients of f found by solving
2 3
2 3
6
6
. . . 7
7 6 . 7
Coefficients of f: a, b, and c 6 72 3 6 7

are found by solving:

6
6
6
. . . 7
7
76 a 7 6
6
6 . 7
7
7
6 uj2 vj2 76 7 6 7
6
6
6 uj vj
76
76
76 b 7 6
7 =6
7 6 hj 7
7
7
6 2 2 74 5 6 7
in the least squares sense 6
6
6 . . .
7
7
7 c 6
6
6 . 7
7
7
6 7 4 5
4 5
. . . .
 Shape Operator
• Now, the shape

operator is
a b
S= (7)
b c
principal directions and curvatures are easily
• since fu(0,0) = fv(0,0) = 0
d by eigenvalue decomposition
• and a,b and c are 2. order derivatives of
f(u,v)= 0.5 (au2 + buv + cv2)
 Principal Curvature
• Directions of min and max principal
curvature vmin and vmax are eigenvectors of
S

•𝜅 min and 𝜅max are the

eigenvalues of S
Curvature Lines
 Ridge Detection

• curvature extrema along

max curvature direction

• Surprisingly hard to
compute in a stable way
 Gaussian and Mean
Curvature
• The Gaussian curvature
K = 𝜅min 𝜅max Gauss Map area to surf area

• The mean curvature

H = 𝜅min + 𝜅max Length of area gradient at a point

• Both of these have far

more meaningful
definitions
 Euler-Poincare

• Remember the Euler-Poincare formula

#V - #E + #F = 𝛘

?
aussian Curvature (alt)

Gauss-Bonnet
Gauss Bonnet Theorem states that
Z Z X
K dA + g ds + i = 2⇡ (R), (3)
R C

where (R) = #Faces #Edges + #Vertices in an

arbitrary triangulation of R. Thus, if we can find a curve
•
with zero Thus,
geodesic
obtain thethe
same
givencurvature around a vertex, then we
a path around
vertex (with zero
geodesic, we can compute P ...✓j
2⇡ j
K (pi ) = 1 P
3 pi 2tj Aj
re (R) = #Faces #Edges + #Vertices in an
Gaussian Curvature
rary triangulation of R. Thus, if we can find a curve
zero geodesic curvature around a vertex, then we
in the same
P
2⇡ ✓
j j
K (pi ) = 1 P
3 p 2t
i j
Aj

ssian Curvature
•
Note, we could also start from a definition
of Gaussian curvature:
AG
K = lim
AS !0 AS

y loop around a single vertex maps to a spherical

Gaussian Curvature
 Integral of Gaussian
The Integral Gaussian Curvature
Curvature
P
Since the Gaussian curvature at a vertex is 2⇡ j ✓j , we
can compute the integral Gaussian curvature of a
•
triangulated
For asurface
closedprecisely:
surface, it is really easy:
2 3
Z X V X
K = 42⇡ ✓j 5
S i=1 j
 Mean Curvature on an
The Mean Curvature of Edges
edge

If we replace an edge e of dihedral angle with a blend

Z
1 1 1
H = (0 + )(r kek) = kek
B 2 r 2

independent of radius r !!
 Integral
The Integral Abs Mean
Absolute Mean
Curvature
Curvature
The integral absolute mean curvature
Z |E|
X
1
• The IAMC is S
|H| =
2
i=1
| i |kei k .

This measure is easily minimized using edge swaps.

The Integral Absolute Mean Curvature
Minimizing IAMC
Minimizing (4)
 Return to the Normal
e Mean Curvature Normal

• A surface is minimal it its area is minimal

given its boundary
A surface is minimal if its area is minimal given its
• If so,such
oundary. For H = 0a surface H is zero. Consequently, H
must be related to the area gradient, and
• Thus H relates to the area gradient. In fact
rA
2H = lim
A!0 A
 Mean Curvature
Normal
The Mean Curvature Normal
P
in a mesh, use the 1-ring area, A1i ring = pi 2tj Atj

Thus:

1 rA1i ring 1 X
H(pi ) = 1 ring
= 1 ring
(cot ↵ij +cot ij )(pi pj ) ,
2 Ai 4Ai pj 2Ni

A spatial average, but where do the cots come from?

Triangle Area Gradient
The Mean Curvature Normal
The gradient of the vertex area is equal to the sum:
1 ring
X
rAi = rAtj (pi )
pi 2tj

kb⇥ak
where Atj (pi ) = 2
and the terms are as shown:
The Mean Curvature Normal
Rewriting.............
The gradient can be found by inspection, and we obtain

(b ⇥ a) ⇥ b
rAtj (pi ) =
2kb ⇥ ak
(bt b)a (bt a)b
=
2kb ⇥ ak
(bt b)a (bt a)a + (bt a)a (bt a)b
=
2kb ⇥ ak
(ct b)a (bt a)c
= +
2kc ⇥ bk 2kb ⇥ ak
(ct b)a (bt a)c
= +
2kc ⇥ bk 2kb ⇥ ak
1
= (a cot + c cot ↵)
2
Mean Curvature
 The Laplace-Beltrami
Operator
• The mean curvature normal is also defined
as the LBO
f = r · rf
applied to the vertex positions
• Not so mysterious
 Litterature
Guide to Computational Geometry Processing, Bærentzen, J.
Andreas ; Gravesen, Jens ; Anton, François ; Aanæs, Henrik, Springer 2012

Discrete Differential-Geometry Operators for Triangulated 2-

Manifolds Mark Meyer, Mathieu Desbrun, Peter Schröder, and Alan H. Barr,
VisMath 2002.

Software: GEL