Geodesic Patterns

Helmut Pottmann Qixing Huang Bailin Deng Alexander Schiftner Martin Kilian Leonidas Guibas Johannes Wallner
KAUST, TU Wien Stanford University TU Wien Evolute GmbH, TU Wien TU Wien Stanford University TU Graz, TU Wien

Figure 1: Geodesic patterns on freeform surfaces. Left: For the cladding of a surface by wooden panels bendable only about their weak axis,
we perform segmentation into parts which can be covered by geodesic strips of roughly constant width. Right: A timber construction derived
from a hexagonal geodesic web has good manufacturing and statics properties.

Abstract 1 Introduction
Geodesic curves in surfaces are not only minimizers of distance, but In recent years it has become apparent that methods from Geomet-
they are also the curves of zero geodesic (sideways) curvature. It ric Computing bear a great potential to advance the field of freeform
turns out that this property makes patterns of geodesics the basic ge- architecture. This fact has created the new research area architec-
ometric entity when dealing with the cladding of a freeform surface tural geometry, which draws from various branches of geometry
with wooden panels which do not bend sideways. Likewise a geo- and which is motivated by problems originating in architectural de-
desic is the favored shape of timber support elements in freeform sign and engineering – see for instance the proceedings volume
architecture, for reasons of manufacturing and statics. Both prob- [Pottmann et al. 2008a]. The topics studied in the present paper
lem areas are fundamental in freeform architecture, but so far only belong to this line of research. They have as a common theme the
experimental solutions have been available. This paper provides a design of a pattern of geodesics on a freeform surface.
systematic treatment and shows how to design geodesic patterns in
different ways: The evolution of geodesic curves is good for local One problem concerns the cladding of a general double curved sur-
studies and simple patterns; the level set formulation can deal with face with wooden panels. Such claddings will be mainly applied
the global layout of multiple patterns of geodesics; finally geodesic to interior spaces (see Fig. 2). Even if the material may be differ-
vector fields allow us to interactively model geodesic patterns and ent from wood, the panels are assumed to be close to developable
perform surface segmentation into panelizable parts. and their development should fit well into a rectangle whose length
is much larger than its width. Hence, each panel should follow a
geodesic curve. The cladding problem can be approached in an ex-
CR Categories: I.3.5 [Computer Graphics]: Computational Ge-
perimental way as illustrated by Fig. 2. Computationally it means
ometry and Object Modeling—Geometric algorithms, languages,
decomposing a given surface into regions, each of which can be
and systems; I.3.5 [Computer Graphics]: Computational Geometry
covered by a sequence of nearly equidistant geodesic curves (see
and Object Modeling—Curve, surface, solid, and object represen-
Fig. 1).
tations
Other applications of geodesic patterns lie in wooden construc-
Keywords: computational differential geometry, architectural ge- tions where the geodesics are used for the supporting structure of
ometry, geometry of webs, timber rib shell, cladding, freeform sur- a curved shell. Extending pioneering technologies by J. Natterer
face, pattern, geodesic, Jacobi field. [2002], ongoing research at the EPF Lausanne aims at the design
of freeform timber rib shells, which are composed of a grid of geo-
desic curves (see Fig. 3). Other innovative timber constructions, as
seen in recent projects by Shigeru Ban (Fig. 4), would also benefit
from an efficient computational approach to the layout of geodesic
patterns on surfaces. One reason why geodesic curves are a pre-
ferred shape is statics: Geodesics – being minimizers of distance
– are the equilibrium shapes of elastic curves constrained to the
surface. Another reason is the manufacturing of laminated beams,
which are much easier to make if the individual boards can simply
be twisted and bent and along the weak axis [Pirazzi and Weinand
2006].
 Our approach to geodesic patterns is closely related to classical
results on the geometry of webs which was developed mainly by
W. Blaschke and his school. We refer to the survey article [Chern
1982] and the monograph [Blaschke and Bol 1938].

Contributions and overview. Inspired by practical problems in

architecture, we study geodesic N -patterns on surfaces. These are
formed by N discrete families of geodesics which are subject to
additional constraints arising from the specific application. We put
particular emphasis on the cases N = 1 (Fig. 2), N = 2 (Fig. 3)
Figure 2: Experimental cladding using paper strips (left) results in and N = 3 (Fig. 4). The overall goal is to provide an efficient
an office space design by NOX Architects [Spuybroek 2004]. computational framework to support the user’s navigation through
design space. Our results include:

Figure 3: Assem- * ways to control the strip width variation in a geodesic 1-pattern
based on geometric concepts such as Jacobi fields and striction
bling screw-laminated curves (Sec. 2);
beams for a timber rib
shell prototype based * two computational approaches for designing geodesic 1-patterns,
namely an evolution algorithm guided by Jacobi fields (Sec. 3)
on a 2-pattern of geo- and a level set approach which is also used for the design of gen-
desics. Image cour- eral N -patterns (Sec. 4);
tesy IBOIS, the tim-
ber construction lab at * the design of geodesic webs (special 3-patterns and 4-patterns)
and the extraction of further patterns from such webs, for instance
EPF Lausanne. in the style of Islamic art (Sec. 5);
* a user-friendly design tool for the solution of the cladding prob-
lem which results in an aesthetically pleasing segmentation of the
Related work. Let us briefly address the literature in geometric design surface into regions covered by geodesic 1-patterns. This
computing, as far as it is obviously related to our problem formula- tool is based on the concept of piecewise-geodesic vector fields
tion. We will later encounter more connections to previous work. (Sec. 6).
A geodesic curve g is a locally shortest path on a surface S. The Each of these approaches solves different instances of the problem
computation of geodesics is a classical topic: For smooth surfaces, of covering a surface with geodesic curves.
pursuing a geodesic curve emanating from a point with a given tan-
gent vector is equivalent to solving an initial value problem for
a 2nd order ODE, while the boundary value problem (connecting 2 Distances between geodesics
two points on a surface by a geodesic path c(t)) can be converted
into the constrained minimization of the quadratic energy kċk2
R In this section, we discuss general se-
[do Carmo 1976]. For triangle meshes, shortest polylines cross quences (1-patterns) {gi } of successive
edges at equal angles, and ambiguities at vertices may be resolved geodesic lines in a surface and the distances
by the concept of “straightest geodesics” [Polthier and Schmies between them. It is not usually possible to
1998]. Finding the truly shortest geodesic paths requires the com- express by formulas such a distance – say gi+1
gi
putation of distance fields, for which several efficient algorithms between the geodesic curve gi and its neigh-
have been developed, see for instance [Chen and Han 1996] or bor gi+1 . Not even an exact definition of ww
w
w
[Kimmel and Sethian 1998]. such a distance is straightforward. Only in
some highly symmetric cases, like evenly
Early research on the cladding of freeform surfaces with devel- distributed meridian curves of a rotational surface (see the figure)
opable panels evolved from the architecture of F. Gehry [Shelden the measurement of distance is elementary. However, a first order
2002]. One contribution to the present cladding problem is pro- approximation of that distance is well known: Start at time t = 0
vided by the geodesic strip models of Pottmann et al. [2008b]. with a geodesic curve g(s), which is parametrized by arc length s,
This name is used for continuous surfaces composed of devel- and let it move smoothly with time. A snapshot at time t = ε yields
opable strips which have nearly straight development. One can a geodesic g + near g (see Fig. 5):
view them as semi-discrete versions of smooth families of geode-
sics on a smooth surface. In general, several geodesic strip models g + (s) = g(s) + εv(s) + ε2 (. . . ).
with different directions are required to cover a freeform surface
(see Fig. 2). The choice of these directions and the initialization of The derivative vector field v is called a Jacobi field. It is known
the optimization in [Pottmann et al. 2008b] has not been systemati- that without loss of generality we may assume that it is orthogonal
cally investigated so far. to the curve g(s), and it is expressed in terms of the geodesic’s

Figure 4: Timber construction for v(s) Figure 5: A geodesic g(s) with

the Yeoju Golf Club by Shigeru Jacobi field v(s), and a neigh-
Ban; CAD/CAM by designtoproduc- boring geodesic g + (s) which is
tion. The beams are no geodesics and +
+ at distance ≈ εkv(s)k. The pa-
manufacturing thus requires CNC ma- ggg+
+
+ (s)
(s)
(s) rameter s is the arc length along
g 0 (s) the geodesic g, and v obeys the
chining. We address the design of 3-
patterns of geodesics and so contribute g(s) Jacobi differential equation (1).
to their simplified manufacturing. g(0) Here v(0) = 0.
tangent vector g 0 as geodesics with start at points s(t) with initial tangent vector v(t),
then s is (part of) the striction curve of this family of geodesics.
v(s) = w(s) · Rπ/2 (g 0 (s)), where w00 + Kw = 0. (1)
Knowing this, the design of geodesic patterns with prescribed stric-
Here Rα is the rotation by the angle α in the tangent plane of the tion curve s(t) is simple: Once the striction curve s is chosen, there
surface, and K is the Gaussian curvature [do Carmo 1976]. is only one remaining degree of freedom in the choice of the par-
allel vector field. From this 1-parameter family of patterns we can
Since distances between infinitesimally close geodesics are gov- take the one which best fits the design intent (see Fig. 7).
erned by (1), that equation also approximately governs the width of
a strip bounded by two geodesic curves of small finite distance.
Note that the Jacobi equation can also be used to fabricate sur- 3 Geodesic 1-patterns by evolution
faces of given Gaussian curvature, by gluing together strips of paper
whose width obeys the Jacobi differential equation (see Fig. 6). This section presents a strategy for the design and computation
of geodesic 1-patterns on triangle meshes which represent general
p freeform surfaces. It is based on an evolution of curves, where the
w(s) = α cosh(s |K|) transfer from a current geodesic curve g to the next one (denoted by
g + ) considers only the local neighborhood of g and can thus nicely
be governed by Jacobi vector fields.
Figure 6: Do it yourself K-sur-
face. We glue together strips Solution of the Jacobi differential equation. The computation
whose width w(s) fulfills the of all possible Jacobi vector fields v(s) orthogonal to a geode-
Jacobi equation for some con- sic curve which is given by an arc length parametrization g(s)
stant value K < 0. This results (0 ≤ s ≤ L) is easy. We only have to compute the function w(s)
in a surface of approximately which gives the length of the Jacobi field, and which satisfies the
constant Gaussian curvature. differential equation w00 + Kw = 0. This is a linear ODE with
coefficient function K = K(g(s)). Any solution w(s) is a linear
A first design method for geodesic 1-patterns. For the design combination
of a geodesic 1-pattern {gi } it is important to control the positions
of points at which the distance of the curve gi+1 from the curve gi w(s) = λ1 w(1) (s) + λ2 w(2) (s)
assumes a minimum or maximum. It turns out that it is not difficult of two linearly independent fundamental solutions w(1) and w(2) .
to design 1-patterns of geodesics where the locus of these points is
prescribed. The Jacobi relation (1) is the key to understanding the We have implemented the solution of the Jacobi differential equa-
local behavior of the strip width. tion as follows: The geodesic under consideration is represented as
a polyline p0 , p1 , . . . , pM with edge lengths Li = kpi+1 − pi k.
Assuming w > 0, we have w00 < 0 whenever the Gaussian curva- The unknown function w is given by its values wi = w(pi ) in the
ture is positive. Therefore in an area where K > 0, the strip width vertices pi . Now the second derivative w00 is approximated by
can have only maxima. If the Gaussian curvature is negative, then
the strip width can have only local minima. The case of constant 2 “ 1 1 ”
wi00 ≈ (wi+1 − wi ) − (wi − wi−1 ) .
strip width is only possible if K = 0, which means developable Li−1 + Li Li Li−1
surfaces.
In this way the Jacobi equation turns into a sparse linear system,
Setting aside the special case of developable surfaces, we get a pic- involving M − 1 equations for w100 , . . . , wM
00
−1 in M + 1 unknowns
ture of the locus of extremal strip width by first looking at the case w0 , . . . , wM . If we prescribe the pair (w0 , w1 ) of values, the sys-
of a smooth family {gt } of curves on a surface. The striction curve tem reduces to a triangular one. For the computation of the funda-
s is the locus of extremal width of the infinitesimal strips defined mental solutions w(1) and w(2) we simply choose these pairs to be
by two neighboring curves of the family (generically s is curve-like, (0, 1) and (1, 0), respectively.
but it may also degenerate). H. R. Müller [1941] showed the follow-
ing: If the curves gt are geodesics, their initial tangent vectors are Selecting a Jacobi field. The selection of the Jacobi field which
geodesically parallel along s. Conversely, if we compute a geodesi- is to guide the next geodesic g + depends on the design intent. If we
cally parallel vector field v(t) along a curve s(t) and then trace the are interested in constant strip width w ≈ W = const., we compute
λ1 , λ2 by minimizing

Fw = ∫[0,L] (λ1 w(1) (s) + λ2 w(2) (s) − W )2 ds, (2)

s
which amounts to a linear 2 × 2 system. Since we have only two
degrees of freedom, we cannot expect λ1 w(1) (s) + λ2 w(2) (s) ≈
sss const. in all cases. It is therefore advisable to check afterwards if
we really have w(s) ∈ [W − ε1 , W + ε2 ], for certain pre-assigned
tolerances ε1 , ε2 . See Figures 8a,d for an example.
Replacing the constant W by a prescribed function W (s) is compu-
tationally the same, but opens up many possibilities: The decompo-
Figure 7: The striction curve s of a geodesic 1-pattern. If the initial sition of a freeform surface into strips might require a denser sam-
tangent vectors of geodesics constitute a geodesically parallel vec- pling if the curvature across the strips is high. We therefore guide
tor field along s, then extremal distances between successive curves strip width by minimizing Fw with W (s) = φ(κn (s)), where φ
occur near s. Left: the piecewise-developable surface of Fig. 6. is a strictly decreasing function and κn is the surface’s normal cur-
Right: The areas of extremal distance of a geodesic curve from its vature in the direction orthogonal to the given geodesic. For an
neighbors are indicated by white color. This surface is taken from example, see Fig. 8b. Prescribed patterns of geodesic curves can be
the Cagliari museum project by Zaha Hadid architects. achieved by appropriate choices of W (s) (see Fig. 8c).
(a) (b) (c) (d) (e)

Figure 8: Evolution of geodesics, starting from a source curve (dark blue). (a) Regular equidistant evolution. Thin blue pieces of curves
show where the distance constraint is violated. (b) Aesthetic reasons might require denser strips if the normal curvature across the strip is
high. (c) Pattern transfer. The small image shows some intended width functions W (s). (d) Because of positive Gaussian curvature, in some
areas evolution is possible only if we allow intersections of adjacent geodesics. (e) Evolution of the same geodesic curve as in (d), but by
introducing breakpoints if the strip width deviates too much from a constant value. The breakpoints are guided along the red curves.

Computing the next geodesic. The function w(s) computed fact that we cannot have a proper non-intersecting next geodesic, or
above approximately describes the distance between the geodesic to consider broken geodesics, which are the topic of the following
curve g and the ‘next geodesic’ g + . We could now find g + by paragraph.
simply moving the endpoints g(0), g(L) sideways by the amounts
w(0), w(L), respectively, and connecting them by a geodesic g + .
The following method takes more information into account:

X1
X0 Y
V
g+
h0 h1
g
g(s0 ) g(s1 )
(a) (b)
i
We sample the original geodesic at parameter values si = N
L
(i = 0, . . . , N ) and move the points g(si ) sideways on geodesics Figure 9: Evolution of a 1-pattern of broken geodesics, which starts
hi orthogonal to g. This results in points Xi . The next geodesic g + with an unbroken one (blue, at extreme left). The threshold for
is represented by a point Y which slides along a curve orthogonal to the introduction of breakpoints in (a) and (b) is that the strip width
the current geodesic g and by an initial tangent deviates more than 7.5% or 5%, resp., from the desired value (this
Pvector V . These two surface is taken from the top of the Lilium tower, Warsaw, by Zaha
degrees of freedom are determined such that i dist(g + ∩hi , Xi )2
(each distance measured along hi ) becomes minimal. Being close Hadid Architects).
to the solution already, this can be done by a Levenberg-Marquardt
method which avoids second derivatives. Extension to broken geodesics Evolving a geodesic curve into
a 1-pattern frequently runs into obstacles, due to the few degrees
Limitations of the evolution method. The method above which of freedom available. We therefore relax the geodesic condition
finds a smooth geodesic at a certain nonzero distance from a given and consider broken geodesic curves. Such a curve g consists of
one works only if one can find a width function w which solves geodesic arcs gi defined in arbitrary subintervals [si , si+1 ]. These
the Jacobi equation and which has no zeros in the considered in- arcs fit together at points Pi = gi−1 (si ) = gi (si ) = g(si ). We
terval. It turns out that we can tell the existence of such ‘useful’ evolve each arc gi separately, with its own width function wi (s)
solutions simply by testing if the fundamental solution w(1) has an- which satisfies the Jacobi equation. Obviously, the single functions
other zero in the considered interval. The proposition below, proved wi are not independent. We draw the following picture:
in [do Carmo 1992], characterizes the two possible cases (good and
Pi+ = g + (si )
bad) and sums up some of their geometric properties.
g + (si+1 )
P ROPOSITION 1. Consider a geodesic curve g(s), where s ∈ [0, L] +
gi−1 α
α
αiiiiii βββiiiiii
s i)

si )

is an arc length parameter. Assume a fundamental solution w(1) (s)

1(

wi (
w i−

of the Jacobi equation with w(1) (0) = 0. Then there are the fol-
)
+1
) 1

lowing two cases:

i
(s i−

(s
wi

case 1 case 2 Pi = g(si )

w i−1

# zeros of w(1) (s) for 0 < s ≤ L: 0 ≥1 gi−1

g(si+1 )
# zeros of any solution in the interval [0, L]: ≤1 ≥1
existence of solution nonzero in [0, L] yes no With the angles αi , βi between curve normals and the line connect-
g locally minimizes distance of g(0), g(L) yes no ing breakpoints, we read off the approximate relation
The inequality K(s) > (π/L)2 for s ∈ [0, L] implies case 2. Anal- wi−1 (si ) wi (si )
ogously K(s) < (π/L)2 implies case 1 (surely true if K ≤ 0). = , (3)
cos αi cos βi
If the geodesic curve under consideration is a case 2 curve we have which assumes that the derivatives of wi−1 , wi are small. We can
two choices (illustrated by Figures 8d,e): Either to put up with the now find the single width functions wi by first choosing the di-
(a) (b) (c) (d)

κg = 10 1 0.1 0.01

Figure 10: 1-patterns of geodesic curves which are found as level sets of a real-valued function defined on the surface. (a) Optimization with
a low weight on Fw yields almost true geodesics. Three geodesics are shown in orange for comparison. (b) A higher weight on Fw generates
strips of even width, but lets level sets deviate from true geodesics and creates higher geodesic curvatures. (c) and (d): Deviation from the
geodesic property for subfigures (a) and (b). We show the geodesic curvature κg on a logarithmic scale; bounding box diagonal equals 4.

rection of movement of the breakpoints, reading off the angles αi , We thus optimize φ by solving a nonlinear least squares problem
βi , and subsequently minimizing the sum of all Fwi as defined by which is governed by a linear combination of three functionals:
(2) with (3) as side condition. The ‘next’ geodesics gi+ are sub- Fκ penalizes deviation from zero geodesic curvature, F∆ is for
sequently fitted to the width functions in a manner similar to the smoothing/regularization, and Fw penalizes deviation from the de-
computation of the next geodesic which is described above. With sired strip width w. Definitions are:
the path of breakpoints already prescribed, the number of degrees
of freedom for the broken geodesic g + is the number of segments
X ` ∇φ ´2
Fκ = A(v) div (v) ,
plus one. v∈V k∇φk
X
In the examples of Figures 8e and 9 breakpoints are automatically F∆ = area(S) A(v)∆φ(v)2 ,
v∈V
inserted whenever a strip would violate the distance constraint, and X “ h ”2
the paths of breakpoints are guided such that they bisect the angle Fw = area(f ) k∇φ(f )k − .
of their adjacent geodesic segments. f ∈F w
Here A(v) means the area of the Voronoi cell of v. All three func-
4 Geodesic N -patterns from level sets tionals are invariant w.r.t. scaling of the geometry. We therefore
formulate our optimization as
The Jacobi field approach is well suited if, in a 1-pattern, we want
to move from one geodesic to the next, in a way which allows us Fκ + λF∆ + µFw → min . (6)
to control the distance between these two geodesics. For global
tasks such as an optimal alignment of a 1-pattern of geodesics with We initialize optimization by a function which minimizes F∆ alone,
a vector field, or design problems involving several 1-patterns, we under the side condition of 3 arbitrary function values. Results of
prefer to represent the geodesics of a 1-pattern as selected level sets this method can be seen in Fig. 10. As is to be expected we cannot
of a real valued function φ which is defined on the given surface S. make make Fκ and Fw vanish at the same time: Only for devel-
opable surfaces (K = 0) the geodesic property is compatible with
Geodesic level sets are not new, in fact they represent the main idea constant width between curves.
in the geodesic active contour method of Caselles et al. [1997]. The
difference to our setting is that we consider global patterns of geo- For an N -pattern of geodesics, each of the N families of curves in-
desics. For level set methods in general we refer to [Osher and volved is represented as level sets of a function φi (i = 1, . . . , N ).
Fedkiw 2002]. We now minimize the sum of the single target functions analogous
to (6). We can also incorporate additional requirements, for in-
Geodesic curves are characterized by vanishing geodesic curvature stance constant intersection angle α between families i and j. To
κg [do Carmo 1976]. If a curve is given in implicit form as a level this end we augment the target functional by
set φ = const., then its geodesic curvature can be computed by
“ ∇φ ” X area(f ) D ∇φi ` ∇φj ´E2
κg = div , (4) ν · Fangle = ν , R π2 −α .
k∇φk area(S) k∇φi k k∇φj k
f ∈F

where div, ∇ are understood in the intrinsic geometry of S (see e.g. Here R π2 −α means a rotation by the angle π2 − α in the respective
[do Carmo 1992], p. 142).
face; ν is the weight given to Fangle . For an example see Fig. 16.
We implement construction of a function φ as follows. The surface
The level set approach is an integral part of the examples and im-
S is represented by a triangle mesh (V, E, F ). The function φ is
ages in Sections 5 and 6.
determined by its values on vertices and is considered to be linear
within each triangular face. The level sets under consideration are Remark 1. Our level set approach works only for simply connected
polylines on the mesh. The vector field ∇φ is constant in each face. surfaces. It is possible to extend it to arbitrary surfaces by using
For any vector field X we evaluate div(X) at a vertex v by com- branched coverings, such as in [Kälberer et al. 2007].
puting the flux of X through the boundary of v’s intrinsic Voronoi
cell, divided by the area of that cell. Implementation details. To compute the function φ which min-
For regularization, we wish to keep ∆φ = div(∇φ) small. Further, imizes the combined functional (6), we employ a classical Gauss-
applications might require the distance between level sets φ = c Newton method with line search for optimization [Madsen et al.
and φ = c + h to equal some value w; for that distance we have 2004]. The variables of this optimization are the values of the un-
‹ known function φ on the vertices of a mesh. All required first order
w ≈ h k∇φk. (5) derivatives are calculated analytically (i.e., are not approximated).
(a) (b) (c) (d) (e)

Figure 11: A hexagonal web of geodesics imposed on the surface of Fig. 6. The level sets of functions φ1 , φ2 , φ3 are geodesics, if they obey
div(∇φi /k∇φi k) = 0. If they also obey φ1 + φ2 + φ3 = 0, they form a hexagonal web. (a) Integer level sets √ for each of φ1 , φ2 , φ3 and true
geodesics for comparison. (b) Flattening of the surface by the mapping ψ(x) = (φ1 (x), (φ1 (x) + 2φ2 (x))/ 3). (c)–(d): Using ψ as texture
mapping, patterns with hexagonal combinatorics can be transferred to the surface. (e) Detail of a geometric pattern inspired by Islamic art.
Long and thin components follow geodesic curves on the surface and can therefore be manufactured by bending of wooden panels.

The linear systems to be solved in each round of iteration are sparse, and φ2 to lines x = const. and y = const., resp., while the level
since the single terms which contribute to (6) involve only local dif- sets of φ3 are mapped to the lines x + y = const. Equation (7) is
ferential operators which are defined in terms of small vertex neigh- known in the theory of webs; in fact every hexagonal web can be
borhoods. For sparse Cholesky factorization we employ C HOLMOD found in this way.
[Chen et al. 2008].
We implement geodesic webs as follows: We consider two func-
tions φ1 , φ2 on the given surface S, and optimize them such that
5 Geodesic webs the sum of target functionals according to (6), evaluated for φ1 , φ2 ,
and φ3 := −φ1 − φ2 becomes minimal. For results, see Figure 1
Timber constructions like the one of Figure 4 follow a curve pattern (right) and Figure 11. A different kind of web is shown by Fig. 19.
with regular combinatorics. For manufacturing and statics reasons,
one would like these guiding curves to be geodesics. Questions
of this kind immediately lead us to the investigation of systems
of curves on surfaces with both geometric and topological proper-
ties. The most important concept here is the hexagonal web, which
means 3 families of curves which can be continuously mapped to 3
families of parallel straight lines in the plane, as shown by Fig 11b.
By selecting a discrete sample of curves from the web we can gener-
ate a variety of patterns – see Figures 1 (right), 4, and 11. The con- (a) (b)
dition that families of curves form a hexagonal web is of a topolog-
ical nature and belongs to the so-called geometry of webs [Blaschke
Figure 12: Hexagonal webs and class 3 curves. (a) The condi-
and Bol 1938; Chern 1982]. Neglecting some details concerning
P that a straighti line
tion u0 + u1 X + u2 Y = 0 is in the web reads
the domain where a web is defined, we quote a result: j k
i+j+k=3 aijk u0 u1 u2 = 0. This equation characterizes the tan-
i,j,k≥0
T HEOREM 2. A complete hexagonal web of straight lines in the
gents of a class 3 curve, and is determined, up to a factor, by 9
plane consists of the tangents of an algebraic curve of class 3 (pos-
given lines in general position. The colors given to the tangents
sibly reducible); any class 3 curve yields a hexagonal web [Graf
identify the 3 families of a hexagonal web. (b) A geodesic web on
and Sauer 1924]. The variety of geodesic hexagonal webs in a sur-
the sphere. It is transformed to a planar one by projection from the
face is the same as in the plane ⇐⇒ the surface has constant
center.
Gaussian curvature [Mayrhofer 1931].

In summary this means that there is a 9-parameter family of pos-

sible hexagonal webs made from geodesics, if the surface under
6 Global solution of the cladding problem
consideration is the plane, or its Gaussian curvature K is constant
When the input surface is more complicated, it is harder to cover it
(see Fig. 12). Note that such surfaces possess mappings into the
with a single geodesic pattern. We have already encountered such
plane where geodesics become straight lines; such surfaces carry
difficulties in Section 3 where we discussed the evolution of a geo-
one of the non-Euclidean geometries if K 6= 0.
desic curve. We saw that the Gaussian curvature of the input surface
Little seems to be known about surfaces which carry a smaller va- is responsible for the maximal length of a strip which is bounded by
riety of hexagonal webs made from geodesics. Anyway, for our geodesic curves. We are thus led to the question of segmentation of
applications it is more important to have a computational tool for the input surface such that each piece can be covered by a geodesic
generating a hexagonal web whose curves are as geodesic as possi- 1-pattern.
ble. The level set approach is well suited for that. We describe the
For this purpose we employ a device not used in previous sec-
three families of curves as level sets of functions φ1 , φ2 , φ3 with
tions, namely geodesic vector fields and piecewise-geodesic vec-
φ1 + φ2 + φ3 = 0. (7) tor fields. This section describes a general setup related to geode-
sic vector fields, shows how to interactively design a near-geodesic
This equation results in a hexagonal web: Mapping the surface S vector field by means of a certain reduced eigenbasis of fields, and
into the plane via x 7→ (φ1 (x), φ2 (x)) maps the level sets of φ1 demonstrates how to subsequently modify (sharpen) it to make it
(a) (b) (c) (d)

Figure 13: Processing pipeline for the global cladding problem. (a) The first two elements of the reduced basis which spans the vector field
design space. (b) User’s selection vf∗1 , vf∗2 , . . . indicated by arrows, and blue design vector field v◦ adapted to this selection. (c) Sharpened
vector field v which is now piecewise geodesic together with the boundaries of macro patches which lie where the vector field is sharp. (d)
Segmentation into finitely many geodesic 1-patterns which are aligned with the user’s selection. This surface is taken from the interior facade
of the Heydar Aliev Merkezi Project by Zaha Hadid Architects.

piecewise geodesic. Finally, we perform segmentation of the input In summary, the vector field v = (vf )f ∈F is a geodesic vec-
surface along the curves where the resulting vector field is sharp. tor field, if its length kvf k is constant for all faces f , and we
To find the actual geodesic curves defined by the vector field, we can find a collection of coefficients g = (gf )f ∈F with gf =
refer back to the level set method of Section 4. (g1,f , g2,f , g3,f ), such that (8) holds with the r’s below some
threshold.
Geodesic vector fields. A vector field v on a surface S is geode-
sic if it consists of tangent vectors of a 1-parameter family of geode- The following functional attempts to quantify how well v satisfies
sic curves covering S. Later we also encounter piecewise-geodesic the geodesic property. It will turn out to be very useful for the next
fields which fulfill the geodesic property in the inside of certain task (interactive editing of geodesic vector fields). We let
patches. The following characterization of the geodesic property
“X ”
Q(v) = min we kre (v, g)k2 + λr kgk2 . (9)
employs the notation ∇x (v), which means the covariant derivative g∈R3|F |
e∈E
of a vector field v in direction of the tangent vector x, and which
is defined as the tangential component of the ordinary directional The term including λr is for regularization; we choose the factor
derivative [do Carmo 1976]. λr proportional to a characteristic edge length in the mesh. The
P ROPOSITION 3. A vector field v of constant length is geodesic edge weights we ≥ 0 are for downweighting areas where we do
⇐⇒ for all points of the surface, the linear mapping x 7→ ∇x (v) not care about the geodesic property and where we (later) want to
in the tangent plane is symmetric. encourage formation of a patch boundary. We chose to downweight
areas of high curvature; so we let we = exp(−κ2e /2µ2 ), where κe
Proof: kvk = const. implies that ∇x (v) is orthogonal to v, for is the normal curvature across the edge e, and µ is computed as a
all x. The geodesic condition is ∇v (v) = 0 everywhere. Among median of all absolute values |κe |. We say that these weights are
the linear mappings which map the entire tangent plane to v⊥ , the median-weights w.r.t. the mapping e 7→ |κe |.
symmetric ones are exactly those which map v to zero.  Since the re ’s which contribute to the value Q(v) depend only on
v, we can also write re (v). The following turns out to be important:
Implementing geodesic vector fields. In our implementation,
the surface S is represented as a triangle mesh (V, E, F ). The vec- P ROPOSITION 4. There is a positive semidefinite symmetric matrix
tor field is represented by unit vectors vf attached to the incenters H, which depends on the given mesh, such that Q(v) = vT Hv.
mf of faces f ∈ F . Consider two adjacent faces f1 , f2 , such that
the face f2 has been unfolded into the plane of f1 (in the following Proof: The bracket expression in (9) has the general form gTAg +
text, the unfolded positions of items associated with f2 are marked 2gTBv + vTCv with symmetric A, B, C =⇒ its minimum is
with a hat). In order to capture the condition of Prop. 3, we endow attained for g = −A−1 Bv, so Q(v) = vT(C − BA−1 B)v. Pos-
each face with a local coordinate system and a Jacobi matrix Jf itivity is obvious from Q(v) ≥ 0. 
such that
Interactive vector field selection. We show how the user can in-
vf1 bf2 = vf1 + Jf1 · (m
v b f2 − mf1 ) + re ,
teractively design a piecewise geodesic vector field, by prescribing
“g
1,f g2,f
” the values
mf1 m where Jf = . (8) vf∗1 , . . . , vf∗k (10)
b f2 g2,f g3,f
e
of that field in user-selected faces f1 , . . . , fk . This procedure,
which is described in the following paragraphs and which is mo-
Here re is a remainder term associated with the edge e = f1 ∩ f2 . tivated by [Huang et al. 2009], first amounts to choosing a vector
field v◦ which is not exactly geodesic but has a reasonably small
This manner of discretizing the symmetry of the covariant deriva-
value of Q(v◦ ). In a subsequent step v◦ is sharpened so as to be-
tive – as postulated by Prop. 3 – comes from two facts: (i) covariant
come a piecewise geodesic vector field v.
differentiation is invariant if an isometric deformation is applied, so
we may unfold the neighboring triangles into the plane of the trian- In order to achieve interactive editing rates, we precompute a re-
gle under consideration; (ii) in a plane, ∇x (v) equals the ordinary duced basis v(1) , . . . , v(n) of vector fields and and try to realize
directional derivative of v w.r.t. the vector x, i.e., it equals multi- the user’s selection by a linear combination
plication of a Jacobi matrix with x. The condition of Prop. 3 now
means symmetry of the Jacobi matrix. v◦ = x1 v(1) + · · · + xn v(n) .
of these eigenvectors alone. Our choice is to take the first n eigen- The face weights wf offer the possibility to keep v close to v◦
vectors of the quadratic form Q (i.e., eigenvectors of the matrix where v◦ already is a geodesic vector field – indicated by smallness
◦
H as defined in Prop. 4). In order to determine x1 , . . . , xn and P re (v ). We◦ employ median-weighting w.r.t. the mapping f 7→
of
to reinforce the condition kvf◦ k = 1 which is not satisfied by the e∈∂f kre (v )k.
eigenvectors, we employ two steps in an alternating way: (i) we let
v̄f◦ = vf◦ /kvf◦ k, and (ii) we determine x1 , . . . , xn such that Surface segmentation and pattern layout. Having found a
piecewise geodesic vector field v, we now define patches by cut-
k
X X ting along the edges where v is sharp. We first collect all edges
Fprox = kvf◦i −vf∗i k2 +λr kv◦ k2 +λn kvf◦ −v̄f◦ k2 → min . e = f1 ∩ f2 where the angle between vf1 and vf2 is greater than a
i=1 f ∈F
threshold value α; such edges indicate patch boundary curves (we
The weights λn , λr govern the influence of normalization and use α = 20◦ ). We then use the method of Pauly et al. [2003] to
regularization terms. The choice of these values is not critical; polish these curves. We omit details since we do not consider seg-
we used λn = 0.1 and λr = 0.01. We initialize this itera- mentation a new result. Segmentation being completed, we end
tion by letting λn = 0 in the first round. Each round amounts up with patches P1 , . . . , PM which we know can be covered by a
to solving a linear system and a matrix-vector multiplication, and smooth geodesic vector field. The layout of evenly spaced geode-
takes O(n3 ) + O(n · |F |) time. For instance, with n = 60 and sics within each Pj is done according to Section 4, augmenting the
|F | = 40 000 we experienced 10 ms per round on a 2 GHz PC. target functional (6) by
Further, we found 3–5 rounds to be sufficient. Thus, vector field X X
h∇φ(f ), vf∗i i2 + λsharp
(j)
selection can be performed in real time. Falign = λuser h∇φ(f ), vf i2 .
i:fi ∈Pj f ∈Pj
Remark 2. The number n determines the degrees of freedom of-
(j)
fered to the user; we found n ≈ 50 to work well. Guidance is given Minimizing Falign means that the geodesics occurring as level sets
by the magnitude of the first n eigenvalues which should be small, of φ are aligned with the sharpened vector field v and/or with the
and by the complexity of the object under consideration. Choos- user’s selection {vf∗i }. The corresponding weights λsharp and λuser
ing the first eigenvectors of Q to span our design space amounts to have to be set accordingly. Results are shown in Figs. 1, 13, and 15.
choosing a design space where Q(v◦ ) has small values under the
‘wrong’ and geometrically meaningless normalization constraint
kv◦ k2 = 1. As all vector fields undergo further treatment any- Figure 15: Instead of properly
way, this wrong normalization does not matter. Our purpose was to segmenting a surface into parts
find a design space which contains enough degrees of freedom, and which can be covered by smooth
which is taken from the low frequency end of the spectrum (thus geodesic 1-patterns, we may look
avoiding unreasonably small patch sizes). for a weaker solution: cutting it
open along curves such that it can
be covered by 1-patterns which
are smooth everywhere except at
Figure 14: Sharpening a vector field the cuts. The present example is
(left) such that it becomes piecewise almost a true segmentation with
geodesic (right). This is a detail of only 1 dangling edge.
Fig. 13.
General remarks. There are several reasons why the interactive
Vector field sharpening. Segmentation of the given surface S procedure of Section 6 is not only nice to have in applications, but
into patches is based on piecewise geodesic vector fields, which is actually necessary: First, the highly nonlinear minimization of
fulfill the geodesic property in the inside of certain patches (i.e., re the function F (v, g) in Equation (11) will typically get stuck in
is small there), but we allow high values of kre k along the patch local minima, so we cannot expect that unguided minimization of
boundaries. We assume that a vector field v◦ is given, and we wish F (v, g) succeeds. Second, there is a great variety of local minima
to find another vector field v which is close to v◦ , but is piecewise which cannot clearly be distinguished by the magnitude of F (v, g)
geodesic. We set up an optimization problem as follows: alone. For this reason it is necessary to let the user choose.

F (v, g) = λv Fdist (v) + λr kgk2 + Fgeod (v, g) → min, (11) 7 Discussion

X x2
Fgeod (v, g) = we ρ(kre (v, g)k), ρ(x) = , (12) Architectural applications are not limited to timber structures, as the
e∈E
1 + αx2
X static properties of geodesics apply to any material. An additional
Fdist (v) = wf kvf − vf◦ k2 . (13) bonus is that beams which follow geodesics, being shortest, need
f ∈F less material than other shapes. Fig. 16 shows an example, featuring
a 2-pattern of geodesics which is designed for realization in steel
The function ρ could be any of the heavy-tailed functions used for rather than in wood.
image sharpening (see for instance [Levin et al. 2007]). Its purpose
is to push the deviations to accumulation areas. We used the robust Comparing the vector field method with other approaches.
estimator of Geman and McClure [1987], with α = 100 in all ex- The segmentation procedure of Section 6 is governed by both user
amples. Global weights λr , λv which govern the influence of the interaction and Gaussian curvature, since it produces patches inside
regularization term and the proximity of v to v◦ have to be set ac- which the Gaussian curvature has a nice behaviour — ‘nice’ being
cording to the application. We maintain the condition kvf k = 1, defined in terms of the capability of being covered with a 1-pattern
so the variables in this optimization are, besides g ∈ R3|F | , the of geodesic curves. Segmentation of surfaces driven by Gaussian
collection of angles θf which define the vector vf in the local coor- curvature as proposed by [Yamauchi et al. 2005] for even distribu-
dinate system of the face f . The optimization problem is solved in tion of Gaussian curvature or by [Julius et al. 2005] for near-de-
the same way as that of Section 4. velopable patches probably produces patches which can be covered
 decision for equal spacing rather than for the geodesic property).
This phenomenon can also cause the segmentation process of Sec-
tion 6 to produce unsatisfactory results, in which case it has to be
iterated (see Fig. 17).
Geodesic webs pose many constraints on the involved curves and
in fact we cannot expect them even to exist in a mathematically
exact way on arbitrary surfaces. Our level set approach produces
webs of curves which are as geodesic as possible, but deviations
Figure 16: Relations between fami- of level curves from true geodesics are inevitable (see Fig. 18). As
lies of geodesics. The level set ap- mentioned in Theorem 2, on constant Gaussian curvature surfaces
proach is capable of accommodating we have the same variety of geodesic webs as for straight line webs
additional requirements, such as a de- in the plane. This is verified by the fact that the level curves in
sired intersection angle α between the Figures 11 and 19 are indistinguishable from true geodesics.
level sets of functions φ1 , φ2 . This 2-
pattern of geodesic curves is imposed
on the Warsaw Lilium Tower by Zaha
Hadid Architects. True geodesics in
orange are shown for comparison.

by geodesic 1-patterns. The method of the present paper, however,

directly works with geodesics in a user-controlled way; Gaussian
curvature is present only implicitly.
This paper contains another way of segmenting a surface into
patches which can be covered by almost equidistant geodesics,
namely the evolution of broken geodesics as illustrated by Fig. 9 (it Figure 17: A situation where two rounds of segmentation are nec-
even yields a consistent spacing of curves along the patch bound- essary because of the unsatisfactory result of the first round.
aries). The method of Section 6 is much more flexible, however,
and treats each part of the surface in the same way, independent of
an initial choice of geodesic to evolve from.
Figure 18: A hexagonal web of
Computation details and timings. For the level set approach, near-geodesics. We show some
details on the choice of weights, the number of variables, etc., as curves of the web Fig. 1 is based
well as timings are given by Table 1. For the geodesic vector field on. Due to the strong variation
approach, the choice of weights is detailed by Section 6. Timings in Gaussian curvature they de-
are given by Table 2 for three examples, each of which have 40k viate from true geodesics (thick
faces. Preprocessing (eigenspace computation with Arpack) needs curves).
30–40 seconds. One round of sharpening takes 1.1 seconds, and
the final segmentation needs about 1 second. That table also shows One major motivation for geodesic patterns is the cladding of
times for the evolution of broken geodesics according to Section 3. freeform surfaces by thin wooden panels. This cladding problem
Here about 70% of the time is used for the piecewise fitting of has been considered previously by Pottmann et al. [2008b], in the
curves, the rest is postprocessing like merge, trim, etc. context of developable strips which certainly are a shape which thin
panels can assume. In that paper the importance of curve’s tangents
Fig. var iter. sec Fκ F∆ Fw Fangle λ µ ν staying away from asymptotic directions was mentioned. We could
1 33k 4 32 .25 20 10−2 0 0 incorporate this condition in our level set approach, but we have
10a 10k 10 16 .0006 2470 25 10−8 10−4 0 completely neglected it in this paper. In the negatively curved areas
10b 10k 11 17 .71 392 .0007 10−4 10+4 0 of surfaces, it would very much obstruct our available degrees of
freedom. The reason why we chose not to aim at developable strips
11 21k 12 77 .005 145 10−6 0 0
is that they are not the shape that bent wooden panels assume in
16 159k 10 419 .66 1050 .26 3·10−4 0 2.8
general; such panels can be twisted (for other materials, however,
this additional aspect has to be observed to a higher degree).
Table 1: Details for the level set approach. We show the number of
variables, number of iterations, total time for a 3GHz PC, the values Conclusion and future research. Motivated by problems in
of functionals and the weights used in optimization. freeform architecture we have described three different approaches
to the layout and interactive design of geodesic patterns on surfaces.
Figure 9a 9b 1 (left ) 13 17 (left) Each of the methods treated in this paper has its specific strength:
seconds 86 113 102 85 73 The Jacobi field approach to the evolution of geodesic curves is best
Method Evolution Vector fields suited to deal with local issues. The level set approach can deal with
the global layout of patterns very well, and is efficient in dealing
Table 2: Representative timings for evolution of geodesics, and for with multiple patterns of geodesic curves which are in some rela-
the geodesic vector field method. Data apply to a 2GHz PC. tion, such as the condition that they form a hexagonal 3-web. The
vector field approach is capable of solving the global layout and
segmentation problem in an interactive way.
Limitations. Since in a fixed surface the geodesics are only a two-
parameter family of curves, often the designer’s request cannot be This paper seems to be the first where the classical geometry of
met and one has to compromise (see Fig. 10b for an example of a webs is employed in a geometry processing context. It is likely that
this theory has yet more potential and applications, in particular in M ÜLLER , H. R. 1941. Über die Striktionslinien von Kurvenscha-
architecture. ren. Monatsh. Math. Phys. 50, 101–110.
NATTERER , J., B URGER , N., AND M ÜLLER , A. 2002. The roof
Acknowledgments. This research has been supported by the structure “Expodach” at the world exhibition Hannover. In Proc.
Austrian Science Fund (FWF) under grants No. S92-06 and S92- 5th Intl. Conf. Space Structures, 185–193.
09 (National Science Network Industrial Geometry), and by the
European Community’s 7th Framework Programme under grant O SHER , S. J., AND F EDKIW, R. P. 2002. Level Set Methods and
agreement 230520 (ARC). The authors gratefully acknowledge the Dynamic Implicit Surfaces. Springer Verlag.
support of NSF grants 0808515 and 0914833, of NIH grant GM-
072970, and of a joint Stanford-KAUST collaborative grant. We PAULY, M., K EISER , R., AND G ROSS , M. 2003. Multi-scale
want to express our thanks to Zaha Hadid Architects, London, to feature extraction on point-sampled surfaces. Comput. Graph.
be able to work on data which come from some of their current Forum 22, 3, 281–290.
projects. P IRAZZI , C., AND W EINAND , Y. 2006. Geodesic lines on free-
form surfaces: optimized grids for timber rib shells. In Proc.
References World Conference on Timber Engineering. 7pp.
P OLTHIER , K., AND S CHMIES , M. 1998. Straightest geodesics on
B LASCHKE , W., AND B OL , G. 1938. Geometrie der Gewebe. polyhedral surfaces. In Mathematical Visualization, Springer,
Springer. H.-C. Hege and K. Polthier, Eds., 391–409.
C ASELLES , V., K IMMEL , R., AND S APIRO , G. 1997. Geodesic P OTTMANN , H., H OFER , M., AND K ILIAN , A., Eds. 2008. Ad-
active contours. Int. J. Comput. Vision 22, 1, 61–79. vances in Architectural Geometry. Proc. of Vienna conference.
C HEN , J., AND H AN , Y. 1996. Shortest paths on a polyhedron. I. P OTTMANN , H., S CHIFTNER , A., B O , P., S CHMIEDHOFER , H.,
Computing shortest paths. Int. J. Comput. Geom. Appl. 6, 127– WANG , W., BALDASSINI , N., AND WALLNER , J. 2008.
144. Freeform surfaces from single curved panels. ACM Trans.
Graphics 27, 3, #76, 1–10.
C HEN , Y., DAVIS , T. A., H AGER , W. W., AND R AJAMAN -
ICKAM , S. 2008. Algorithm 887: C HOLMOD , supernodal sparse S HELDEN , D. 2002. Digital surface representation and the con-
Cholesky factorization and update/downdate. ACM Trans. Math. structibility of Gehry’s architecture. PhD thesis, M.I.T.
Softw. 35, 3, #22, 1–14. S PUYBROEK , L. 2004. NOX: Machining Architecture. Thames &
C HERN , S. S. 1982. Web geometry. Bull. Amer. Math. Soc. 6, 1–8. Hudson.
S UTTON , D. 2007. Islamic Design: A Genius for Geometry.
DO C ARMO , M. 1976. Differential Geometry of Curves and Sur-
Walker Publ. Comp.
faces. Prentice-Hall.
YAMAUCHI , H., G UMHOLD , S., Z AYER , R., AND S EIDEL , H. P.
DO C ARMO , M. 1992. Riemannian Geometry. Birkhäuser. 2005. Mesh segmentation driven by Gaussian curvature. Vis.
G EMAN , S., AND M C C LURE , D. E. 1987. Statistical methods for Computer 21, 659–668.
tomographic image reconstruction. Bull. Inst. Internat. Statist.
52, 4, 5–21.
G RAF, H., AND S AUER , R. 1924. Über dreifache Geradensysteme. (a) (b)
Sitz. Bayer. Akad. Math.-nat. Abt., 119–156.
H UANG , Q., W ICKE , M., A DAMS , B., AND G UIBAS , L. 2009.
Shape decomposition using modal analysis. Comput. Graph. Fo-
rum 28, 2, 407–416.
J ULIUS , D., K RAEVOY, V., AND S HEFFER , A. 2005. D-charts:
Quasi-developable mesh segmentation. Comput. Graph. Forum
24, 3, 581–590.
K ÄLBERER , F., N IESER , M., AND P OLTHIER , K. 2007. Quad-
Cover – surface parameterization using branched coverings.
Comput. Graph. Forum 26, 3, 375–384.
K IMMEL , R., AND S ETHIAN , J. A. 1998. Computing geodesic
paths on manifolds. PNAS 95, 8431–8435.
L EVIN , A., F ERGUS , R., D URAND , F., AND F REEMAN , W. T. Figure 19: Geodesic 4-pattern. For a combination of two hexag-
2007. Image and depth from a conventional camera with a coded onal webs, functions φ1 , φ2 are optimized such that level sets of
aperture. ACM Trans. Graphics 26, 3, #70, 1–9. φ1 , φ2 , φ1 ± φ2 are geodesic. (a) Level sets (thin) and true geode-
sics (thick). Level sets of all four functions are almost truly geo-
M ADSEN , K., N IELSEN , H. B., AND T INGLEFF , O., 2004. Meth-
desic, because the Gaussian curvature is nearly constant for this
ods for non-linear least squares problems. Lecture Notes. http://
particular surface. (b) Transfer of an Islamic art pattern according
http://www.imm.dtu.dk/courses/02611/nllsq.pdf.
to [Sutton 2007] to the surface by the simple texture coordinates
M AYRHOFER , K. 1931. Sechseckgewebe aus Geodätischen. Mo- u = φ1 , v = φ2 . The four principal directions in the pattern follow
natsh. Math. Phys. 38, 401–404. geodesic curves.