Volume 25 (2006), number 4 pp.

789–807 COMPUTER GRAPHICS forum

Differential Representations for Mesh Processing

Olga Sorkine

School of Computer Science, Tel Aviv University, Israel

sorkine@gmail.com

Abstract
Surface representation and processing is one of the key topics in computer graphics and geometric modeling,
since it greatly affects the range of possible applications. In this paper we will present recent advances in geometry
processing that are related to the Laplacian processing framework and differential representations. This framework
is based on linear operators defined on polygonal meshes, and furnishes a variety of processing applications, such
as shape approximation and compact representation, mesh editing, watermarking and morphing. The core of the
framework is the definition of differential coordinates and new bases for efficient mesh geometry representation,
based on the mesh Laplacian operator.
Keywords: discrete Laplacian, surface representation, detail preservation, geometry compression, mesh editing
ACM CCS: I.3.5 Computer Graphics: Computational Geometry and Object Modeling — Boundary representa-
tions

1. Introduction meshes are studied through their differential properties, de-

rived from certain linear operators defined on the mesh. These
Surface representation and processing is one of the key top- operators are usually different variants of the mesh Laplacian,
ics in computer graphics, geometric modeling and computer- and they provide means to represent a surface by new bases
aided design. To put it simply, the way a 3D object is defined that benefit various processing applications. In contrast to the
greatly affects the range of things one can do with it. Different traditional global Cartesian coordinates, which can only tell
object representations may reveal different geometric, com- the spatial location of each point, a differential surface rep-
binatorial or even perceptive properties of the shape at hand. resentation carries information about the local shape of the
For instance, the very popular piecewise-linear surface rep- surface, the size and orientation of local details. Therefore,
resentation (triangular meshes) provides immediate means defining operations on surfaces that strive to preserve such
to display the surface, learn about its topological properties a differential representation, results in detail-preserving op-
(through Euler’s characteristic) and, with some effort, differ- erations. The linearity of the processing framework makes it
ential properties of the smooth surface it approximates. On very efficient, and it has become an attractive and promising
the other hand, many modeling tasks are difficult to perform research direction, as is evident from the amount of recent
on meshes. Moreover, nowadays the common scanned sur- publications on the subject.
faces usually come as very detailed, complex and at times
noisy meshes that require geometry processing, such as fil- In the following section we will closely look into the
tering, resampling and compression for efficient storage and Laplacian operator, the differential representations and the
streaming. associated surface reconstruction problems. We will then re-
view some results and applications that benefit from this
In this paper, we describe recent work on mesh process- framework, such as shape approximation, geometry com-
ing and modeling that is based on the Laplacian framework pression and watermarking (Section 3) and interactive mesh
and differential representations. In this framework, surface editing and interpolation(Section 4).


c 2006 The Author
Journal compilation 
c 2006 The Eurographics Association and
Blackwell Publishing Ltd. Published by Blackwell Publishing, Submitted September 2005
9600 Garsington Road, Oxford 2DQ, UK and 350 Main Street, Revised April 2006
Malden, MA 02148, USA. Accepted May 2006
789
790 O. Sorkine / Differential Representations for Mesh Processing

2. Laplacian Operator and Differential Surface

Representation

In the following, we review the definition of differential coor-

dinates (δ-coordinates), the associated mesh Laplacian oper-
ator, describe the surface reconstruction framework and dis-
cuss various properties of the Laplacian operator.

Figure 1: The vector of the differential coordinates at a ver-

2.1. Basic definitions tex approximates the local shape characteristics of the sur-
face: the normal direction and the mean curvature.
Let M = (V , E, F) be a given triangular mesh with n ver-
tices. V denotes the set of vertices, E denotes the set of edges
and F denotes the set of faces. Each vertex i ∈ M is con-
ventionally represented using absolute Cartesian coordinates, From a differential geometry perspective, the δ-
denoted by v i = (xi , yi , z i ). We first define the differential or coordinates can be viewed as a discretization of the con-
δ-coordinates of v i to be the difference between the absolute tinuous Laplace–Beltrami operator [3], if we assume that our
coordinates of v i and the center of mass of its immediate mesh M is a piecewise-linear approximation of a smooth
neighbors in the mesh, surface. We can write the differential coordinate vector at
  vertex v i as
(y) 1 
δi = δi(x) , δi , δi(z) = vi − vj
di j∈N (i) 1 
δi = (vi − v j ).
di j∈N (i)
where N (i) = { j |(i, j) ∈ E} and di = |N (i)| is the number
of immediate neighbors of i (the degree or valence of i). The sum above is a discretization of the following curvilinear
integral: |γ1 | v∈γ (vi − v) dl(v), where γ is a closed simple
The transformation of the vector of absolute Cartesian co- surface curve around v i and |γ | is the length of γ . It is known
ordinates to the vector of δ-coordinates can be represented in from differential geometry that
a matrix form. Let A be the adjacency (connectivity) matrix
of the mesh 1
 lim (vi − v) dl(v) = −H (vi ) ni
1 (i, j) ∈ E
|γ |→0 |γ | v∈γ
Ai j =
0 otherwise where H (v i ) is the mean curvature at v i and n i is the surface
normal. Therefore, the direction of the differential coordinate
and let D be the diagonal matrix such that Dii = di . Then vector approximates the local normal direction and the mag-
the matrix transforming the absolute coordinates to relative nitude approximates a quantity proportional to the local mean
coordinates is curvature [4] (the normal scaled by the mean curvature is
termed mean-curvature normal). Intuitively, this means that
L = I − D −1 A. the δ-coordinates encapsulate the local surface shape (see
Figure 1).
It is often more convenient to consider the symmetric version
of the L matrix, defined by L s = DL = D − A, It should be noted that geometric discretizations of the
⎧ Laplacian have better approximation qualities. Meyer et al.
⎪ d i= j
⎨ i [5] propose employ the cotangent weights, first proposed by
(L s )i j = −1 (i, j) ∈ E Pinkall and Polthier [6], instead of uniform weights:
⎪
⎩
0 otherwise 1  1
δic = (cot αi j + cot βi j )(vi − v j )
That is, L s x = Dδ (x) , L s y = Dδ (y) and L s z = Dδ (z) , where |i | j∈N (i) 2
x is an n-vector containing the x absolute coordinates of all
the vertices and so on. See Figure 3 for an example of a mesh where | i | is the size of the Voronoi cell of i and α i j , β i j
and its associated matrices. denote the two angles opposite of edge (i, j) (see Figure 2),
to approximate mean-curvature normals. These geometry-
The matrix L s (or L) is called the topological (or graph) dependent weights lead to vectors δ ic with normal compo-
Laplacian of the mesh [1]. Graph Laplacians have been ex- nents only, unlike the previously defined δ i which also have
tensively studied in algebra and graph theory [2], primarily tangential components and may be nonzero on planar 1-rings.
because their algebraic properties are related to the combi- The cotangent weights may be negative and problematic to
natorial properties of the graphs they represent. We will look define for very large angles due to the properties of cotangent
into some of these properties later on, in Sections 2.2 and 3. near π; convex weights that mimic the cotangent weights are

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 791

It follows that

rank(L) = n − k

where k is the number of connected components of M,

because each connected component has one (translational)
degree of freedom.

In order to uniquely restore the global coordinates, we need

to solve a full-rank linear system. Assuming M is connected,
we need to specify the Cartesian coordinates of one vertex
to resolve the translational degree of freedom. Substituting
Figure 2: The angles used in the cotangent weights and the the coordinates of vertex i is equivalent to dropping the ith
mean-value coordinates formulae for edge (i, j). row and column from L, which makes the matrix invertible.
However, as we will see shortly, usually we place more than
one constraint on spatial locations of Ms vertices. Denote by
called mean-value coordinates [7] C the set of indices of those vertices whose spatial location
is known. We have therefore |C| additional constraints of the
tan(θi1j /2) + tan(θi2j /2) form
wi j = ,
vi − v j 
v j = c j , j ∈ C. (1)
where the θ angles are depicted in Figure 2. These geometric
If we assume w.l.o.g. that the vertices are indexed such that
discretizations are beneficial for applications where the mesh
C = {1, 2, . . . , m} then our linear system now looks like this
geometry is always known, such as mesh filtering and edit-
  
ing, since they approximate the differential properties more L δ (x)
accurately. However, in compression applications we will de- x= . (2)
ω Im×m | 0 ω c1:m
fine effective bases for geometry representation based on the
topological Laplacian, so that the weights in the operator def- The same goes for y and z coordinate vectors, of course. We
inition, and thus the bases themselves, are implicitly encoded denote the system matrix in (2) by L̃. Note that we treat the
in the mesh connectivity and no extra information is required positional constraints (1) in the least-squares sense: instead
to store them. of substituting them into the linear system, we add the above
equations as additional rows of the linear system. This way,
for each j ∈ C we are also able to retain the “smoothness”
2.2. Surface reconstruction from δ-coordinates term, that is, the equation Lv j = δ j . Note that in our discus-
In the following, we will be working with the above defined sion we will keep addressing (1) as “positional constraints”
differential surface representation. We will perform different or “modeling constraints,” although they may not be fully
operations on the δ-coordinates, depending on the task at satisfied by the least-squares solution. The weight ω > 0 can
hand. The final step of any such manipulation must be surface be used to tweak the importance of the positional constraints
reconstruction, or, in other words, we need to recover the (generally, every constraint may have its own weight, and we
Cartesian coordinates of M’s vertices. can furthermore apply varying weights on the rows of the ba-
sic Laplacian matrix as well). The additional constraints make
Given a set of δ-coordinates, can we uniquely restore the our linear system over-determined (more equations than un-
global coordinates? The immediate answer is no, because the knowns) and in general no exact solution may exist. However,
matrix L (or L s ) that we use to transform from global to dif- the system is full-rank and thus has a unique solution in the
ferential coordinates is singular, and therefore the expression least-squares sense
x = L −1 δ (x) is undefined. The sum of every row of L is zero,  
 2 
which implies that L has a nontrivial zero eigenvector (1,  (x) 
x̃ = argmin Lx − δ  + ω |x j − c j | . (3)
2 2
1, . . . , 1)T . The matrix is singular since the δ-coordinates are x j∈C
translation invariant: if we use weights w i j that sum up to 1
to define the Laplacian operator and we translate the mesh The analytical least-squares solution is expressed in terms
by vector u to obtain the new vertices vi then of the rectangular (n + m) × n system matrix L̃ from
 Equation (2):
L(vi ) = wi j (vi − vj )
j∈N (i) x̃ = ( L̃ T L̃)−1 L̃ T b

= wi j ((vi + u) − (v j + u)) where b = (δ, ω c 1 , . . . , ω cm )T is the right-hand side vector
j∈N (i) of the linear system. See Figure 3 for an example of a small
= j∈N (i) wi j (vi − v j ) = L(vi ). mesh and its associated matrix L̃.

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
792 O. Sorkine / Differential Representations for Mesh Processing

Table 1: Time measurements of manipulations of the (extended)

Laplacian matrices for several typical meshes, on a 3.4 GHz Pentium
4 computer. Factorization stands for Cholesky factorization of L̃ T L̃
and Solve denotes solving by back-substitution for one mesh function.

No. of Factor Solve

Model vertices (seconds) (seconds)

Eight 2,718 0.085 0.004

Horse 19,851 0.900 0.032
Camel 39,074 2.096 0.073
Feline 49,864 2.750 0.110
Max Planck 100,086 7.713 0.240

The factorization takes the bulk of the computation time, and

the back-substitution step is usually very fast. Using advanced
linear solvers, such as TAUCS library [9], makes the computa-
tions very efficient. See Table 1 for timing data on several typ-
ical meshes. Another important advantage of direct solvers is
the separation between the matrix processing (factorization)
and the right-hand side, which is only involved in the fast
back-substitution step. This permits frequently changing the
right-hand sides at the price of the factorization pre-process,
which is done only once per fixed matrix. This property is
extremely useful for interactive mesh editing, as described
Figure 3: A small example of a triangular mesh and its as- in Section 4. When the system matrix is very large (million
sociated L s matrix (top right). Second row: a 2-anchor in- variables and more), the size of the factor becomes too large
vertible Laplacian and a 2-anchor L̃ matrix. The anchors are to fit into memory. In this case, it is possible to either use the
denoted in the mesh by red. out-of-core version of Cholesky factorization (available, e.g.,
with [9]) or iterative solvers whose complexity scales linearly
with the size of the system (for instance, see the recent work
on multigrid solvers [10,11]).
An important practical aspect of surface reconstruction
from δ-coordinates is the availability of efficient and accurate Note that the accuracy of the solution is highly depen-
numerical methods for solving the least-squares problem (3). dent on the conditioning of the linear system. The Laplacian
A recent comparative study [8] shows that direct methods matrix is notoriously ill-conditioned: when we add only one
prove to be superior for these types of problems on mod- anchor, the largest singular value of L̃ is proportional to the
erately sized meshes (up to a few hundreds of thousands of maximal degree in the mesh, while a good estimate for the
vertices). The system (3) can be solved by applying Cholesky smallest singular value is one over the product of the maxi-
factorization to the associated normal equations mum topological distance of a vertex from the single anchor
vertex and the number of vertices in the mesh [12,13]. For
( L̃ T L̃) x = L̃ T δ. a typical n-vertex 3D mesh, the condition number is there-
fore likely to be (n −1.5 ), and it is squared when we consider
The matrix L̃ is sparse, and M = L̃ T L̃ is also sparse (al- the normal equations. However, adding anchors to the sys-
though not as sparse as L̃) and positive definite. By using tem improves its conditioning and makes the factorization
fill-reducing reordering, it is possible to compute a sparse of the normal equations stable. A rigorous way to bound the
Cholesky factorization of M smallest singular value of L̃ from below is shown in [14]; the
bound depends on the number of anchors and their spacing
M = RT R
across the mesh.
where R is an upper-triangular sparse matrix. The factoriza- In summary, the transition from the global Cartesian rep-
tion is computed once, and we can solve for several mesh resentation to differential representation is performed by a
functions (x, y and z) by back substitution linear operator with local support: the mesh Laplacian. And
vice versa: in order to recover the Cartesian coordinates
R T ξ = L̃ T δ (x) from the differential representation, one needs to solve a
R x = ξ. sparse linear least-squares problem. This provides a powerful

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 793

framework for surface manipulation: we will apply different are well studied in both continuous and discrete settings, and
modifications to the differential coordinates (such as quan- the sampling theorem tells us the number of basis functions
tization for compression purposes) and/or pose additional we need to use for perfect signal reconstruction. Things are
modeling constraints, and then reconstruct the surface by more involved for the irregular setting of arbitrary triangu-
solving the least-squares problem. The advantages of this lar meshes. One possible approach of direct application of
framework can be summarized as follows: the signal processing theory on meshes is to first perform
semi-regular remeshing by parameterizing the mesh over a
r It strives to preserve local surface detail as much as pos- simple base complex, and then apply variations of 2D signal
sible under the constraints. processing methods [18,19,20,21,22]. Another possibility is,
r The least-squares solution smoothly distributes the error
instead of altering the mesh, to work directly on the irreg-
ular mesh and develop generalizations of the methods from
across the entire domain, providing a graceful reconstruc-
regular settings.
tion.
r Sparse linear systems can be solved very efficiently. Karni and Gotsman [15] introduce a spectral method where
the mesh is approximated by reconstructing its geometry us-
ing a linear combination of a number of basis vectors. The
2.3. Spectral properties basis is the spectral basis E of the L s matrix, which can be
regarded as a generalization of the discrete Fourier basis (see
Let us consider the L s matrix, since it is symmetric and thus
Section 2.3). Similarly to JPEG compression of images, the
simpler to analyze. The matrix L s is symmetric positive semi-
mesh geometry functions x, y, z are decomposed in the basis
definite and thus has an orthonormal eigenbasis
E:
E = {e1 , e2 , . . . , en }.
x = α1 e1 + α2 e2 + · · · + αn en (4)
Denote the eigenvalues by λ i ,
and the high-frequency components (coefficients of the last
0 = λ1 < λ2 ≤ λ3 ≤ · · · ≤ λn . eigenvectors) are truncated. The coefficients α i are quantized
and subsequently entropy coded. It is assumed that the mesh
It is known that the first nonzero λ i are very small, and the
connectivity is known to the decoder prior to receiving the
largest eigenvalue λ n is bounded by the largest vertex va-
geometry coefficients, so that the decoder first computes the
lence in M times 2. The first eigenvectors (corresponding
spectral basis and then reconstructs the geometry by combin-
to small eigenvalues) are smooth, slowly varying functions
ing the coefficients. Note that this method is readily made for
on the mesh, and the last eigenvectors have high frequency
progressive streaming, since if one sends the α i coefficients in
(rapid oscillations). For example, the first eigenvector e 1 is
ascending order of i, the decoder can first reconstruct a coarse
the constant vector, that is, the “smoothest” mesh function
(very smooth) approximation of the geometry, and then grad-
that does not vary at all. In fact, the Laplacian eigenbasis is
ually add high-frequency detail, as more coefficients become
an extension of the discrete Fourier basis to irregular domains
available.
[4]. The eigenvalues are considered as mesh-frequencies. We
will see in the following how this fact is exploited for sig- An example of the progressive encoding of [15] is shown
nal processing on meshes, compact geometry representation, in Figure 4. The geometry of the Horse model is first approx-
watermarking and more. See also [16] for an earlier state-of- imated using only a few of the first eigenvectors e i , and the
the-art report on this subject. resulting approximation is very smooth. As more and more
eigenvectors participate in the approximation, the obtained
surface receives more high-frequency detail.
3. Efficient Shape Representation
The spectral method of [15] provides high compression
In this section we review several methods for efficient and ratios, since most of the energy is concentrated in the low-
compact geometry representation that benefit from the mesh frequency coefficients (the first α i ), and the high-frequency
Laplacian operator and the framework presented in Section 2. components add negligible weight. However, computing
For an extensive recent survey on general compression meth- even the first few eigenvectors of typical meshes (with tens
ods, we refer the reader to [17]. of thousands of vertices) is an extremely computationally
expensive task of superlinear complexity, that has to be car-
3.1. Efficient bases for geometry representation ried out both on the encoder and the decoder side. Karni
and Gotsman [15] propose to divide the mesh into patches
Finding a good basis for compact shape representation means and compute the spectral decomposition separately for each
that one can use only a fraction of the basis functions to ap- patch. This speeds up the computation, but the quality of
proximate a given geometry well. This idea is extensively the low bit-rate reconstructions suffers from artifacts along
used in signal processing and in the image domain, by defin- the boundaries between patches. Later, Karni and Gotsman
ing Fourier or wavelet bases. The regular 1D and 2D settings developed another technique, where fixed spectral bases are

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
794 O. Sorkine / Differential Representations for Mesh Processing

Figure 4: Progressive approximations of the Horse geometry using the eigenvectors of the topological Laplacian [15]. The left
figure shows an approximation using only a small number of the first eigenvectors (those that correspond to small eigenvalues).
When more eigenvectors are added to the approximation, the surface attains more high-frequency detail (middle and right).
Data courtesy of Zachi Karni.

employed [23]. The spectral basis of a regular mesh with

n vertices is used to approximate the Laplacian eigenbasis,
and the computational cost of spectral analysis is avoided on
the decoder side altogether. Of course, this technique works
better for meshes that are close to being regular.
In [24,25], a different class of shape approximation tech-
niques for irregular triangular meshes was introduced. The
method approximates the geometry of the mesh using a lin-
ear combination of a small number of basis vectors that are
functions of the mesh connectivity and of the mesh indices
of a number of anchor vertices. The initial motivation was
to improve the shortcomings of the spectral basis [15]. One
must bear in mind that, together with their appealing proper-
ties, the spectral bases are geometry-oblivious, since the basis
vectors are functions of the connectivity alone. In contrast, the
new geometry-aware method derives the basis both from the
mesh connectivity and geometry. The basis vectors in [25] are
centered around selected “geometrically important” anchor
vertices. This allows a terse capturing of important features of
the surface and leads to compact and efficient representation
of the mesh geometry. Figure 5 illustrates the reconstruction
using such basis vectors on a 2D curve example. The bottom
row shows the meshes reconstructed using geometry-aware
bases. The locations of the anchors are marked by small dots.
Figure 5: Reconstruction of the swallow curve (simple
The reconstructed mesh passes close to the original locations
closed path) using different bases. The top row shows re-
of the anchor points, which enables good approximation of
construction using the Laplacian eigenvectors, which are the
such features as the tips of the bird’s wings and tail. For
discrete Fourier basis function in this case. The bottom row
comparison, reconstruction of this mesh using an analogous
displays reconstruction with geometry-aware basis vectors.
number of spectral basis vectors misses out the features. This
The reconstructed mesh is shown in black, while the original
behavior is evident in large as well as in small scale.
mesh is tinted in blue. The geometry-aware bases better ap-
It should be noted that explicit computation of the basis proximate the features of the shape, on a large as well as on
vectors is, generally speaking, too expensive for large meshes. a small scale.
The method in [25] avoids explicit computation of the under-
lying basis. Instead of directly representing the geometry by squares sense: it attains the value 1 at one of the anchors and
the coefficients of the linear combination of the basis vectors, 0 at the other anchors, and it is the smoothest among all the
the reconstruction problem is reduced to solving a sparse lin- functions that satisfy these requirements (a slightly different
ear least-squares system of the type (2). Each basis function is definition for strictly interpolatory anchors is also proposed,
selected so that it fulfills the following conditions in the least- but the principle is the same). The smoothness of a function is

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 795
1.2
5 first vectors out of 5 anchor vectors
1.2
5 first vectors out of 20 anchor vectors
of Cartesian coordinates is possible without causing visible
1 1
artifacts (10–16 bits per coordinate).
0.8
0.8

0.6
0.6 In [26], a different approach to geometry quantization is
0.4

0.2
0.4
proposed. Instead of directly quantizing the Cartesian coor-
0
0.2
dinates, the quantization is applied to the δ-coordinates, and
−0.2
0
the geometry of the mesh can be restored on the decoder
−0.4 −0.2

−0.6 −0.4
side by solving a linear least-squares system defined by the
extended Laplacian matrix (discussed in Section 2.2). Intro-
0 50 100 150 200 250 300 0 50 100 150 200 250 300

(a) (b) ducing high-frequency errors by quantizing the δ-coordinates

results in low-frequency errors in the reconstructed Cartesian
Figure 6: Geometry-aware basis functions on a 1D domain. coordinates. By considering a visual error metric between
The mesh here is a simple closed path with 274 vertices. Plot meshes, which takes into account not only the Euclidean dis-
(a) displays the five basis functions corresponding to a set of tance between corresponding vertices (or the “Metro” dis-
five anchors; (b) shows the first five basis functions out of a tance [27]) but also the smoothness error, it can be argued
20-anchor basis. that low-frequency displacements in the surface geometry
are less noticeable than high-frequency displacements that
modify the local characteristics of the surface such as nor-
defined in a discrete manner using the Laplacian of the mesh. mals and curvature. Consequently, strong quantization of the
Specifically, the basis functions are defined as follows: Given δ-coordinates yields a small visual error, in contrast to stan-
a set of k anchor vertex indices 1 ≤ a 1 , a 2 , . . . , ak ≤ n, the dard Cartesian coordinate quantization.
ith function f i in the basis minimizes
The strategy of [26] is called high-pass quantization, to
 
 2   2 emphasize the fact that it tends to concentrate the quantization
2  2 
L fi  + ω ( f i )ai − 1 +
2
ω ( f i )a j − 0 error at the low-frequency end of the spectrum, in contrast
j=i to the spectral method of [15]. Mesh frequencies are defined
as in Section 2.3. Let us look what happens to quantization
where ω is a positive weight. See Figure 6 for a 2D example
errors. The δ-coordinates are obtained by applying the mesh
of these basis functions (note that they were computed here
Laplacian, δ = L s x. When the δ-coordinates are quantized, an
for illustration purposes only; the algorithm in [25] does not
error ε (with small norm) is added. To reconstruct the mesh
require the bases explicitly). The definition results in smooth
function x from δ + ε, we write
basis functions that are easy to combine into an approxima-
tion that attains specific values at the anchors. The compact x = L −1 −1 −1 −1
s (δ + ε) = L s δ + L s ε = x + L s ε
geometry representation thus consists of the indices of the
anchor vertices and the coefficients associated with the basis (We assume a nonsingular L s by adding one anchor.) Thus,
functions. To recover the surface geometry, we simply need the quantization error is q ε = L −1
s ε, and its norm is no longer
to solve the system (3), where instead of δ we put zero small. In fact, if we look at q ε in the spectral basis, we will
  see that its low-frequency component is amplified
 k  k  2
2 
x̃ = ci fi = argmin Lx + 2
ω xai − ci  . ε = c1 e1 + c2 e2 + · · · + cn en
i=1 x i=1
1 1 1
Taking a sufficiently large value for ω (in the order of n) makes qε = L −1
s ε = c1 e1 + c2 e2 + · · · + cn en .
λ1 λ2 λn
the reconstruction effectively interpolatory at the anchors.
Figure 7 demonstrates an example of shape approximation The last statement holds since a matrix and its inverse share
obtained with these basis functions. the same set of eigenvectors, and the corresponding eigen-
values are inversed. The first λ i are very small, thus the first
1
λi
are large. This means that q ε has very strong components
3.2. High-pass quantization in the direction of the first e i , that is, low frequencies. The
high-frequency components are damped, because the last λ i
One of the main ways to compress floating-point data, such as are relatively large, so for large i, λ1i < 1. Thus we can con-
mesh geometry, is by quantization. Quantization necessarily clude that when quantizing δ-coordinates, we get mostly low-
introduces errors and causes a certain loss of data. Loosely frequency errors. Note that the eigendecomposition is used
speaking, quantizing the Cartesian coordinates of the mesh solely for analysis purposes; the quantization method only
produces high-frequency errors across the surface. This es- requires linearly computing the δ-coordinates and linear sur-
pecially damages the fine-sampled areas, since the relative face reconstruction on the decoder side.
error is greater when the polygons are smaller. Aggressive
quantization significantly alters the surface normals, causing To reduce the low-frequency error, more anchor ver-
the irritating “jaggies” effect. Thus, only mild quantization tices, whose spatial location is known, can be added to the

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
796 O. Sorkine / Differential Representations for Mesh Processing

Figure 7: Reconstruction of the Feline model using an increasing number of geometry-aware basis vectors. The sizes of the
encoded geometry files are displayed below the models. The letter e denotes the L 2 error value, given in units of 10−4 .

3.3. Mesh watermarking

The visual properties of low-frequency errors in geometry

prove to be useful not only for geometry compression. The
insensitivity of the human visual system to low-frequency
errors (such as smooth shifts or rotations of parts of the
surface) is exploited in mesh watermarking techniques
[28,29,30]. Mesh watermarking, similar to image water-
marking, is essential for copyright protection and ownership
authentication of geometric data. The goal is to embed an
invisible bit string into the geometric representation of the
object, so that the actual owner of the authentic digital object
would be able to robustly compare the marked copy with the
real (unmodified) version of the 3D object and to verify the
presence of the correct watermark.
The idea of Ohbuchi et al. [29,30] was to embed a wa-
termark bit string into the spectral representation of the 3D
mesh at hand, by slightly modifying the low-frequency co-
Figure 8: The δ-coordinates quantization to 5 bits/ efficients of the spectral decomposition (4). As previously
coordinate (left) introduces low-frequency errors, whereas discussed, small changes in the low-frequency component of
Cartesian quantization to 11 bits/coordinate (right) intro- the surface geometry, or low-frequency errors, do not alter
duces noticeable errors. The upper row shows the quantized the local surface appearance, and thus they are less apparent
model, and the bottom figures use color to visualize the cor- to the human eye. Therefore, in applications where the 3D
responding quantization errors. It is recommended to zoom meshes are used primarily for display, such an approach to
in into the electronic version to see the quantization artifacts. watermarking is definitely plausible. In brief, the watermark-
ing algorithm works as follows: the input mesh is partitioned
into balanced patches to reduce the cost of spectral decom-
representation. These vertices “nail” the geometry in place position, and then the eigendecomposition of the symmetric
and prevent large shifts or rotations. Algebraically, anchors topological Laplacian matrix is computed for each patch (see
improve the conditioning of the Laplacian matrix by making Equation (4)). Next, the low-frequency coefficients (the first
the smallest singular value larger. Adding anchors requires α i ) are modulated by the selected watermark bit string, and
more storage space, and therefore their placement needs to the mesh geometry is reassembled by summing up the right-
be optimized so as to minimize the error. A greedy placement hand side of (4) with the modified coefficients. This com-
technique is given in [26], where the anchors are chosen one- pletes the watermarking process. To extract the watermark
by-one, each new anchor is placed at the vertex that achieved from a “suspected” mesh, it is first aligned with the original
highest reconstruction error in the previous iteration. A dif- mesh and remeshed so that it has the same connectivity as
ferent placement strategy that conforms with provable upper the original mesh. Then, spectral analysis is performed, and
bounds on the geometry error is introduced in [14]. the spectral coefficients of the suspected mesh are subtracted
from the original (unmodified) coefficients, so that the wa-
Figures 8 and 9 show some quantization results from [26]. termark bit string can be retrieved.

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 797

Figure 9: Visualization of the visual error across the mesh surface. The surface was reconstructed from δ-coordinates quantized
to 7 bits/coordinate using 2 anchors (a), 4 anchors (b) and 20 anchors (c). The anchor points are shown as small red spheres.
Each vertex v is colored according to its visual error value Evis (v). We have also added a sign to these values, so that vertices
that move outside of the surface have positive error values (colored by red), and vertices that move inwards have negative error
values (colored by blue). In (d), the visual error of direct quantization of the Cartesian coordinates is shown.

Recently, the above watermarking scheme was extended From the algorithmic standpoint, the handle vertices func-
to point-set data where explicit connectivity is not present tion as boundary conditions for the calculation of the edited
(see [31]). surface. These conditions change every time the user manip-
ulates the handle. An additional set of boundary constraints
expresses the static parts of the surface that remain unchanged
4. Mesh Editing and Shape Interpolation during the editing process. Usually, this set of constraints con-
sists of some padding belt of vertices around the ROI bound-
Manipulating and modifying a surface while preserving the
ary, and it ensures a gradual transition between the ROI and
geometric details is important for various surface editing op-
the rest of the mesh (which, as mentioned above, is com-
erations, including free-form deformations [32,33], cut and
pletely static and is usually not included the editing process
paste [34,35,36], fusion [37], morphing [38] and others. Note
at all for efficiency reasons). Thus, if we denote the group
that the absolute position of the vertices in a mesh is not
of constrained vertices (those that belong to the handle and
important for these operations; what is needed is a local de-
the static ones) by U, the boundary constraints are written
scription of the shape that is not dependent on the particular
simply as
placement of the shape in the Euclidean space. We call such
shape representation “shape-intrinsic”: it is invariant under vi = ui , ∀i ∈ U (5)
rigid transformations of the space in which the shape is em-
bedded, and it allows to reconstruct the shape uniquely, up where u i are the desired positions for the constrained vertices.
to a rigid transformation. In the following we will describe
several types of local surface representation that employ the Any editing algorithm in fact changes the shape of the
Laplacian framework and other differential representations, ROI so as to satisfy the modeling constraints (5) while striv-
and show how they are used for mesh editing. In all these rep- ing to give the modified surface a natural shape. Here, we
resentation methods the main question is how to make them will concentrate on one particular definition of natural defor-
behave in a shape-intrinsic manner. mation, namely: a deformation that preserves the local details
of the original surface. In the next sections we will see how
the above modeling metaphor is implemented using various
4.1. The modeling metaphor surface representations.

Let us first briefly describe the traditional way of direct sur-

face manipulation. The modeling metaphor is very simple 4.2. Multiresolution mesh editing
from the user’s point of view [39]. First, the user marks the
desired region of interest (ROI) on the mesh that is to be Partially shape-intrinsic surface mesh representa-
modified. The rest of the mesh will remain unchanged (and tions are given by multiresolution decompositions
thus does not participate in the calculations). Next, the user [19,21,39,40,41,42,43]. In a multi-resolution mesh, the
selects a group of vertices that will serve as the manipulation geometry is encoded as a base mesh and several levels
handle. The user can then drag the handle around and apply of refinement. The base mesh is the low-frequency com-
arbitrary transformations to it (such as rotation, scaling, etc.), ponent of the shape, and it is typically represented in
and the surface of the ROI is modified on the fly, following Cartesian coordinates. Thus it is not shape-intrinsic, but the
the handle manipulation. shape details are represented in an intrinsic manner. The

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
798 O. Sorkine / Differential Representations for Mesh Processing

refinements are described locally, so that the geometric

details are mostly captured in a discrete set of translation-
and rotation-invariant coordinates. Using this representation,
modeling operations can be performed on an appropriate
user-specified level-of-detail. The multiresolution hierarchy
can be constructed for both the connectivity and the
geometry (as in, e.g., [19,21]), or the only geometry (as
in, e.g., [39,42,43]), where the connectivity of the mesh
remains full and unchanged in all the levels-of-detail in the
hierarchy. Different levels of refinement are usually obtained
by gradual smoothing of the mesh.

We will outline the editing mechanism of the multireso-

lution framework of Kobbelt and colleagues [39,41,43] as
an exemplary multiresolution technique that also uses the
Laplacian framework. For simplicity, let us assume that the
mesh hierarchy consists of two levels only, sharing the same
connectivity: the smooth base mesh and the detailed mesh.
The detailed level is encoded with respect to the base level
by computing the offset of each vertex and representing this
offset in the local frame of the corresponding vertex in the
base mesh. The choice of the local frames is rotation- and
translation-invariant, and therefore the representation of the
offsets (which are, in fact, the surface details in this case)
is independent of the location of the mesh in space. Mesh
editing is performed as follows: first, the smooth base mesh Figure 10: The importance of rotation-invariant represen-
is deformed according to the constraints (5); then the local tation. (a) the original shape, where the base shape is drawn
frames are recomputed on the deformed base mesh, and fi- in brown. In (b), the base shape was bent, while the details
nally, the original detail offsets (in their local frame represen- were encoded in a rotation-invariant manner, and thus their
tation) are added to the base mesh. This way, the local details orientation with respect to the base was preserved. In (c), the
are reproduced correctly in the deformed mesh: for example, same deformation leads to distorted orientation of the details
if the surface was bent, the local frames of the smooth base (stretch) because the encoding was not rotation-invariant.
mesh rotate, and thus so do the offsets. This key concept of
invariance in local representation is illustrated in Figure 10.
hard constraints. Note that the factorization of the system
How is the base mesh deformed? It is assumed that the base matrix can be computed once per choice of ROI and handle
mesh is smooth enough and does not contain local details that vertices; each time the user transforms the handle, only the
need to be preserved. Therefore, the base mesh is deformed right-hand side of the system changes and thus only a back-
by solving an optimization problem with two objectives: substitution step is required. This allows interactive response
(i) satisfying the user constraints (5) and (ii) keeping the of the editing mechanism for fairly large ROIs (up to several
surface of the base mesh smooth. The latter goal is achieved tens of thousands of vertices, refer to Table 1).
by describing smoothness in terms of the mesh Laplacian op-
The multiresolution approach is probably the most intuitive
erator, as we already saw in Section 3.1. Botsch and Kobbelt
method for mesh editing, and its advantage is the straightfor-
[43] propose different orders of smoothness, obtained via dif-
ward representation of the details of the shape in a man-
ferent orders of the Laplacian operator. The optimization is
ner that is invariant to the global coordinate system. On the
formulated as
other hand, this approach defines details explicitly, or in other
words, one may need to explicitly set the required amount of
Lkx = 0
smoothing to create a satisfactory base mesh. In meshes with
complex details, many levels of multi-resolution hierarchy
leading to C k−1 smoothness of the resulting surface. Proba-
may be required in order to handle the details correctly.
bly the most common value is k = 2, producing the so-called
“thin-plate” surface. The Equations (5) are incorporated as
hard constraints into the above system by substitution, secur- 4.3. Single-resolution local representations
ing its nontrivial solution. When k = 2, this optimization is for mesh editing
equivalent to solving the least-squares problem (3) where all
the δ-coordinates are set to zero and sufficiently large weight Local details can be defined in a more implicit way with
is put on the positional constraints, to effectively make them a single-resolution mesh, by using differential coordinates.

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 799

Mesh editing using this representation was pioneered by cal rotations based on translational, or positional constraints
Alexa [44]. The basic idea is simply to add the positional alone; the user is not expected to provide rotational trans-
constraints (5) to the system that reconstructs the Cartesian formations of the handle. So even in the simplest case of
coordinates of the mesh from the differential coordinates (3). grabbing one handle vertex and “pulling” it out of the sur-
Once again, note that the system matrix needs to be con- face, the technique will produce rotation estimates for the
structed and factored only once per choice of ROI and handle; surface details and they will be transformed accordingly.
during user manipulation one only needs to solve the linear
system by back-substitution, which is very efficient. Yu et al. [47] propagate the rotation (or any other linear
transformation) of the editing handle, defined by the user, to
The above idea for surface representation and reconstruc- all the gradient vectors of the ROI. This is the major differ-
tion was further elaborated in [45,46,47]. Yu et al. [47] in- ence from [45], because here the user is expected to specify
troduce a technique called Poisson Editing, formulated by compatible translation and rotation constraints on the han-
manipulation of the gradients of the coordinate functions dle; translational constraints alone will result in no local
(x, y, z) of the mesh. The surface is reconstructed by solv- rotations (and thus the surface details may be significantly
ing the least-squares system resulting from discretizing the distorted). The propagated transformations are interpolated
Poisson equation with the identity, where the interpolation weight for each tri-
angle is proportional to the geodesic distance of the triangle
∇2 f =  f = ∇ · w (6)
from the handle. As pointed out in [48], this solution may also
with Dirichlet boundary conditions. Here, f is the unknown produce unexpected results for meshes with complex features
scalar function on the mesh, that is, one of the coordinate protruding from the smooth base surface. This is because the
functions x, y, z of the deformed mesh. The vector field w tip of a protruding feature is geodesically further from the
is the gradient field of the corresponding original coordi- handle than the vertices around the feature’s base, and thus
nate function. The set of boundary constraints consists of the tip receives a smaller share of the deformation than the
both positional constraints of the form (5) and constraints base (whereas for a natural result, it should be roughly the
on the gradients of the handle vertices, directly expressed by same amount of deformation).
the desired linear transformation on the handle. The spatial
Laplacian operator , restricted to the mesh, is discretized A different interpolation scheme for deformation propa-
using the cotangent weights. By inspecting Equation (6), we gation was recently proposed by Zayer et al. [49]: it avoids
can see that it has a similar form to the Laplacian surface re- geodesics computation, creating a harmonic scalar field on
construction system (3); the difference lies on the right-hand the mesh instead. The harmonic field attains the value of 1
side, that is expressed in terms of a vector field rather than a at the handle vertices and 0 at the boundary of the ROI, and
scalar field, and the constraints are substituted into the system varies monotonically on the rest of the ROI, thus providing a
(so that this is not a least-squares optimization). valid and smoother alternative to the geodesic weights. The
harmonic field is obtained by solving (2), only the 1 and 0
In [45], the surface is reconstructed from the Laplacian constraints are substituted into the system rather than added in
δ-coordinates of the mesh and spatial boundary conditions least-squares sense. An additional advantage of this approach
by solving (3). Both Lipman et al. [45] and Yu et al. [47] is that it uses the same system matrix both for the computation
point out the main problem of the above approach: the need of the weights and for the actual editing (only with different
to rotate the local frames that define the Laplacians, or the gra- right-hand sides), thus the factorization of the matrix serves
dients, to correctly handle the orientation of the local details. both tasks at once. As a side note, we mention that the mesh
If this issue is not properly addressed, the system in (3) (or in Laplacian operator is tightly coupled with Hodge theory and
(6)), tries to preserve the Laplacian vectors (or the gradients) computation of globally conformal (harmonic) mesh param-
in their original orientation with respect to the global coor- eterizations for arbitrary topology. This topic is outside of
dinate system. Therefore a result in the spirit of the sketch in the scope of this paper; the reader is kindly referred to recent
Figure 10c would be obtained. publications [6,50,51,52] for further information.

Yu et al. and Lipman et al. propose remedy to this prob- It is important to note that all the above-mentioned methods
lem by explicit assignment of the local rotations. Lipman [45,47,49] are fast: the size of the system matrices involved in
et al. [45] estimate the local rotations of the frames on the the reconstruction is n × n, where n is the number of vertices
underlying smooth surface by smoothing the “naive” solu- in the ROI. This size will grow in the follow-up works, as we
tion of (3); then the original Laplacians are rotated by these will see shortly.
estimated rotations and the reconstruction system is solved
again. This heuristic achieves good results on meshes with Subsequent research continued to look for truly shape-
relatively small details; applying it to complex meshes where intrinsic representation suitable for mesh editing. Sheffer and
the details are far from being height-fields above a smooth Kraevoy [53] introduced the so-called pyramid coordinates:
base surface would require a lot of smoothing iterations. An at each vertex i of the mesh, the normal and the tangential
important feature of this approach is that it deduces the lo- components of the local frame are stored, independently of

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
800 O. Sorkine / Differential Representations for Mesh Processing

the global coordinate frame. To construct this representa-

tion, a local projection plane is fitted to the vertices of the
1-ring by averaging the face normals of the 1-ring. The nor-
mal component of the center vertex i is its distance to the
projection plane, whereas the tangential component is en-
coded as the edge lengths and the angles between the edges
of the 1-ring, projected onto the plane. Since this representa-
tion is evidently rotation-invariant, it allows performing nat-
ural and large deformations on meshes based on translational
(or other) constraints on the handle, as well as linear blend
between different shapes (given the correspondence between
them). However, the reconstruction of the Cartesian coor-
dinates from the pyramid coordinates is clearly a nonlinear Figure 11: Laplacian mesh editing [46]. The red belt bounds
optimization (involving cross-product and norm computa- the region of interest; editing is performed by manipulating
tions to construct the projection planes), which makes it time the handle object represented by the yellow sphere.
consuming.
In [46], the Laplacian editing method that implicitly trans- The last term is quadratic, and is dropped, which results in
forms the differential δ-coordinates is proposed, based on the following linearized form
finding an optimal transform for each vertex. The editing
paradigm implies positional constraints on some vertices, ⎛ ⎞
s −h 3 h2 tx
and it deduces the local rotations based on translations of the ⎜ ⎟
⎜ h3 s −h 1 ty ⎟
handle, similarly to [45]. The local transforms are defined Ti = ⎜
⎜ −h
⎟
by comparing 1-rings in the original mesh and the unknown ⎝ 2 h1 s tz ⎟
⎠
deformed mesh, and are formulated as linear expressions in 0 0 0 1
the vertices of the unknown mesh. The reconstruction Equa-
tion (3) is thus altered to make the differential coordinates For further details on the derivation of the expressions for
similarity-invariant: we solve for the reconstructed geometry Ti ’s see [54]. The linearization causes anisotropic scaling by
V  = {v1 , . . . , vn } by minimizing cos θ along the rotation axis h, where θ is the rotation angle.
Therefore, when large rotations are involved in the editing

n    constraints, this would lead to scaling artifacts. To rectify the
E(V  ) = Ti (V  )δi − L(v )2 + vi − ui 2 (7)
i anisotropy, [46] propose to remove this anisotropic scaling
i=1 i∈U
in postprocessing (by scaling the transformed differential co-
where U is the set of constrained vertices. The transforma- ordinates back to their original size) and reconstruct again
tions Ti are themselves unknowns that linearly depend on the using the basic Equation (3). Note that this repair step also
unknown vertices V  . Each Ti is a transformation that takes handles the isotropic scaling that may have been caused due
the original 1-ring of vertex i to the newly reconstructed one to the scale-invariance of (7). Also note that the size of the
 linear system that corresponds to (7) is 3n × 3n because in
Ti = argmin Ti v j − vj 2 the expressions for Ti ’s there is a dependence between the x,
Ti 
j∈{i} N (i) y and z coordinates of the mesh, and thus we can no longer
solve the system separately for each coordinate. Figure 11
In addition, the Ti ’s are constrained to be translations, rota- shows an example of editing with this method.
tions and scales only, in order to enforce local detail preserva-
tion and to disallow shears. In 2D, this is easy since similarity A 2D variant of the editing algorithm of [46] was success-
transformations have linear form fully applied to textured 2D shapes with meshed interior in
⎛ ⎞ the work of Igarashi et al. [55]. As mentioned, the 2D version
s −h tx of the linear reconstruction problem (7) contains exact linear
⎜ ⎟
Ti = ⎝ h s ty ⎠ expressions for transformations Ti that consist of rotations,
0 0 1 isotropic scale and translations only. Igarashi et al. propose a
different solution to the elimination of isotropic scale in the
In 3D, however, a linearized version of rotation constraint is deformed shape: after the initial solution to (7) given the user
needed, since true 3D rotation matrices cannot be expressed constraints, each triangle of the original shape is fitted into
linearly in 3D. If h represents the rotation axis and H is its corresponding triangle in the deformed shape, mimicking
the skew-symmetric matrix such that H x = h × x, ∀x ∈ R3 , the new orientation and translation but retaining the original
then the class of translations, rotations and scales has the form size of the triangle. Then, another linear system is solved to
consolidate all these fitted triangles together, similar to the
s exp H = s(α I + β H + γ hT h). approaches in [56,57]. Since the 2D editing method works on

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 801

sical differential geometry notions of the first and the second

fundamental forms. Assuming a normal to the tangent plane
is given at each vertex of the mesh, Lipman et al. define two
discrete forms: the first represents a scalar product in the tan-
gent plane of each vertex, and is represented by edge lengths
and the angles between the edges of the 1-ring, projected onto
the tangent plane. The second form is linear and expresses
the height function of the mesh with respect to the tangent
plane. It is represented by the signed distance of the 1-ring
vertices to the tangent plane. Clearly, such a representation
is invariant to the global coordinate frame, because it con-
sists solely of distances and angles (similarly to the pyramid
coordinates [53]).

Lipman et al. [48] prove that, given the values of the dis-
crete forms that were taken from an existing mesh, the mesh
Figure 12: Examples of sketch-based editing of silhouettes Cartesian geometry can be uniquely restored up to global
[58]. The sketched curve hints at the new desired shape of translation and rotation. The linear reconstruction process is
the silhouette. By weakly weighting the sketch constraints in the key issue: to achieve it, two decoupled steps are needed.
the reconstruction process (Equation (7)), the details of the First, the local frames of each vertex are reconstructed. This
shape are preserved, and the deformation follows the hint of can be done by constructing and solving a linear system of
the user. equations that encodes the changes between the local frames
of adjacent vertices. For each mesh edge (i, j) we need an
equation that expresses the difference between {bi1 , bi2 , Ni }
j j
a “volumetric” representation (meshed area of the interior of (the local frame at i) and {b1 , b2 , N j } (the local frame at j)
the shape in the 2D case), the deformations exhibit realistic in the coordinates of the local frame at i
rigidity and shape preservation. Additional important contri- j ij ij ij
b1 − bi1 = α11 bi1 + α12 bi2 + α13 Ni
bution of this work is the integration of new input devices for j ij ij
direct interaction with the shapes, such as two-handed mouse b2 − bi2 = α21 bi1 + α22 bi2 + α23
i
Ni
ij ij ij
devices and touch-pads. N j − Ni = α31 bi1 + α32 bi2 + α33 Ni
The 3D Laplacian editing [46] has been recently extended ij
The coefficients α km of this equation can be completely de-
and adapted for advanced modeling metaphors. Nealen et al. fined by the discrete forms representation. Note that since
[58] employ the Laplacian framework for sketch-based edit- at each vertex we have three local frames vectors, and the
ing that provides an intuitive yet powerful interface for the system of equations is separable in the three coordinates, the
user. The main building block of the interface is manipula- width of the matrix is 3n. By adding modeling constraints on
tion of sketched curves: the user can edit silhouette curves the local frames we arrive at a least-squares system whose
or any other paths on the mesh via direct sketching of the normal equation matrix is of size 3n × 3n. The next step is to
new shape for the curves. The underlying machinery is an reconstruct the actual vertex positions from the local frames.
adaptation of Equation (7), where the positional constraints This is done by solving a second linear system of equations,
are weighted differently according to the intention of the which simply encodes the edge vectors in the coordinates of
user: strong weighting if the intention is to closely follow the local frames
the sketch, or weak weighting for approximate sketching.
ij ij ij
See Figure 12 for an example of such approximate silhou- v j − vi = β1 bi1 + β2 bi2 + β3 Ni , ∀(i, j) ∈ E
ette sketching. In addition, the new editing system allows to
ij
create sharp or smooth ridges and ravines over the sketched Here, the right-hand side is known: the coefficients β k are
curves, by constraining the Laplacians to specific values (at- defined by the discrete forms and the local frame vectors were
tained by user-controlled scaling). Finally, suggestive con- already computed from the previous system. It is possible to
tours (see [59]) can be created by assigning specific rotations add positional modeling constraints to this system and to
to the Laplacians in the region of interest around the sketched solve it in the least-squares sense (it is an n × n system).
curve.
The power of the method of Lipman et al. [48] lies in
The pursuit after a local differential representation that is the fact that the proposed geometry representation is com-
invariant to rigid transformations and enables linear recon- pletely rotation-invariant, and the reconstruction process is
struction mechanism achieved a significant progress in the linear. Figure 13 shows strong deformations obtained with
work of Lipman et al. [48]. In this paper, a new differential this method, while preserving the local details of the sur-
representation is proposed, which bears similarity to the clas- face. The price is solving two linear systems, and it remains

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
802 O. Sorkine / Differential Representations for Mesh Processing

Figure 13: Some editing results using the rotation-invariant representation of [48]. Note the large rotational deformations
achieved and the preservation of local surface details.

unclear whether it is possible to avoid this. Another draw- Local differential representations described above allow
back is that the manipulation of the frames and the positions basic mesh manipulation, as well as advanced editing op-
is decoupled (as in the Poisson editing method of [47]) and erations. By mixing and interpolating the differential repre-
the user must adjust the transformation of the handle frame to sentations of several meshes, it is possible to blend between
the transformation of the handle position. Sole translation of different shapes, perform detail transfer from one shape onto
the handle will not affect the local frames of the mesh here, another, transplant parts of meshes onto other meshes, etc.
unlike in Laplacian editing, for instance, where the needed Local shape morphing with differential δ-coordinates was
local rotations are automatically deduced from the handle first proposed by Alexa [44] (see also the extended pa-
translation. per [38]). When simple linear interpolation of Laplacians
is performed, the result is identical to linear interpolation
It is important to note that all the above-described surface- of Cartesian coordinates, which is known for its artifacts
based deformation methods do not take the volume of the [56,63]. However, Alexa proposed to use varying interpola-
object into account. As a result, self-intersections might hap- tion weights, which lead to better results. For example, when
pen during the editing process, and in general the volume attaching one mesh part to another, the weighting can vary
of the shape cannot be preserved. The problem of local self- proportionally to the distance from the merge region, which
intersections was treated, for example, in [60], where a mul- leads to a more natural, gradual blend. This idea was further
tiresolution framework was used, with details encoded as explored in [46], where the “coating transfer” tool was in-
volumetric elements rather than displacement vectors. How- troduced: by “peeling” mesh detail via smoothing, one can
ever, rigorously treating volume conservation in this manner transfer the details onto another mesh. The peeled coating
leads to time-consuming computations that hinder interactive is represented by differential coordinates; their orientation is
response. Recently, Zhou et al. [61] proposed to augment the adapted to the orientation of the corresponding local frames
Laplacian surface representation with a volumetric graph. on the target mesh, and then they are added to the differential
More precisely, they place a 3D grid of vertices inside the coordinates of the target mesh. Finally, the target mesh with
shape, as well as a layer of additional vertices wrapping the the new coating is reconstructed via Equation (3). Figure 14
shape on the exterior. These additional vertices are connected shows some coating transfer results and mesh transplanting
between themselves as a standard grid, and also linked to the results from [46].
surface mesh. Zhou et al. then apply the Laplacian editing
technique to this volumetric mesh graph. It is observed that When interpolating between different shapes (given they
such deformation tends to preserve the shape volume better have the same connectivity in full correspondence), correct
while retaining local surface detail, at the expense of enhanc- handling of rotations is extremely important. In the simplest
ing the complexity of the representation. Yet, the challenge case, when the target shape is a rigid transformation of the
remains to find a theoretically sound deformation approach source shape (say, rotation by 90 degrees about some axis),
that would tie the surface properties with the volumetric ones. we expect the blending between the two shapes to be gradual
rotation of the source shape towards the target. This does not
happen if we linearly blend the Cartesian (or the differential)
4.4. Shape interpolation using differential coordinates of the two shapes. We need a correct interpolation
representations of the orientation of the shape. In general, naturally-looking
shape interpolation exhibits rigidity and minimizes elastic
For a recent survey on morphing techniques the reader distortion of the in-between shapes [56].
is referred to [62]; below we summarize the newest ap-
proaches related to the Laplacian framework and differential As mentioned, the Laplacians (or the gradients [47]) are
representations. linear functions of the Cartesian coordinates and thus their

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 803

Figure 14: Examples of shape blending using the Laplacian framework [46]. (a-b) show coating transfer (the sources for
coating are framed). In (c-d), the wings of the Feline model were transplanted onto the Bunny.

linear interpolation is equivalent to simple Cartesian interpo- Observe that in order to correctly interpolate representa-
lation. When a volumetric shape representation is available, tions that are linear functions of Cartesian geometry, we have
Alexa et al. [56] proposed to consider the local linear trans- to perform nonlinear interpolation. This is due to the fact
formation between the elements of the simplicial complexes that those representations are not rotation-invariant. The im-
that describe the two shapes, and interpolate separately be- portance of having a linear interpolation method with the
tween the rotational and the shearing parts of this transforma- possibility of linear reconstruction is evident in the work of
tion. This was also employed by Sumner and Popović [57] Sumner et al. [65]. Sumner et al. propose a new algorithm
to transfer local deformations from one mesh animation se- for mesh editing, which relies on a space of example edits.
quence onto another, and by Xu et al. [64] for surface mesh This idea enables easy creation of naturally behaving shapes
morphing. For each pair of source and target mesh triangles, when we have an input data base (e.g., key frames of a manu-
the affine transformation between their local frames is com- ally designed animation or key-frame output of a simulator).
puted by considering the triangles’ edges and normals (this Roughly speaking, given a rest mesh M 0 in some represen-
transformation is in fact the so-called deformation gradient, tation F(M 0 ), the example space is a span of the deformed
see [57]). This affine transformation Hi is factored using the states F(M 1 ), . . . , F(Mk ) of the mesh M 0 , again, using the
polar decomposition: Hi = Ri Si , where Ri is a rotation and Si chosen representation. The representation F(·) should be lo-
is symmetric and represents the elastic (scaling) part of Hi . cal (at least translation-invariant), such as the Laplacian co-
To interpolate between the identity and Hi , Ri is interpolated ordinates or the deformation gradients (which was the choice
in the Lie algebra (using e.g., quaternions) and Si is linearly of Sumner et al.). When the user interactively edits the mesh
interpolated. Therefore, for a given interpolation parameter M 0 , he/she poses positional modeling constraints of the form
t, for each triangle i we have a local transformation (5). The algorithm searches for a shape that (i) satisfies the
modeling constraints and (ii) whose representation is as close
Hi (t) = LIE INTERP(Ri , I , t) · ((1 − t)Si + t I ). (8) as possible to the example space. In other words, the algo-
rithm computes a mesh M and a set of scalar weights w 1 , . . . ,
wk that minimize the following energy:
This transformation Hi (t) is applied to the gradients of the
 
source triangle i, and the intermediate mesh is reconstructed  F(M) − INTERPw ,...,w · (F(M1 ), . . . , F(Mk ))2 (9)
1 k
from the transformed gradients as in [47]. This approach re-
sults in a more natural interpolation of rotations of surface under positional constraints of the form (5). When the exam-
meshes. ple space is very rich (meaning, when a dense set of example
edits is available), it is enough to assume that it is a linear
A notable limitation of the above approach is its failure to space, that is,
correctly interpolate very large rotations: it is impossible to
determine how many laps of rotations an element has gone 
k

through by considering solely the source and target state of INTERPw1 ,...,wk (F(M1 ), . . . , F(Mk )) = wi F(Mi ).
i=1
the element (we will not be able to distinguish between rota-
tion of α and α + 2π k for any integer k). The above approach Since F(·) is a linear function of the coordinates of M, the
always chooses the “shortest path” interpolation, assuming global optimization in (9) can be solved linearly. However,
the minimal angle between each pair of elements. Thus, for when we do not have a dense set of examples, a nonlin-
example, it is impossible to achieve a natural morphing be- ear interpolation is required, so that the generated shapes
tween a straight line and a spiral in 2D, or a straight bar and include correct rotations and extrapolate well. Sumner et al.
its multiply twisted form (as in Figure 15). [65] employ the interpolation method (8), which turns the

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
804 O. Sorkine / Differential Representations for Mesh Processing

Figure 15: Linear shape interpolation sequence of the Bar mesh using the discrete forms representation [48]. Note the natural
rotation of the in-between shapes.

above problem (9) into a global nonlinear optimization, and from its linearity, coupled with the availability of advanced
thus more expensive. linear solvers. We hope that our paper will help familiarizing
researches with this area of study and believe that similar ap-
As mentioned by the authors of [65], a valid alternative for proaches will lead to new useful tools in geometric modeling
representation of the example space could be the pyramid co- yet to be discovered.
ordinates [53]. Since they are rotation-invariant, they can be
simply linearly interpolated in (9). However, the optimization
problem would be still nonlinear because the reconstruction Acknowledgments
of Cartesian coordinates from pyramid coordinates is not lin-
ear (F(·) is not a linear representation in this case). I would like to express my deepest gratitude to my col-
leagues and co-authors, whose work constitutes the core of
The discrete forms [48] provide an especially powerful this paper: Marc Alexa, Daniel Cohen-Or, Doron Chen, Dror
representation for shape interpolation. They are rotation- Irony, David Levin, Yaron Lipman, Andrew Nealen, Chris-
invariant and can thus be linearly interpolated, and the re- tian Rössl, Hans-Peter Seidel and Sivan Toledo. I am also
construction process requires solving two linear systems. In grateful to the anonymous reviewers for their extensive help
the case of the discrete forms, the representation F(·) is not in improving this paper. This work was supported in part by
linear in the Cartesian coordinates, but, figuratively speak- grants from the Israel Science Foundation (founded by the
ing, it can be “factored” so that the reconstruction is done in Israel Academy of Sciences and Humanities) and the Israeli
two linear stages. In the context of morphing, this represen- Ministry of Science. The author’s PhD studies are supported
tation can be viewed as an extension of the 2D representation by the Colton Foundation.
of Sederberg et al. [63]: they represented polygonal curves
by their edge lengths and angles between successive edges,
whereas Lipman et al. [48] represent 3D meshes by edge References
lengths and angles, projected onto the tangent plane, plus the
additional height component. Both methods produce natural 1. M. Fiedler: Algebraic connectivity of graphs. Czech.
shape interpolations. Figure 15 demonstrates a morphing se- Math. Journal, 23, 298–305, 1973.
quence from [48]. Since interpolation of discrete forms does
not require explicit computation of global rotations between 2. F. R. K. Chung: Spectral Graph Theory. American Math-
pairs of mesh elements, but rather only considers the dif- ematical Society, 1997.
ference between adjacent local orientations (which contains
3. M. P. do Carmo: Differential Geometry of Curves and
rotations of up to π radians), it permits correct interpolation
Surfaces. Prentice-Hall, 1976.
of large (“multiple-lap”) rotations.
4. G. Taubin: A signal processing approach to fair sur-
face design. In Proceedings of ACM SIGGRAPH, ACM
5. Conclusions Press, pp. 351–358, 1995.

In this paper, we have described the Laplacian mesh pro- 5. M. Meyer, M. Desbrun, P. Schröder, A. H. Barr: Dis-
cessing framework and reviewed the recent applications of crete differential-geometry operators for triangulated 2-
differential surface representations. The advantages of such manifolds. In Visualization and Mathematics III, Hege,
surface representations are in their capability for local dif- H.-C., Polthier, K., (Eds.). Springer-Verlag, Heidelberg,
ferential shape description and definition of effective bases pp. 35–57, 2003.
for geometry representation. The framework is useful for ap-
plications in geometric modeling where surface details are 6. U. Pinkall, K. Polthier: Computing discrete minimal sur-
important, especially when processing complex scanned sur- faces and their conjugates. Experiment. Math., 2(1): 15–
faces with abundant detail. The power of the framework stems 36, 1993.

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 805

7. M. S. Floater: Mean value coordinates. CAGD, 20(1): 21. I. Guskov, W. Sweldens, P. Schröder: Multiresolution
19–27, 2003. signal processing for meshes. In Proceedings of ACM
SIGGRAPH, ACM Press, pp. 325–334, 1999.
8. M. Botsch, D. Bommes, L. Kobbelt: Efficient linear sys-
tem solvers for mesh processing. IMA Mathematics of 22. A. Khodakovsky, P. Schröder, W. Sweldens: Progres-
Surfaces XI, Lecture Notes in Computer Science, 3604: sive geometry compression. In Proceedings of ACM SIG-
62–83, 2005. GRAPH, pp. 271–278, 2000.

9. S. Toledo: Taucs: A Library of Sparse Linear Solvers, 23. Z. Karni, C. Gotsman: 3D mesh compression using fixed
version 2.2. Tel-Aviv University, Available online at spectral bases. In Proceedings of Graphics Interface,
http://www.tau.ac.il/∼stoledo/taucs/, Sept. 2003. Canadian Information Processing Society, pp. 1–8, 2001.

10. B. Aksoylu, A. Khodakovsky, P. Schröder: Multilevel 24. O. Sorkine, D. Cohen-Or: Least-squares meshes. In Pro-
solvers for unstructured surface meshes. SISC, 26(4): ceedings of Shape Modeling International, IEEE Com-
1146–1165, 2005. puter Society Press, pp. 191–199, 2004.

11. L. Shi, Y. Yu, N. Bell, W.-W. Feng: A fast multigrid 25. O. Sorkine, D. Cohen-Or, D. Irony, S. Toledo:
algorithm for mesh deformation. In Proceedings of ACM Geometry-aware bases for shape approximation. IEEE
SIGGRAPH, To appear, 2006. Transactions on Visualization and Computer Graphics,
11(2): 171–180, 2005.
12. E. G. Boman, B. Hendrickson: Support theory for pre-
conditioning. SIAM Journal on Matrix Analysis and Ap- 26. O. Sorkine, D. Cohen-Or, S. Toledo: High-pass quan-
plications, 25(3): 694–717, 2003. tization for mesh encoding. In Proceedings of the Eu-
rographics/ACM SIGGRAPH Symposium on Geometry
13. S. Guattery, G. L. Miller: Graph embeddings and Lapla- processing, Eurographics Association, pp. 42–51, 2003.
cian eigenvalues. SIAM Journal on Matrix Analysis and
Applications, 21(3): 703–723, 2000. 27. P. Cignoni, C. Rocchini, R. Scopigno: Metro: Measuring
error on simplified surfaces. Computer Graphics Forum,
14. D. Chen, D. Cohen-Or, O. Sorkine, S. Toledo: Algebraic 17(2): 167–174, 1998.
analysis of high-pass quantization. ACM Transactions on
Graphics, 24(4): 1259–1282, 2005. 28. E. Praun, H. Hoppe, A. Finkelstein: Robust mesh water-
marking. In Proceedings of ACM SIGGRAPH, pp. 49–
15. Z. Karni, C. Gotsman: Spectral compression of mesh 56, 1999.
geometry. In Proceedings of ACM SIGGRAPH, pp. 279–
286, 2000. 29. R. Ohbuchi, S. Takahashi, T. Miyazawa, A. Mukaiyama:
Watermarking 3D polygonal meshes in the mesh spectral
16. G. Taubin: Geometric signal processing on polygonal domain. In Proceedings of Graphics Interface, Morgan
meshes. In Proceedings of Eurographics (STAR volume), Kaufmann Publishers, pp. 9–17, (June 2001).
State of the Art Report, 2000.
30. R. Ohbuchi, A. Mukaiyama, S. Takahashi: A frequency-
17. P. Alliez, C. Gotsman: Recent advances in compression domain approach to watermarking 3D shapes. Computer
of 3D meshes. In Advances in Multiresolution for Geo- Graphics Forum (Proceedings of Eurographics), 21(3):
metric Modelling, Dodgson, N., Floater, M., Sabin, M., 373–382, 2002.
(Eds.), Springer-Verlag, pp. 3–26, 2005.
31. D. Cotting, T. Weyrich, M. Pauly, M. Gross: Robust wa-
18. M. Eck, T. DeRose, T. Duchamp, H. Hoppe, M. Louns- termarking of point-sampled geometry. In Proceedings
bery, W. Stuetzle: Multiresolution analysis of arbitrary of Shape Modeling International, pp. 233–242, 2004.
meshes. In Proceedings of ACM SIGGRAPH, pp. 173–
182, 1995. 32. T. W. Sederberg, S. R. Parry: Free-form deformation of
solid geometric models. In Proceedings of ACM SIG-
19. D. Zorin, P. Schröder, W. Sweldens: Interactive mul- GRAPH, pp. 151–160, 1986.
tiresolution mesh editing. In Proceedings of ACM SIG-
GRAPH, pp. 259–268, 1997. 33. S. Coquillart: Extended free-form deformation: A sculp-
turing tool for 3D geometric modeling. In Proceedings
20. A. W. F. Lee, W. Sweldens, P. Schröder, L. Cowsar, D. of ACM SIGGRAPH, 90, pp. 187–196, 1990.
Dobkin: MAPS: Multiresolution adaptive parameteriza-
tion of surfaces. In Proceedings of ACM SIGGRAPH, 34. M. Ranta, M. Inui, F. Kimura, M. Mäntylä: Cut and paste
ACM Press, pp. 95–104, 1998. based modeling with boundary features. In SMA ’93:

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
806 O. Sorkine / Differential Representations for Mesh Processing

Proceedings of the Second Symposium on Solid Model- Shum: Mesh editing with Poisson-based gradient field
ing and Applications (May 1993), pp. 303–312. manipulation. ACM Transactions on Graphics (Proceed-
ings of ACM SIGGRAPH), 23(3): 644–651, 2004.
35. S. Kuriyama, T. Kaneko: Discrete parametrization for
deforming arbitrary meshes. In Proceedings of Graphics 48. Y. Lipman, O. Sorkine, D. Levin, D. Cohen-Or: Linear
Interface, pp. 132–139, 1999. rotation-invariant coordinates for meshes. ACM Trans-
actions on Graphics (Proceedings of ACM SIGGRAPH),
36. H. Biermann, I. Martin, F. Bernardini, D. Zorin: Cut-and- 24(3): 479–487, 2005.
paste editing of multiresolution surfaces. In Proceedings
of ACM SIGGRAPH, pp. 312–321, 2002. 49. R. Zayer, C. Rössl, Z. Karni, H.-P. Seidel: Harmonic
guidance for surface deformation. In Computer Graph-
37. T. Kanai, H. Suzuki, J. Mitani, F. Kimura: Interac-
ics Forum (Proceedings of Eurographics), Eurographics,
tive mesh fusion based on local 3D metamorphosis. In
Blackwell, pp. 601–609, 2005.
Proceedings of Graphics Interface, pp. 148–156, (June
1999. 50. X. Gu, S.-T. Yau: Global conformal surface param-
38. M. Alexa: Differential coordinates for local mesh morph- eterization. In Proceedings of the Eurographics/ACM
ing and deformation. The Visual Computer, 19(2): 105– SIGGRAPH Symposium on Geometry processing, Eu-
114, 2003. rographics Association, pp. 127–137, 2003.

39. L. Kobbelt, S. Campagna, J. Vorsatz, H.-P. Seidel: Inter- 51. X. Ni, M. Garland, J. C. Hart: Fair morse functions for
active multi-resolution modeling on arbitrary meshes. extracting the topological structure of a surface mesh.
In Proceedings of ACM SIGGRAPH, ACM Press, pp. ACM Transactions on Graphics (Proceedings of ACM
105–114, 1998. SIGGRAPH), 23(3): 613–622, 2004.

40. D. Forsey, R. Bartels: Hierarchical B-spline refinement. 52. S. Dong, S. Kircher, M. Garland: Harmonic functions for
In Proceedings of ACM SIGGRAPH, ACM Press, pp. quadrilateral remeshing of arbitrary manifolds. CAGD,
205–212, 1988. 22(5): 392–423, 2005.

41. L. Kobbelt, J. Vorsatz, H.-P. Seidel: Multiresolution hier- 53. A. Sheffer, V. Kraevoy: Pyramid coordinates for mor-
archies on unstructured triangle meshes. Computational phing and deformation. In Proceedings of the Second
Geometry: Theory and Applications, 14: 5–24, 1999. International Symposium on 3DPVT (3D Data Process-
ing, Visualization, and Transmission), IEEE Computer
42. S. Lee: Interactive multiresolution editing of arbitrary Society Press, pp. 68–75, 2004.
meshes. Computer Graphics Forum (Eurographics 99),
18(3): 73–82, 1999. 54. Y. Lipman, O. Sorkine, M. Alexa, D. Cohen-Or, D.
Levin, C. Rössl, H.-P. Seidel: Laplacian framework for
43. M. Botsch, L. Kobbelt: An intuitive framework for real-
interactive mesh editing. International Journal of Shape
time freeform modeling. In Proceedings of ACM SIG-
Modeling, 11(1): 43–62, 2005.
GRAPH, pp. 630–634, 2004.
55. T. Igarashi, T. Moscovich, J. F. Hughes: As-rigid-as-
44. M. Alexa: Local control for mesh morphing. In Proceed-
possible shape manipulation. ACM Transactions on
ings of the International Conference on Shape Modeling
Graphics (Proceedings of ACM SIGGRAPH), 24(3):
and Applications (SMI-01) (Los Alamitos, CA, May 7–
1134–1141, 2005.
11 2001), Werner, B. (Ed.), pp. 209–215.

45. Y. Lipman, O. Sorkine, D. Cohen-Or, D. Levin, C. 56. M. Alexa, D. Cohen-Or, D. Levin: As-rigid-as-possible
Rössl, H.-P. Seidel: Differential coordinates for inter- shape interpolation. In Proceedings of ACM SIG-
active mesh editing. In Proceedings of Shape Modeling GRAPH, pp. 157–164, 2000.
International, IEEE Computer Society Press, pp. 181–
190, 2004. 57. R. W. Sumner, J. Popović: Deformation transfer for tri-
angle meshes. ACM Transactions on Graphics (Proceed-
46. O. Sorkine, Y. Lipman, D. Cohen-Or, M. Alexa, C. Rössl, ings of ACM SIGGRAPH), 23(3): 399–405, 2004.
H.-P. Seidel: Laplacian surface editing. In Proceedings
of the Eurographics/ACM SIGGRAPH Symposium on 58. A. Nealen, O. Sorkine, M. Alexa, D. Cohen-Or: A
Geometry processing, ACM Press, pp. 179–188, 2004. sketch-based interface for detail-preserving mesh edit-
ing. ACM Transactions on Graphics (Proceedings of
47. Y. Yu, K. Zhou, D. Xu, X. Shi, H. Bao, B. Guo, H.-Y. ACM SIGGRAPH), 24(3): 1142–1147, 2005.

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.
 O. Sorkine / Differential Representations for Mesh Processing 807

59. D. DeCarlo, A. Finkelstein, S. Rusinkiewicz, A. Santella: 64. D. Xu, H. Zhang, Q. Wang, H. Bao: Poisson shape in-
Suggestive contours for conveying shape. ACM Trans- terpolation. In Proceedings of the ACM Symposium on
actions on Graphics (Proccedings of ACM SIGGRAPH), Solid and Physical Modeling, ACM Press, pp. 267–274,
22(3): 848–855, 2003. 2005.

60. M. Botsch, L. Kobbelt: Multiresolution surface repre- 65. R. W. Sumner, M. Zwicker, C. Gotsman, J. Popović:
sentation based on displacement volumes. Computer Mesh-based inverse kinematics. ACM Transactions on
Graphics Forum, 22(3): 483–492, 2003. Graphics (Proceedings of ACM SIGGRAPH), 24(3):
488–495, 2005.
61. K. Zhou, J. Huang, J. Snyder, X. Liu, H. Bao, B. Guo, H.-
Y. Shum: Large mesh deformation using the volumetric
graph Laplacian. ACM Transactions on Graphics (Pro- About the author
ceedings of ACM SIGGRAPH), 24(3): 496–503, 2005.
Olga Sorkine received the BSc degree in mathematics and
62. M. Alexa: Recent advances in mesh morphing. Com- computer science from Tel Aviv University in 2000. Cur-
puter Graphics Forum, 21(2): 173–196, 2002. rently, she is a PhD student at the School of Computer Science
at Tel Aviv University. Her research interests are in computer
63. T. W. Sederberg, P. Gao, G. Wang, H. Mu: 2-D shape graphics and include shape modeling, mesh processing and
blending: an intrinsic solution to the vertex path prob- approximation. She has worked in the area of using Laplacian
lem. In Proceedings of ACM SIGGRAPH, ACM Press, transforms for mesh processing and will submit a PhD thesis
pp. 15–18, 1993. on this topic.

c 2006 The Author

Journal compilation 
c 2006 The Eurographics Association and Blackwell Publishing Ltd.