Noname manuscript No.

(will be inserted by the editor)

Meshless Methods for Physics-Based Modeling and

Simulation of Deformable Models

Xiaohu Guo · Hong Qin

Received: date / Accepted: date

Abstract As 3D digital photographic and scanning devices produce higher resolu-

tion images, acquired geometric data sets grow more complex in terms of the modeled
objects’ size, geometry, and topology. As a consequence, point-sampled geometry is
becoming ubiquitous in graphics and geometric information processing, and poses new
challenges which have not been fully resolved by the state-of-art graphical techniques.
In this paper, we address the challenges by proposing a meshless computational frame-
work for dynamic modeling and simulation of solids and thin-shells represented as
point samples. Our meshless framework can directly compute the elastic deformation
and fracture propagation for any scanned point geometry, without the need of con-
verting them to polygonal meshes or higher order spline representations. We address
the necessary computational techniques, such as Moving Least Squares, Hierarchical
Discretization, and Modal Warping, to effectively and efficiently compute the physical
simulation in real-time. This meshless computational framework aims to bridge the
gap between the point-sampled geometry with physics-based modeling and simulation
governed by partial differential equations.

Keywords Meshless Method · Physics-Based Modeling · Physics-Based Simulation ·

Deformable Models

Xiaohu Guo is partially supported by National Science Foundation CCF-0727098. Hong Qin
is partially supported by National Science Foundation IIS-0710819.

X. Guo
Department of Computer Science
University of Texas at Dallas
Tel: +1-972-883-4723
Fax: +1-972-883-2349
E-mail: [email protected]
H. Qin
Department of Computer Science
Stony Brook University
Tel: +1-631-632-8450
Fax: +1-631-632-8334
E-mail: [email protected]
2

1 Introduction

A fundamental question in digital modeling and simulation is the choice of geometric

and functional representation: What mathematical description should be chosen to
represent the surface or volume of a 3D object on a digital computer? The diversity
of the application fields of 3D geometry is reflected in a wide variety of geometric
representations that have been proposed in the past. For example, industrial design
applications for car and airplane construction are mostly based on B-splines or NURBS
[9]; medical applications make frequent use of implicit representations such as level
sets [19]; game and movie industry has focused on polygonal representations such as
triangle meshes [7]. The choice of geometric representation is typically guided by the
following considerations: the number of individual primitives, the descriptive power
or approximation order, the power of handling complicated geometry, and the ease of
performing topological changes.

– B-splines and NURBS can represent complex shapes with relatively few primitives,
since each individual spline patch has a high approximation order. However, com-
bining different patches to a consistent surface model of arbitrary topology can be
difficult, since strict global continuity constraints need to be observed. Furthermore,
it is difficult to model high-frequency features or even discontinuity. Some attrac-
tive properties such as local adaptivity and multi-resolution are rather difficult to
achieve.
– Polygonal meshes are easier to handle as they are defined by a set of vertices with
a consistent adjacency graph, but require a higher number of primitives to achieve
the same geometric accuracy. With increasing resolution of the scanning devices,
geometric data sets are becoming more and more complex, lately reaching billions of
sample points for a single model [17]. Since scanners typically produce a collection
of point samples, many reconstruction methods are used to convert the point cloud
into a polygonal representation. Unfortunately, most of these methods do not scale
well with model size, e.g., many techniques based on the Delaunay triangulation
have a worst-case complexity of O(n2 ), where n is the number of sample points.
Being able to directly process, animate, or simulate point cloud representations
would avoid the need for surface reconstruction entirely.

Nowadays three-dimensional acquisition is an increasingly popular means of cre-

ating surface models. With the dramatic increase of the polygonal complexity of the
acquired graphical models, large scaled point-sampled geometry is becoming ubiqui-
tous in graphics and geometric information processing pipeline. Points have several
unique advantages over traditional primitives such as triangle meshes in modeling, an-
imation, and simulation. For example, they are free of connectivity concerns, which
makes them very suitable for dynamic shape manipulation and physical simulation
when the underlying geometry is undergoing topological changes.
Traditionally in computer animation and physical simulation, meshes are inevitable
for physical simulations since the finite element methods (FEM) require an explicit
mesh structure. However, for some complex physical effects, such as large deforma-
tion and fracture, it will pose grand technical challenges in terms of maintaining the
topological consistency of the underlying meshes. In these cases, efficient and consis-
tent surface (and volumetric) representations are necessary to facilitate geometric and
topological operations.
 3

– The finite element interpolation functions are then built upon the mesh, which
ensures the compatibility of the interpolation. However, this procedure is not always
advantageous, because the numerical compatibility condition is not the same as the
physical compatibility condition of a continuum. For instance, in a Lagrangian type
of computations, one may experience mesh distortion, which can either end the
computation altogether or result in drastic deterioration of accuracy. In computer
simulations of very large deformation, a distorted mesh introduces severe errors
in numerical computations. Therefore, it would be computationally efficacious to
discretize a continuum by only a set of nodal points without mesh constraints.
This is the leitmotif of contemporary meshless methods in mechanical engineering,
scientific computing, and computer animation.
– In simulations of failure processes [23], we need to model the propagation of cracks
along arbitrary and complex paths. This problem, in particular, becomes a notori-
ously difficult task using conventional mesh-based computational techniques such
as finite element method or finite difference method. In essence, the underlying
structure of these methods, which stem from their reliance on meshes, impedes the
flexible modeling and natural handling of discontinuities that do not coincide with
the original mesh lines. Therefore, the most viable strategy for dealing with moving
discontinuities in these methods is to remesh in each time step of the integration so
that mesh lines remain coincident with the discontinuities throughout the simula-
tion. However, this can introduce numerous difficulties for data management, such
as the strong need to map between meshes in consecutive stages of the simulation,
which inevitably results in degradation of both accuracy and complexity for system
implementation. In addition, model remeshing becomes an unavoidable burden. In
sharp contrast, meshless methods can overcome the above difficulties associated
with mesh structure, and avoid complex remeshing operations and the associated
problems of element cutting and mesh alignment sensitivity common in FEM.

In this paper, we try to build high-fidelity meshless computational models of the

point-sampled scanned geometry in the physical world. The objective is to unify com-
plex point-sampled geometry, easy topological change, realistic physical properties and
behavior in a single meshless framework. Unlike meshes or splines, the point-sampled
geometry representation has no intrinsic differential structure, which make it very hard
to incorporate rigorous physical model from continuum mechanics. In this paper, we
introduce our meshless simulation framework, in which the partial differential equa-
tions (for dynamic physical simulation) can be applied and solved directly over point
samples via Moving Least Squares (MLS) shape functions defined on the parametric
domain without explicit connectivity information. Our meshless framework can per-
form efficient simulation for elastic thin-shell and solid deformations. For thin-shell
simulation, the simulation domain is not naturally given by the point geometry, which
make it necessary to utilize parameterizations for the point-sampled surfaces. For solid
deformation, the number of volumetric nodes is typically prohibitive for real-time sim-
ulation. We exploit the methodology of Modal Analysis and adapt the Modal Warping
technique into our volumetric meshless simulation framework to achieve real-time ma-
nipulation and deformation.
The remainder of this paper is organized as follows. In section 2 we briefly review
the prior work utilizing meshless methods for physics-based modeling and simulation.
In section 3 we describe the detailed theoretical derivations for the meshless method,
particularly the Moving Least Squares method. In section 4 we describe the hierarchical
4

discretization methods for our meshless framework. For solid simulation, we use octree
discretization based on implicit functions of the point-set surface (section 4.1). For
thin-shell simulation, we use quadtree discretization based on the global conformal
parameterization of the surfaces (section 4.2). In section 5 we describe the differential
equations used to represent the solid and thin-shell dynamics, and the Modal Warping
technique that we use to solve the dynamic system in real-time. Section 6 shows the
statistics and performance data of our experimental results. Finally, we conclude this
paper with several possible directions for future work in section 7.

2 Previous Work

Meshless (mesh-free) methods [4, 18] have been developed in the field of mechanical
engineering to enable solving partial differential equations (PDEs) numerically. The
meshless computation is based on a set of scattered nodes without having to recourse
to an additional mesh structure. The meshless methods require only a set of nodes
distributed across the entire analysis domain. The shape function associated with each
node is then constructed to approximate (or interpolate) the field functions using their
values at the sampling nodes in the analysis domain without explicit connectivity,
while satisfying certain basic requirements, such as compact support for computational
accuracy and efficiency, stability and consistency to ensure numerical convergence, etc.
Meshless method was introduced into the graphics and animation field by Desbrun
and Cani [8] using a particle system coated with a smooth iso-surface for animating soft
inelastic substance which undergo topological changes. Later they applied Smoothed
Particle Hydrodynamics (SPH) to simulate highly deformable bodies. Müller et al. [21]
presented a method for modeling and animating elastic, plastic, and melting volumetric
objects based on the MLS approximation of the displacement field. Most recently, they
presented a geometrically motivated approach in [22] for simulating deformable point-
based objects. Pauly et al. [27] simulated volumetric meshless fracture with a highly
dynamic surface and volume sampling method that affords complex fracture patterns
of interacting and branching cracks. Guo and Qin [11] combined meshless method with
modal analysis framework to provide real-time deformation of volumetric objects. The
most relevant work to our meshless thin-shell simulation is the approach proposed
by Wicke et al. [29], which uses locally defined fibers to approximate the differential
surface operators of the thin shell functional.

3 Meshless Methods

During the last two decades, meshless methods have been developed that enable solv-
ing PDEs numerically, based on a set of scattered nodes without having recourse to an
additional mesh structure (which must be put in place for the traditional finite element
methods). The unique advantages of meshless methods are multifold: (1) there is no
need to generate a mesh of nodes – they only need to be scattered within the analysis
domain, which is much easier to handle in principle; (2) moving discontinuities such as
cracks can be naturally facilitated, since no new mesh needs to be constructed as in
finite element methods, and the computational cost of remeshing at each time step can
be avoided entirely; (3) properties such as spatial adaptivity (node addition or elimi-
nation) and shape function polynomial order adaptivity (approximation/interpolation
 5

types) can streamline the adaptive model refinement and simulation in both time and
space; and (4) data management overhead can be minimized during simulation. In
1994, Belytschko et al. proposed the Element Free Galerkin (EFG) method [2] to solve
linear elastic problems, specifically the fracture and crack growth problems [3], in which
the Moving Least Squares (MLS) interpolant was employed in a Galerkin procedure.
In this paper we focus on the Element Free Galerkin method, mainly because it has
been well-developed with mature techniques, and it has shown a superior rate of con-
vergence and high efficiency in modeling moving interfaces. Other variants of available
mesh-free methods can also be adopted into our prototype system in a straightforward
way without theoretical obstacles.
In a nutshell, the meshless methods require only a set of nodes distributed across
the entire analysis domain. The shape function associated with each node is then
constructed to approximate (or interpolate) the field functions using their values at
the sampling nodes in the analysis domain. For finite element methods, however, the
shape functions are constructed utilizing the mesh of the elements. In sharp contrast,
the shape functions in meshless methods are constructed using only the sampling nodes
without any connectivity, while satisfying certain basic requirements, such as compact
support for computational accuracy and efficiency, stability and consistency to ensure
numerical convergence, etc. The shape functions in the EFG method are constructed
by using the MLS technique, or alternatively on the basis of reproducibility conditions
(note that both approaches can arrive at the same expressions for the shape functions),
and it can provide continuous and smooth field approximation throughout the analysis
domain with any desirable order of consistency. To make our paper self-contained,
we present a brief introduction of the MLS approximation for the definition of shape
functions in the following subsection.

3.1 Moving Least Squares Shape Functions

The moving least squares method can be traced back to its original application in
scattered data fitting, where it has been studied under different names (e.g., local
regression, “LOESS”, weighted least squares, etc.) [16].
We associate each node I with a positive weight function wI of compact support.
The support of the weight function wI defines the domain of influence of the node: ΩI =
{x : wI (x) = w(x, xI ) > 0}, where w(x, xI ) is the weight function associated with node
I evaluated at position x. The approximation of the field function f at a position x
is only affected by those nodes whose weights are non-zero at x. We denote the set of
such nodes the active set A(x). Figure 1 illustrates the meshless computational model
with rectangular and circular local influence domains, respectively.
Let f (x) be the field function defined in the analysis domain Ω, and f h (x) the
approximation of f (x) at position x. In the MLS approximation, we let

m
X
f h (x) = pi (x)ai (x) = pT (x)a(x), (1)
i=1

where pi (x) are polynomial basis functions, m is the number of basis functions in the
column vector p(x), and ai (x) are their coefficients, which are functions of the spatial
coordinates x. In our implementation, we utilize 3-D linear basis functions: pT (m=4) =
6

Influence Domain Node Influence Domain Node

Analysis Domain Object Boundary Analysis Domain Object Boundary

(a) (b)
Fig. 1 The meshless computational model with rectangular (a) and circular (b) support,
respectively.

{1, x, y, z} in the interest of time performance. We can derive a(x) by minimizing a

weighted L2 norm:
X
J= w(x − xI )[pT (xI )a(x) − fI ]2 , (2)
I∈A(x)

where fI is the nodal field value associated with the node I. We can rewrite Equation
(2) in the form:
J = (Pa − f )T W(x)(Pa − f ), (3)
where
f T = (f1 , f2 , . . . fn ),
 
p1 (x1 ) p2 (x1 ) · · · pm (x1 )
 p1 (x2 ) p2 (x2 ) · · · pm (x2 ) 
P=
 .. .. .. .. ,

. . . .
p1 (xn ) p2 (xn ) · · · pm (xn )
and  
w(x − x1 ) 0 ··· 0
 0 w(x − x2 ) ··· 0 
W(x) = 
 .. .. .. .. .

. . . .
0 0 · · · w(x − xn )
To find the coefficients a(x), we obtain the extremum of J by setting
∂J
= A(x)a(x) − B(x)f = 0, (4)
∂a
where the m × m matrix A is called moment matrix :

A(x) = PT W(x)P,

B(x) = PT W(x).
 7

So we can obtain:
a(x) = A−1 (x)B(x)f . (5)
And the shape functions are given by:

φ(x) = [φ1 (x), φ2 (x), . . . φn (x)] = pT (x)A−1 (x)B(x). (6)

If we consider the field function as a function of both space and time f (x, t), the
approximation in the analysis domain Ω can be written as:
X
f (x, t) ≈ f h (x, t) = φI (x)fI (t), (7)
I∈A(x̂)

The moment matrix A may be ill-conditioned when (i) the basis functions p(x) are
(almost) linearly dependent, or (ii) there are not enough nodal supports overlapping
at the given point, or (iii) the nodes whose supports overlap at the point are arranged
in a special pattern, such as a conic section for a complete quadratic polynomial basis
p(x). Note that the necessary condition for the matrix A to be invertible is

∀x ∈ Ω card{I : x ∈ ΩI } > m, (8)

which can be automatically guaranteed using the octree-based (for volumetric models)
and quadtree-based (for surface models) node placement method to be explained in
the next section. Furthermore, the spatial derivatives of the shape functions can be
obtained by noting that the differentiation of Equation (4) yields:

A,i a + Aa,i = B,i f ,

∂a
where a,i denotes ∂xi . Then we can obtain the derivative of a,i by:

a,i = A−1 (B,i f − A,i a). (9)

So the factorization of A in Equation (6) can be re-used for the computation of the
derivatives with little extra cost.

3.2 Basis and Weight Functions

To obtain a certain consistency of any desirable order of approximation, it is necessary

to have a complete basis. The basis functions p(x) may include some special terms such
as singularity functions, in order to ensure the consistency of the approximation and
to improve the accuracy of the results. The following gives two examples of complete
bases in 3 dimensions for first and second order consistency:

Linear : pT
(m=4) = {1, x, y, z}, (10)

Quadratic : pT 2 2 2
(m=10) = {1, x, y, z, x , xy, xz, y , yz, z }. (11)
The weight functions w(x, xI ) play important roles in constructing the shape func-
tions. They should be positive to guarantee a unique solution for a(x); they should
decrease in magnitude as the distance to the node increases to enforce local neigh-
bor influence; they should have compact supports, which ensure sparsity of the global
8

matrices. They can differ in both the shape of the domain of influence (e.g., paral-
lelepiped centered at the node for tensor-product weights, or sphere), and in functional
form (e.g., polynomials of varying degrees, or non-polynomials such as the truncated
Gaussian weight). In our implementation, we choose the parallelepiped domain of in-
fluence for the ease of performing numerical integrations, and utilize the composite
quadratic tensor-product weight function:

x − xI y − yI z − zI
w(x, xI ) = c( )c( )c( ), (12)
dmxI dmyI dmzI

where dmxI , dmyI , and dmzI denote half of the length of the supporting parallelepiped
sides (for the tensor-product weights) along three directions, respectively. The compo-
nent function c(s) is analytically defined as:

 (1 − 2s2 ) f or 0.5 > s ≥ 0
c(s) = 2(1 − s)2 f or 1 > s ≥ 0.5 (13)

0 f or s≥1

One key attractive property of MLS approximations is that their continuity is

directly related to the continuity of the weighting functions. Thus, a lower-order poly-
nomial basis p(x) such as the linear one can still be used to generate highly continuous
approximations by choosing appropriate weight functions with certain smoothness re-
quirements. Therefore, compared to the finite element method, there is no need for
post-processing to generate smooth stress and strain fields. This can facilitate the
direct and fast visualization of the physical properties of volumetric objects for me-
chanical analysis. Note that the FEM equivalents can also be reached if the weight
functions are defined as piecewise-constant entities over each influence domain.

4 Hierarchical Discretization for Meshless Dynamics

The general idea of meshless methods is to create overlapping patches ΩI comprising a

cover {ΩI } of the domain Ω with shape function φI subordinate to the cover ΩI . One
way to create the meshless discretization is to start from an arbitrarily distributed set of
nodes. No fixed connections between the nodes are required. The nodes are the centers
of the overlapping patches Ωi , which can be either parallelepiped or spherical domains.
However, due to the rather unstructured distribution of nodes over the domain some
algorithmic issues may arise: (1) a discretization without structure does not allow
determination of the patches that contribute to a certain integration point without
performing an expensive global search; (2) the moment matrix A in moving least
squares shape function may become invertible if the patch covering conditions (e.g.,
Equation 8) are not satisfied; (3) the effective handling of the interaction between
scattered nodes with the geometric boundary (e.g. the surface of a volumetric model in
our solid simulation system) becomes very difficult. From the implementation’s point of
view, it is very important that the patches are clearly defined. The interaction between
the patches themselves, and between the patches and the boundary, has to be well
understood and easily accessible during the runtime of the system execution. These
problems can be solved perfectly with the assistance of octree discretization (for solid
models) and quadtree discretization (for thin-shell models).
 9

4.1 Octree-Based Discretization for Solid Models

In our meshless volumetric simulation system, the input data is an unstructured point
cloud comprising a closed manifold surface. If we conduct our dynamic simulation
solely on surface points, many difficulties arise. First, performing inside/outside tests
based entirely on surface point information is a forbidding task with many ambiguities.
Second, point insertion is unavoidable if a deformation is large and in fact spreads out
across the model rather significantly, in which case gaps will occur at the current
resolution. To ameliorate, we compute a volumetric distance field for the input surface
points. Such a distance field, which expands to the entire volumetric domain, will also
aid in the selection of volumetric points at the interior of solid objects for the dynamic
simulation governed by the EFG method. Let us first briefly review some relevant
work which leads to the construction of octree-based distance fields for point surfaces.
Pauly et al. [26] rely on the moving least squares surface projection operator for both
inside/outside tests and point insertion. However, ambiguities would still occur in many
degenerate cases if we only use the moving least squares surface projection operator
[30]. Guo et al. [10] proposed to embed the point set surfaces into volumetric scalar
fields to facilitate surface representation, surface editing, dynamic point re-sampling,
collision detection, etc. Implicit surfaces can be considered a natural and powerful
tool for modeling unstructured point set surfaces for the following reasons: (i) the
inside/outside test can be performed by directly utilizing the implicit function; (ii) the
topology of the implicit surface can be easily updated without any ambiguity. Figure
2 (a) shows the visualization of distance fields using a color-coded 2D slice.

(a) (b) (c)

Fig. 2 (a) Distance field visualization using a color-coded 2D slice; (b) Octree-based dis-
cretization; (c) Volumetric node placement.

In our implementation, we utilize multi-level partition of unity(MPU) implicit sur-

face construction method proposed by Ohtake et al. [24]. The multi-level approach
allows us to construct implicit surface models from large point sets based on an octree
subdivision method that adapts to variations in the complexity of the local shape.
We also observed that the octree discretization of the volume can provide a structure
to construct the patches which would provide a priori information with respect to
the size and interactions of the patches [15]. The octree subdivides the volume of an
object represented as point set surface into cubes, giving a non-overlapping discrete
representation of the domain, on which efficient numerical integration schemes can be
employed. The octants serve as the basic unit from which to construct the patches and
10

allow the efficient determination of patch interactions. In the following subsection, we

will describe the use of the octree structure as the basic building block to help us define
our mesh-free patches and integration cells.

4.1.1 Octree-Based Volumetric Node Placement

An octree structure can be defined by enclosing the object domain of interest Ω in a

cube which represents the root of the octree, and then subdividing the cube into eight
octants of the root by bisection along all three directions. The octants are recursively
subdivided to whichever levels are desired. Note that the terminal level used for our
node placement does not need to coincide with the terminal level of the MPU implicit
surface construction. Actually, in our implementation, the size of the terminal octant
used for our volumetric node placement (for mesh-free simulation) is much larger than
the terminal octant used for MPU implicit surface reconstruction because the surface
point density is much larger compared to the volumetric node density. Figure 2 (b)(c)
shows the octree-based discretization for the MPU implicit surface construction and
volumetric node placement. We restrict the octree to be a one level adjusted octree,
where the level difference of all terminal octants and their face and edge neighbors is
no more than one. This restriction can facilitate the automatic satisfaction of patch
covering condition (Equation (8)) as we will discuss later.

( (%(% (%
(% (
( ( ( ( ( (
( ,%,% ,%
,% (
( ( (
( ( ,%
,% (%
(%
( ( (
,LQWHULRU
( ( ( (
(
( (%
(%
(%% ,
,%
,%
,
,%
,%
(%
(
( ( ( ( (H[WHULRU
((%
%
((%
%
,,%
,,%
%
%
((%
%
,
(%
,,%% ((%
(%% ,
% ( ((%% ,
,
,
((%
((%
%
%((% % ((%
%
(
((%%
(%%
,,%
%
((%%
,
((%
((%
%
% ,%LQWHULRUERXQGDU\
(
(
( (%
(
,,%
%
((%
, , , ,
% ,%
,%
,
,
,
,
((%
,%
,%
,
%
,
(%%
,
((%
% ,%
,%
, ,
((%
%
((%%
(% (%H[WHULRUERXQGDU\
( (% , ,%
,% ( ( ( ( ( (
( ,%
,% ((%% (
(
(%%
(
,%
,%
(% ,%
(% ,%
, , ,%
,
,% ((%
%
,
% , ,%
((%
((%
%
,% ,%
,%
(
(%
(
(
(
0RGHO
( (
,%
,% ,%
,% ((%% (% ( ( ( (%(% ,%
,% , ((%
% (
( (%% ,%
,% (% ( (%% ,%
,% (
( ( ( (
( (%% (%
(% ( ( (%% (%
(% (

Fig. 3 The definition of interior, exterior, interior boundary, and exterior boundary octants
for mesh-free simulation.

Since we already have the implicit surface representation of the object, we can
easily classify each terminal octants as interior (I) octants OI , exterior (E) octants
OE , and boundary (B) octants OB (see Figure 3). Interior octants are those that
are fully embedded in the interior of the geometric domain Ω. Exterior octants are
those that are located totally outside of Ω, and boundary octants are those that are
intersected by the boundary of Ω. The boundary octants are further classified into
interior boundary (IB) OIB and exterior boundary (EB) OEB octants. The simple rule
is that the centroid of an IB octant is located within the domain, whereas the centroid
of an EB octant is located outside the domain. After the geometric classification, we
can place a volumetric node (for mesh-free dynamics) at the center of each interior (I)
and boundary (IB, EB) octant. For an EB octant, the node should be displaced by
projecting from its center onto the implicit surface to ensure that each node resides in
 11

Ω. Let octant Oi ∈ OI ∪ OB and node i reside in Oi , the open cover associated with
node i is a cube of size α · size(Oi ) centered around node i (see Figure 4 (a)). Both the
volumetric nodes and their open cover regions are necessary constituents for mesh-free
dynamics.

(
a size OI)

open cover
integration cell

size(OI)
integration point
node I

octant

(a) (b)
Fig. 4 (a) The definition of open cover {ΩI } regions based on the octree structure for mesh-
free patches; (b) The interaction between open covers and integration cells, and the integration
points for Gaussian quadrature.

The open cover construction based on terminal octants can provide the structure
needed to perform efficient neighboring search and patch intersection test. It has been
proved in [15] that by choosing a suitable size for α, the validity of the open cover can
be guaranteed a priori. For example, for a linear basis p(x)T (m=4) = {1, x, y, z}, any
point in the domain will be covered by at least 4 patches if we choose α to be 3. The
generation of an octree is much more efficient than a finite element mesh in practice.
Furthermore, the octree allows refinement of the discretization in areas of singularities
if necessary (e.g., near the crack surface).

4.1.2 Octree-Based Gaussian Integration for Matrix Assembly

In order to assemble the entries of the system matrices, such as the mass matrix or
stiffness matrix, we need to integrate over the problem domain. This can be performed
through numerical techniques such as Gaussian quadrature, using the underlying in-
tegration cells. The integration cells can be totally independent of the arrangement of
nodes. The integration cells are used merely for the integration of the system matrices
but not for field value interpolation. In our octree-based discretization scheme, since
the terminal octants do not overlap (except on their shared boundaries), we can fur-
ther subdivide the terminal octants OI and OB into smaller cells and use them as the
integration cells (see Figure 4 (b)). There may exist some integration cells that do not
entirely belong to the analysis domain. We can easily separate the portion of the cell
which lies outside of the domain by evaluating the implicit function (used for repre-
senting the surface distance field). The creation of the open cover and the integration
cells, as described here, eliminates any global searching for members of the open cover
during matrix assembly and time integration. With the prior knowledge of the value
α and utilizing the direct face neighbor links, all patches covering a integration point
x ∈ Ω can be found in O(1) time.
12

4.2 Parameterization-Based Quadtree Discretization for Thin-Shell Models

The implicit surface representation mentioned above is essentially volumetric embed-

ding. It does not admit a natural, 2-dimensional domain for the effective analysis
of point-sampled surfaces. As a result, simulating physical behavior of point-sampled
thin-shells becomes rather difficult, compared with the direct simulation of volumetric
models. Although the volumetric simulation mechanism is both natural and intuitive, it
is essentially plagued by the additional computational burden in both time and space.
In many real-world applications, such as the deformation of the wings of the gargoyle,
or even open surfaces like a plate, thin-shell simulation is much better than volumetric
one, since one dimension of the surface body is much smaller than the other two, and
the volumetric simulation may fail if the neighboring volumetric nodes are arranged
in degenerate locations, such as a plane. To combat the deficiency associated with the
dimensional increase, we can utilize the global conformal parameterization [12] of the
point-sampled surfaces, to obtain a natural 2-dimensional parametric domain for simu-
lation node placement and numerical integration. Figure 5 shows the global conformal
parameterization for the David head surface model. Figure 6 shows the simulation
nodes (and their support radii) placed on the Moai surface model and it corresponding
parametric domain.

(a) (b) (c) (d)

Fig. 5 The global conformal parameterization (b) for the David head model (a). The checker-
board texture on the parametric domain (d) can be mapped onto the surface (c).

(a) (b) (c)

Fig. 6 The simulation nodes placed on the Moai surface model (a) and its parametric do-
main (b). The larger translucent white hemispheres represent the support radii of two of the
simulation nodes. (c) shows the deformed Moai surface using the set of simulation nodes.
 13

One important advantage of meshless method is the flexibility of the sampling pat-
tern. In the thin-shell simulation, it would be more desirable to have a initial sampling
scheme such that the sampling nodes are uniformly distributed on the manifold surface
(Figure 8 (a)). Similar to the idea of using octree structure to facilitate the volumetric
sampling of the volumetric space in Section 4.1, we utilize a quadtree structure on
the parametric domain. The subdivision depth of the quadtree is dependent on the
conformal factor λ. Suppose the size of the quadtree cell is l. If λl is larger than a
threshold (i.e. the surface patch corresponding to the cell is still large enough), we
keep subdividing the cell into 4 child-cells. We place the sampling nodes based on the
quadtree discretization of the parametric plane. Since the conformal factors in u and v
directions are equivalent, we can choose simple “symmetric” supporting region (such
as squares, or disks) for each simulation nodes. In our implementation, we use square-
shaped supporting region for the ease of performing numerical integrations. For a node
i reside in quadtree cell Qi , its supporting region is a square of size η ·size(Qi ) centered
around node i. We restrict the quadtree to be a one-level adjusted quadtree, where the
level difference of all terminal cells and their edge neighbors is no more than one. This
restriction can facilitate the automatic satisfaction of patch covering condition in order
to make the moment matrix invertible. It has been proved in [15] that by choosing a
suitable size for η, the validity of the supporting region can be guaranteed a priori. For
example, for a linear basis p(x)T = [1, θ 1 , θ 2 ], any point in the domain will be cov-
ered by at least 3 patches if we choose η to be 3. The supporting region construction
based on terminal quadtree cells can provide the structure needed to perform efficient
neighboring search and patch intersection test. The quadtree cells can be also utilized
as integration cells to perform numerical integration similar to the volumetric case in
Section 4.1.2.

(a) (b)
Fig. 7 (a) The quadtree structure on the parametric plane for placing simulation nodes; (b)
the dynamic re-sampling near the crack line (blue curve).

In the parameterization stage, the seams between different parametric patches are
represented by additional point set curves, which are generated by tracing integral
curves. These additional point set curves are only used in the parameterization step,
i.e., they are not added to the original point set surface after the parameterization. In
the simulation stage, the parametric seams are maintained in a table indicating the
connection correspondence of different parametric patches. The simulation nodes are
not duplicated on the patch boundaries. The support of the shape function associated
with each simulation node is not restricted to the parametric patch where it resides.
In fact, the support can be expanded to other parametric patch if the node is close to
14

the boundary of its patch (see Figure 8 (c)). So the behaviors of the nodes across the
boundaries are consistent with non-boundary nodes without any unnatural artifact.

(a) (b) (c)

Fig. 8 (a) Uniform sampling nodes on the surface; (b) the 4 parametric planes; (c) the sup-
porting regions of two simulation nodes are shown in pink and green; their overlapping region
is shown in yellow.

When a crack is generated in a body, the dependent variables (e.g. the displace-
ments, etc.), must be discontinuous across the crack. Furthermore, the support of the
nodes affected by the discontinuities need to be modified accordingly to incorporate
the proper behavior of the shape functions and its derivatives. Similar to the approach
of [27], we use the transparency criterion proposed in [25] to allow partial interaction of
nodes in the vicinity of the crack front. Note that all these operations can be easily per-
formed on the 2D parametric domain. We need to perform dynamic up-sampling (node
insertion) during the crack simulation, in order to maintain numerical stability even for
an initially adequately sampled model. New simulation nodes need to be inserted in the
vicinity of crack lines. We take the same criterion as [27] to determine under-sampling
at each node based on transparency weights. If the transparency weights becomes too
small due to a nearby crack line, we subdivide the quadtree cell associated with the
simulation node into 4 child-nodes (see Figure 7 (b)).

5 Dynamic Deformation System

In our meshless approximation, the motion parameters of the material point x, i.e.,
the displacements u, velocity u̇, and acceleration ü, can be approximated by using the
moving least squares shape functions φI (x) as:
X
u(x, t) = φI (x)uI (t), (14)
I

X
u̇(x, t) = φI (x)u̇I (t),
I
X
ü(x, t) = φI (x)üI (t).
I
 15

Note that uI , u̇I , and üI are not the nodal values of displacements, (velocities, etc.),
but rather nodal parameters without a direct physical interpretation, because the shape
functions φI (X) produce approximation, not interpolation of the field values. The par-
tial derivatives with respect to the referencing coordinates xk can be obtained simply
as: X
u,k (x, t) = φI,k (x)uI (t).
I

5.1 Dynamic Elastic Solids

We use the Euler-Lagrange equations for the elastic deformation of solid models:
d ∂T (u̇) ∂V (u)
( ) + µu̇ + = Fext , (15)
dt ∂ u̇ ∂u
where the kinetic energy T and elastic potential energy V are functions of u̇ and u,
respectively. The term µu̇ is the generalized dissipative force, and Fext is a generalized
force arising from external body forces, such as gravity.
The kinetic energy of the moving body can be expressed as:
Z
1 1 X IJ
T = ρ(x)u̇ · u̇dΩ = M u̇I · u̇J , (16)
2 Ω 2
I,J
R
where ρ(x) is the mass density of the body, and M IJ = Ω ρ(x)φI (x)φJ (x)dΩ. Then
we can have:
d ∂T (u̇)
( ) = Mü, (17)
dt ∂ u̇
where the matrix M composed of the elements MIJ is called the mass matrix.
The elastic potential energy of a body can be expressed in terms of the strain tensor
and stress tensor. The strain is the degree of metric distortion of the body. A standard
measure of strain is Green’s strain tensor:
∂ui ∂uj ∂u ∂ul
²ij = + + δkl k . (18)
∂Xj ∂Xi ∂Xi ∂Xj
Forces acting on the interior of a continuum appear in the form of the stress tensor,
which is defined in terms of strain:
ν
τij = 2G{ tr(²)δij + ²ij }, (19)
1 − 2ν
P
where tr(²) = ij δij ²ij . The constant G is called the shear modulus, which determines
how strongly the body resists deformation. The coefficient ν, called Poisson’s ratio, de-
termines the extent to which strains in one direction are related to those perpendicular
to it. This gives a measure of the degree to which the body preserves volume. The
elastic potential energy V (u) is given by the formula:
Z X
ν
V =G { tr2 (²) + δij δkl ²ik ²jl }dΩ, (20)
Ω 1 − 2ν
ijkl

By combining the above Equations (14), (18), and (23), we can formulate the
derivatives of V (with respect to u) as polynomial functions of u, the coefficients of
which are integrals that can be pre-computed using the techniques mentioned in Section
4.1.2.
16

5.2 Dynamic Elastic Thin-Shells

For any point-sampled surface, if we assume that one dimension (i.e. the thickness),
of the surface body is significantly smaller than the other two dimensions, we can
consider the point-sampled surface as a thin-shell. In the Kirchhoff-Love thin shell
framework, the deformation of the surface body is fully described by the deformation
of the middle surface represented by point-samples. Let ϕ denote the position of a
point on the middle surface of the shell, and let a3 be the unit director vector which
is normal to the shell surface. Given the global parameterization of the point-sampled
surfaces, we can describe the positions of any material point in the reference (denoted
r̄) and deformed (denoted r) configurations of the shell by:

r̄(θ1 , θ2 , θ3 ) = ϕ̄(θ1 , θ2 ) + θ3 ā3 (θ1 , θ2 ), (21)

r(θ1 , θ2 , θ3 ) = ϕ(θ1 , θ2 ) + θ3 a3 (θ1 , θ2 ), (22)

where θ1 and θ2 are parameters of the point-sampled middle surface, and θ3 (− h2 ≤
θ3 ≤ h2 ) is in the thickness direction.
The precise form of the membrane and bending strain and stress matrices are given
in [6]. Similar to the elastic solids, we use the Euler-Lagrange equation (24) for our
dynamic thin-shell deformation. The kinetic energy of the moving thin-shell is defined
similar to the elastic solids in equation (16), except that the integration domain Ω is
the global conformal parametric domain. The elastic potential energy V is given by
the formula: Z
Eh Eh3
V = [ 2
αT H̃α + 2)
β T H̃β]dΩ. (23)
Ω 1 − ν 12(1 − ν

where α and β are the membrane and bending strains respectively, H̃ is the stan-
dard constitutive matrix, the constant E is Young’s modulus, and the coefficient ν is
Poisson’s ratio. More specific derivations can be found in [6].

5.3 Solving the Dynamic System

The system of ordinary differential equations which results from the application of
the Element-free Galerkin discretization of the spatial domain can either be integrated
directly, or analyzed by mode superposition. That is, the time dependent solution can
be expressed as the superposition of the natural (or resonant) modes of the system.
In the following section, we will briefly introduce some basics of Modal Analysis. More
detailed discussions can be found elsewhere [28, 14, 13, 5].

5.3.1 Basics of Modal Analysis:

Consider the discretized Euler-Lagrange equations for elastic deformation:

Mü(t) + Cu̇(t) + Ku(t) = F(t) (24)

where M, C, and K are the mass, damping and stiffness matrices, respectively, F is the
external load vector and u(t) is the vector of nodal displacements. Under the commonly
adopted Rayleigh damping assumption, we can replace the damping matrix with C =
αM + βK, where α and β are weighting coefficients. For linear elasticity models, both
 17

M and K are constants. Let the columns of Ψ be the solution to the generalized
eigenvalue problem Kx = λMx, and Λ be the diagonal matrix of eigenvalues, then
equation (24) can be transformed to:

z̈ + (αI + βΛ)ż + Λz = ΨT F, (25)

where z = Ψ−1 u is the vector of modal amplitudes, and Ψ is called modal displacement
matrix whose i-th column represents the i-th mode shape (see Figure 9). The decou-
pled ODEs in Equation (25) can be computed independently and combined by linear
superposition. The computational loads can be further reduced by removing modes
that are too stiff to be observed (corresponding to higher eigenvalues). So we can take
only l dominant columns of Ψ, to reduce the amount of computation significantly.

(a) (b) (c) (d) (e)

Fig. 9 (a) is the beam model before deformation; (b)-(e) show the 6th , 7th , 8th , and 11th
mode shapes, respectively.

5.3.2 Modal Warping for Rotational Deformation:

Our Meshless Modal Analysis framework is build upon the Modal Warping technique
proposed by Choi and Ko [5]. Their innovative approach tracks the local rotations that
occur during the deformation based on the infinitesimal rotation tensor, and warps the
pre-computed modal basis in accordance with the local rotations of the mesh nodes.
For the space limit, we only briefly introduce their general ideas here. More specific
technical details and proofs can be found in [5].
Considering an infinitesimal deformation with displacement u, the rotation tensor
is defined as:
1
ω = (∇ × u)× = w×, (26)
2
where ∇ × u is the curl of the displacement, w× denotes the standard skew-symmetric
matrix of vector w. Here w = 21 (∇ × u) can be considered as a rotation vector that
causes the rotation by angle kwk around the unit axis w/kwk. In the Modal Analysis
setting, the rotation vector can be expressed in terms of the modal amplitude z:

1
w(x) = (∇×)Φ(x)Ψz (27)
2

where Φ(x) is the vector of MLS shape functions evaluated at position x.

The basic idea of the Modal Warping approach is to embed a local coordinate
frame at each simulation node (see Figure 10). The rotation matrix Ri of the local
coordinate frame associated with node i can be computed from its rotation vector wi .
For a general non-linear elastic deformable model, the stiffness matrix K(u) is not a
18

(a) (b)
Fig. 10 A solid rabbit model before (a) and after (b) deformation. The green spheres are the
discretized simulation nodes inside the rabbit surface. The local coordinate frames associated
with the simulation nodes are rotated during the deformation.

constant. In order to apply the linear Modal Analysis method, it has been shown in [5]
that the non-linear Euler-Lagrangian equations

Mü + Cu̇ + K(u)u = F (28)

can be approximated using the displacement uL measured from each local orientation
frame:
MüL + Cu̇L + KuL = RT F (29)
where R = [δij Ri ] is the block diagonal rotation matrix for all the nodes. Actually
there are two basic assumptions to guarantee the validity of this approximation (please
refer to [5] for the details and proofs). And we found that these assumptions can be
directly applied to our meshless setting without influencing its validity. Using modal
decomposition: uL (t) = ΨzL (t), the linear elastodynamic equation (29) for uL can be
reduced to a set of decoupled ODEs:

z̈L + Cz żL + Kz zL = ΨT (RT F). (30)

where Cz = (αI + βΛ) and Kz = Λ are both diagonal matrices. We solve the above
decoupled ODEs using implicit time integration. We take an approach similar to [1] by
making a first-order approximation of the total force at the next time step, to get the
following linear system:
½
∆z = h(ż0 + ∆ż)
∆ż = h(F0 − Kz (z0 + ∆z) − Cz (ż0 + ∆ż))

where h is the size of the time step, z0 and ż0 are the current modal amplitude and
velocity, and ∆z and ∆ż are their expected change in the next time step. By regrouping,
we obtain
Az ∆ż = bz (31)
where
Az = I + hCz + h2 Kz
and
bz = h (F0 − Kz z0 − (Cz + hKz )ż0 )
Note that Az is a diagonal matrix, which makes equation (31) to be solved efficiently.
 19

5.3.3 Manipulation constraints

In order for the users to interact with the simulated objects, position and orientation
constraints are important and must be enforced. In general, the MLS shape functions
lack the Kronecker delta function property and result in u(xI ) 6= uI . The position and
orientation constraints (Cp and Co , respectively) can be formulated as:

Cp (xc ) = Φ(xc )u(t) − dc (t) = 0

1
Co (xc ) = (∇×)Φ(xc )u(t) − wc (t) = 0
2
where xc is the constrained position of the object, dc (t) is the desired displacement,
and wc (t) is the desired orientation, which are known a priori. If we express both the
position and orientation constraints in terms of the modal amplitude z, they can be
simply written as:
C = Ac z − bc = 0 (32)
where Ac is a k × n constraint matrix (k is the number of constraints), and each row of
Ac represents a linear constraint on z, and the vector bc represents the values of these
constraints. The constraint condition (32) can be integrated into the system equation
(31) by Lagrange multipliers. In our implementation, we replace the constraint equation
C = 0 by the damped second-order equation C̈ + 2η Ċ + γ 2 C = 0, where η and γ are
stabilization factors [20]. So we can obtain the constrained equations of motion:
· ¸· ¸ · ¸
Az AT
c ∆ż bz
= . (33)
Ac 0 λh h(−2η Ċ − γ 2 C)

Since both the number of selected modes l and the number of constraints k are typically
small (l ≤ 128, k ≤ 20 in all of our examples), equation (33) could be solved in real-
time.

6 Experimental Results

The meshless simulation and rendering parts of our system are implemented on a
Windows XP PC with dual Intel Xeon 2.8GHz CPUs, 2.0GB RAM, and an nVidia
GeForce Fx 5900 Ultra GPU. The entire point-based rendering pipeline is built upon
Pointshop3D [32]. We have conducted extensive experiments on various point-sampled
surface data sets, for both solid and thin-shell simulations.
Table 1 shows the model statistics and performance data for our meshless solid
simulation based on the modal warping technique. For all the data sets in the elastic
solid simulation, the MLS pre-computation for the system matrices takes less than 10
minutes, while the modal decomposition takes less than 1 minute. Figure 11 shows the
deformation of the Santa Claus model based on users’ manipulation constraints on its
hands and feet. Figure 12 shows the deformation of the balljoint model under users’
manipulation with the bottom fixed.
Table 2 shows the performance data for our meshless thin-shell simulation based
on global conformal parameterization. We use the implicit time integration method
for the thin-shell simulation, which has been demonstrated to be very stable. The
initial numbers of nodes in the examples of gargoyle, bunny, rocker arm, and Iphigenie
simulations are only around 500. The elastic deformations of the gargoyle (Figure 13),
20

Table 1 The model statistics and performance data for the meshless elastic solid simulation
based on modal warping.

Model Points Nodes Modes Simulation

Beam 5,634 1,008 64 0.013 s/f
Balljoint 137,062 357 64 0.026 s/f
Rabbit 67,038 1,251 64 0.019 s/f
Santa 75,781 1,150 128 0.028 s/f

(a) (b) (c)

Fig. 11 A solid Santa model before (a) and after (b,c) elastic deformation.

(a) (b) (c)

Fig. 12 A solid balljoint model before (a) and after (b,c) elastic deformation.

bunny (Figure 14), and rocker arm (Figure 15) are simulated in real-time. Figure 16
shows the global conformal parameterization and explosion simulation of the Iphinenie
model. The computational load at each time step of the fracture propagation is much
higher than pure elastic deformation. We need to perform transparency tests for the
integration points, surface points, and simulation nodes in the neighborhood of new
crack lines. And new simulation nodes need to be inserted if necessary. We have to
update the mass and stiffness matrices at each time step to accomodate these changes.

7 Conclusion and Future Work

Large-scale point-sampled geometry is becoming ubiquitous in graphics, visualization,

and geometric information processing, due to the rapid advancement of the 3D scanning
 21

Table 2 The model statistics and performance data for the meshless elastic thin-shell simu-
lation based on global conformal parameterization.

Model Points Parameterization Nodes (Initial/Final) Simulation

Rockerarm 40,000 88.63 s 405/405 0.020 s/f
Gargoyle 40,002 62.09 s 520/520 0.026 s/f
Bunny 50,002 99.85 s 440/440 0.022 s/f
Iphigenie 200,219 866.45 s 420/650 1.2 s/f

(a) (b) (c) (d)

Fig. 13 A thin-shell gargoyle model before (a) and after (b-d) elastic deformation.

(a) (b) (c)

Fig. 14 A thin-shell bunny model before (a) and after (b,c) elastic deformation.

(a) (b) (c)

Fig. 15 A thin-shell rocker arm model before (a) and after (b,c) elastic deformation.

devices. Developing new computational techniques for point-centered digital process-

ing has become a common and long-term mission in computer graphics. In this paper,
we present our recent research results in meshless simulation of deformable shells and
solids. Our meshless framework is build upon continuum mechanics, which provide a
rigorous foundation to simulate the physical behavior of the modeled point geometry.
We presented several key computational techniques: (1) Moving Least Squares for func-
22

(a) (b) (c) (d)

Fig. 16 (a) The global conformal parameterization of the Iphigenie model. (b-d) The meshless
thin-shell fracture simulation based on the parameterization. The bottom parts show the crack
pattern in the parametric domain.

tional approximation and interpolation using discrete nodes; (2) hierarchical meshless
discretization for shells and solids; (3) Modal Warping for real-time simulation of large
deformations.

There are many avenues for possible future work, including exploring the theoretical
foundation of the meshless framework, designing efficient and accurate computational
algorithms, developing new functionalities, e.g. integrating haptic interaction between
human beings and virtual environments. Some of the most exciting applications of
HCI and haptic feedback have been found in surgical simulation, including a variety
of medical procedures, e.g. limb surgery, plastic surgery, gastrointestinal endoscopy,
etc. Traditionally mesh-based FEM is used to simulate real-time visual and haptic
rendering of soft tissues. A potential solution to the remeshing problems faced by the
FEM techniques is to use meshless methods articulated in this paper. Simulating surgi-
cal operations such as cutting and coagulation requires frequently updating the organ
mesh structure which is a very challenging process to be implemented robustly with-
out numerical singularities. Surgical cutting simulation is a simplified version of the
fracture simulation, since the cutting path is human-guided. However, surgical simu-
lations pose more critical time constraints since they need to provide timely sensory
information for the user in order to achieve immersive realism. A balance is needed to
ensure the simulation is fast enough for generating real-time feedbacks whilst the level
of realism is reasonably acceptable. Numerical methods, including the solvers used in
our meshless methods, can be very computationally expensive. Fortunately, many such
numerical techniques can be parallelized in order to accelerate the computation of rel-
evant quantities. In particular, the hierarchical structure (quadtree and octree) used in
our thin-shell and volumetric simulations can facilitate the parallelization by comput-
ing strains and other relevant quantities locally and by assembling the results globally.
Graphics processing unit (GPU) has seen fast developing in recent years. More and
more researchers have tried to use the GPU to solve general-purpose problems. Some
GPU-based mesh deformation algorithms have already been proposed [31]. However,
the GPU pipeline is specifically designed for mesh-based computations. Using the par-
 23

allel computing ability of the GPU to accelerate meshless simulation is an interesting

direction for the future work.

References

1. Baraff D, and Witkin A. Large steps in cloth simulation. Proceedings of SIGGRAPH’98,

pp. 43-54, 1998.
2. Belytschko T, Lu Y Y, and Gu L. Element free galerkin methods. International Journal for
Numerical Methods in Engineering, 37: 229-256, 1994.
3. Belytschko T, Lu Y Y, and Gu L. Fracture and crack growth by element-free galerkin
methods. Modeling Simulations for Materials Science and Engineering, 2: 519-534, 1994.
4. Belytschko T, Krongauz Y, Organ D, Fleming M, and Krysl P. Meshless methods: an
overview and recent developments. Computer Methods in Applied Mechanics and Engineer-
ing, 139: 3-47, 1996.
5. Choi M G, and Ko H S. Modal warping: real-time simulation of large rotational deformation
and manipulation. IEEE Transactions on Visual Computing and Graphics, 11(1): 91-101,
2005.
6. Cirak F, Ortiz M, and Schröder P. Subdivision surfaces: a new paradigm for thin-shell
finite element analysis. International Journal for Numerical Methods in Engineering, 47(12):
2039-2072, 2000.
7. DeRose T, Kass M, and Truong T. Subdivision surfaces in character animation. Proceedings
of SIGGRAPH’98, pp. 85-94, 1998.
8. Desbrun M, and Cani M P. Animating soft substances with implicit surfaces. Proceedings
of SIGGRAPH’95, pp. 287-290, 1995.
9. Farin G. Curves and Surfaces for CAGD. Morgan-Kaufmann, 5th edition, 2001.
10. Guo X, Hua J, and Qin H. Scalar-function-driven editing on point set surfaces. IEEE
Computer Graphics and Applications, 24(4): 43-52, 2004.
11. Guo X, and Qin H. Real-time meshless deformation. Computer Animation and Virtual
Worlds, 16(3-4):189-200, 2005.
12. Guo X, Li X, Bao Y, Gu X, and Qin H. Meshless thin-shell simulation based on global
conformal parameterization. IEEE Transactions on Visualization and Computer Graphics,
12(3): 375-385, 2006.
13. Hauser K, Shen C, and O’Brien J. Interactive deformation using modal analysis with
constraints. Proceedings of Graphics Interface, pp. 247-255, 2003.
14. James D, and Pai D. Dyrt: dynamic response textures for real time deformation simulation
with graphics hardware. Proceedings of SIGGRAPH’02, pp. 582-585, 2002.
15. Klaas O, and Shephard M. Automatic generation of octree-based three-dimensional dis-
cretizations for partition of unity methods. Computational Mechanics, 25: 296-304, 2000.
16. Lancaster P, and Salkauskas K. Curve and surface fitting: an introduction. Academic Press,
London, 1986.
17. Levoy M, Pulli K, Curless B, Rusinkiewicz S, Koller D, Pereira L, Ginzton M, Anderson
S, Davis J, Ginsberg J, Shade J, and Fulk D. The digital Michelangelo project: 3d scanning
of large statues. Proceedings of SIGGRAPH’00, pp. 131-144, 2000.
18. Li S, and Liu W K. Meshfree particle methods and their applications. Applied Mechanics
Review, 54: 1-34, 2002.
19. Malladi R, Sethian J A, and Vemuri B C. Shape modeling with front propagation: A level
set approach. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17: 158-175,
1995.
20. Metaxas D, and Terzopoulos D. Dynamic deformation of solid primitives with constraints.
Proceedings of SIGGRAPH’92, pp. 309-312, 1992.
21. Müller M, Keiser R, Nealen A, Pauly M, Gross M, and Alexa M. Point-based animation
of elastic, plastic, and melting objects. Proceedings of ACM SIGGRAPH/ Eurographics
Symposium on Computer Animation, 2004.
22. Müller M, Heidelberger B, Teschner M, and Gross M. Meshless deformations based on
shape matching. ACM Transactions on Graphics, 24(3):471-478, 2005
23. O’Brien J F, and Hodgins J K. Graphical modeling and animation of brittle fracture.
Proceedings of SIGGRAPH’99, pp. 137-146, 1999.
24. Ohtake Y, Belyaev A, Alexa M, Turk G, and Seidel H-P. Multi-level partition of unity
implicits. Proceedings of SIGGRAPH’03, pp. 463-470, 2003.
24

25. Organ D, Fleming M, Terry T, and Belytschko T. Continuous meshless approximations

for nonconvex bodies by diffraction and transparency. Computational Mechanics, 18: 1-11,
1996.
26. Pauly M, Keiser R, Kobbelt L, and Gross M. Shape modeling with point-sampled geometry.
Proceedings of SIGGRAPH’03, pp. 641-650, 2003.
27. Pauly M, Keiser R, Adams B, Dutré P, Gross M, and Guibas L. Meshless animation of
fracturing solids. Proceedings of SIGGRAPH’05, pp. 957-964, 2005.
28. Pentland A, and Williams J. Good vibrations: model dynamics for graphics and animation.
Proceedings of SIGGRAPH’89, pp. 215-222, 1989.
29. Wicke M, Steinemann D, and Gross M. Efficient animation of point-sampled thin shells.
Computer Graphics Forum, 24(3):667-676, 2005.
30. Xie H, Wang J, Hua J, Qin H, and Kaufman A. Piecewise C1 continuous surface recon-
struction of noisy point clouds via local implicit quadric regression. Proceedings of IEEE
Visualization’03, pp. 91-98, 2003.
31. Zhou K, Huang X, Xu W, Guo B, and Shum H Y. Direct manipulation of subdivision
surfaces on GPUs. ACM Transactions on Graphics, 26(3): 91, 2007.
32. Zwicker M, Pauly M, Knoll O, and Gross M. Pointshop3D: an interactive system for
point-based surface editing. ACM Transactions on Graphics, 21(3): 322-329, 2002.