Surface representations for 3D face recognition 273

Surface representations for 3D face recognition

Thomas Fabry, Dirk Smeets and Dirk Vandermeulen
0
Surface representations for 3D face recognition
Thomas Fabry
*
, Dirk Smeets
*
and Dirk Vandermeulen
Katholieke Universiteit Leuven
Belgium
1. Introduction
For long, face recognition has been a 2D discipline. However, 2D face recognition has shown
to be extremely difcult to be robust against a.o. lighting conditions and pose variations
(Phillips et al., 2003). At the same time, technological improvements are making 3D surface
capturing devices affordable for security purposes. As a result of these recent developments
face recognition shifts from 2D to 3D. This means that in the current state-of-the-art face recog-
nition systems the problem is no longer the comparison of 2D color photos, but the compari-
son of (textured) 3D surface shapes.
With the advent of the third dimension in face recognition, we think it is necessary to investi-
gate the known surface representations from this point of view. Throughout recent decades, a
lot of research focused on nding an appropriate digital representation for three dimensional
real-world objects, mostly for use in computer graphics (Hubeli & Gross, 2000; Sigg, 2006).
However, the needs for a surface representation in computer graphics, where the primary con-
cerns are visualization and the ability to process it on dedicated computer graphics hardware
(GPUs), are quite different from the needs of a surface representation for face recognition.
Another motivation for this work is the non-existence of an overview of 3D surface represen-
tations, altough the problem of object representation is studied since the birth of computer
vision (Marr, 1982).
With this in mind, we will, in this chapter, try to give an overview of surface representations
for use in biometric face recognition. Also surface representations that are not yet reported
in current face recognition literature, but we consider to be promising for future research
based on publications in related elds such as 3D object retrieval, computer vision, computer
graphics and 3D medical imaging will be discussed.
What are the desiderata for a surface representation in 3D face recognition? It is certainly use-
ful for a surface representation in biometric applications, to be accurate, usable for all sorts of
3D surfaces in face recognition (open, closed. . . ), concise (efcient in memory usage), easy to
acquire/construct, intuitive to work with, have a good formulation, be suitable for computa-
tions, convertible in other surface representations, ready to be efciently displayed and useful
for statistical modelling. It is nevertheless also certainly necessary to look further than a list of
desiderata. Herefore, our approach will be the following: we make a taxonomy of all surface
representations within the scope of 3D face recognition. For each of the of the representations
in this taxonomy, we will shortly describe the mathematical theory behind it. Advantages
and disadvantages of the surface representation will be stated. Related research using these
representations will be discussed and directions for future research will be indicated.
*
The rst two authors have equal contribution in this work .
15
Face Recognition 274
The structure of this chapter follows the taxonomy of Fig. 1. First we discuss the explicit
(meshfree) surfaces in section 2, followed by the implicit surfaces in section 3. We end with
some conclusions regarding surface representations for 3D face recognition.
Surface representations
Explicit Implicit
Point
cloud
Contour
and prole
curves
Mesh
Parametric
Spherical
harmonics
Information
Theoretic
Radial
Basis
Function
Blobby
model
Euclidean
distance
function
Random
Walk
Function
Fig. 1. Overview of surface representations
2. Explicit surface representations
In this section, several explicit surface representations are discussed. Strictly speaking, explicit
functions f (x) : R
m
R
n
are functions in which the n dependent variables can be written
explicitly in terms of the m independent variables. Simple shapes (spheres, ellipsoids,. . . ) can
be described by analytic functions. Unfortunately, it is mostly not possible to represent real
world objects by analytical surfaces. Therefore, this section mainly focusses on discretised
surface representations.
2.1 Point clouds
The point cloud is without doubt the simplest surface representation. It consists of an un-
ordered set of points that lie on the surface, in the 3D case an unordered set of x, y, and
z-coordinates. While a point cloud is not a real surface representation, but a (sometimes very)
sparse approximation of the surface at certain well-dened points, we consider this represen-
tation for a number of reasons. Firstly, point clouds are most often the output created by 3D
scanners. Secondly, the point cloud can be the base of most of the following surface repre-
sentations. Another reason to incorporate the point cloud in this document is its increased
popularity because of the ever-increasing memory capacities in todays computers. Earlier
the amount of stored information was to be minimized, so a minimal amount of points were
stored. As these point clouds were very sparse and as such a very coarse approximation of the
surface, they were then interpolated using, for instance, tensor-product splines. Today, mem-
ory shortage is of less concern, so more points can be stored, making point clouds approximate
other surface representations on ner and ner levels of detail.
Figure 2 gives an example of a 3D face represented as a point cloud, containing approximately
2000 points.
Fig. 2. An example of a 3D face represented as a point cloud.
One big advantage of the point cloud surface representation is the easy editing of point clouds:
because of the lack of a global connectivity graph or parameterization, point insertion, dele-
tion, repositioning,. . . is trivial. Another important advantage is the large amount of algo-
rithms developed for point clouds. A very popular method for 3D surface alignment, the
Iterative Closest Point (ICP) algorithm (Besl & McKay, 1992), uses only points on both sur-
faces and iterates between closest points search for correspondence nding and transforma-
tion calculation and application. Afterwards, many variants of the original algorithm were
developed (Rusinkiewicz & Levoy, 2001). The main drawback of point clouds is the incom-
pleteness of the surface description: only at some sparse point locations the surface is known.
Therefore, by representing a surface by a point cloud, a trade-off has to be made between ac-
curacy and amount of stored information. Also, rendering a set of points as a smooth surface
needs special processing, as explained in (Fabio, 2003).
The earliest use of point clouds was as a rendering primitive in (Levoy &Whitted, 1985). Point
clouds have also been used for shape and appearance modeling in e.g. (Kalaiah & Varshney,
2003; Pauly et al., 2003).
In 3D face recognition, point clouds are frequently used. Mostly for surface registration with
ICP (Alyz et al., 2008; Amberg et al., 2008; Chang et al., 2006; Faltemier et al., 2008; Kakadiaris
et al., 2007; Lu & Jain, 2006; Maurer et al., 2005; Russ et al., 2006; Wang et al., 2006). Bronstein
et al. (2005) represent faces by an expression-invariant canonical form, i.e. a point cloud where
Surface representations for 3D face recognition 275
The structure of this chapter follows the taxonomy of Fig. 1. First we discuss the explicit
(meshfree) surfaces in section 2, followed by the implicit surfaces in section 3. We end with
some conclusions regarding surface representations for 3D face recognition.
Surface representations
Explicit Implicit
Point
cloud
Contour
and prole
curves
Mesh
Parametric
Spherical
harmonics
Information
Theoretic
Radial
Basis
Function
Blobby
model
Euclidean
distance
function
Random
Walk
Function
Fig. 1. Overview of surface representations
2. Explicit surface representations
In this section, several explicit surface representations are discussed. Strictly speaking, explicit
functions f (x) : R
m
R
n
are functions in which the n dependent variables can be written
explicitly in terms of the m independent variables. Simple shapes (spheres, ellipsoids,. . . ) can
be described by analytic functions. Unfortunately, it is mostly not possible to represent real
world objects by analytical surfaces. Therefore, this section mainly focusses on discretised
surface representations.
2.1 Point clouds
The point cloud is without doubt the simplest surface representation. It consists of an un-
ordered set of points that lie on the surface, in the 3D case an unordered set of x, y, and
z-coordinates. While a point cloud is not a real surface representation, but a (sometimes very)
sparse approximation of the surface at certain well-dened points, we consider this represen-
tation for a number of reasons. Firstly, point clouds are most often the output created by 3D
scanners. Secondly, the point cloud can be the base of most of the following surface repre-
sentations. Another reason to incorporate the point cloud in this document is its increased
popularity because of the ever-increasing memory capacities in todays computers. Earlier
the amount of stored information was to be minimized, so a minimal amount of points were
stored. As these point clouds were very sparse and as such a very coarse approximation of the
surface, they were then interpolated using, for instance, tensor-product splines. Today, mem-
ory shortage is of less concern, so more points can be stored, making point clouds approximate
other surface representations on ner and ner levels of detail.
Figure 2 gives an example of a 3D face represented as a point cloud, containing approximately
2000 points.
Fig. 2. An example of a 3D face represented as a point cloud.
One big advantage of the point cloud surface representation is the easy editing of point clouds:
because of the lack of a global connectivity graph or parameterization, point insertion, dele-
tion, repositioning,. . . is trivial. Another important advantage is the large amount of algo-
rithms developed for point clouds. A very popular method for 3D surface alignment, the
Iterative Closest Point (ICP) algorithm (Besl & McKay, 1992), uses only points on both sur-
faces and iterates between closest points search for correspondence nding and transforma-
tion calculation and application. Afterwards, many variants of the original algorithm were
developed (Rusinkiewicz & Levoy, 2001). The main drawback of point clouds is the incom-
pleteness of the surface description: only at some sparse point locations the surface is known.
Therefore, by representing a surface by a point cloud, a trade-off has to be made between ac-
curacy and amount of stored information. Also, rendering a set of points as a smooth surface
needs special processing, as explained in (Fabio, 2003).
The earliest use of point clouds was as a rendering primitive in (Levoy &Whitted, 1985). Point
clouds have also been used for shape and appearance modeling in e.g. (Kalaiah & Varshney,
2003; Pauly et al., 2003).
In 3D face recognition, point clouds are frequently used. Mostly for surface registration with
ICP (Alyz et al., 2008; Amberg et al., 2008; Chang et al., 2006; Faltemier et al., 2008; Kakadiaris
et al., 2007; Lu & Jain, 2006; Maurer et al., 2005; Russ et al., 2006; Wang et al., 2006). Bronstein
et al. (2005) represent faces by an expression-invariant canonical form, i.e. a point cloud where
Face Recognition 276
point wise Euclidean distances approximately equal the point wise geodesic distances
1
in
the original face representation. Three-dimensional face shapes are often modeled with a
point cloud based statistical model, mostly a PCA model as in (Al-Osaimi et al., 2009; Amberg
et al., 2008; Lu & Jain, 2006; Russ et al., 2006). In Mpiperis et al. (2008), a bilinear model,
based on point clouds, is used to seperate intra-class and inter-class variations. In Fabry et al.
(2008), point clouds are treated in an information theoretic way, leading to probability density
function as surface representation, which is discussed in more detail in section 2.6.
2.2 Contour and prole curves
Contour and prole curves are also very sparse surface representations. They can even be
sparser than point clouds, but can also be made to approximate the surface as good as wanted.
The main idea is to represent shapes by the union of curves. The curves itself can be repre-
sented by a set of connected points or as a parametric curve.
Contour curves are closed, non intersecting curves on the surface, mostly of different length.
Depending on the extraction criterion, different types of contour curves can dened. Iso-
depth curves are obtained by translating a plane through the 3D face in one direction and
considering n different intersections of the plane and the object. n is the number of contours
that form the surface representation. Mostly the plane is positioned perpendicular to and
translated along the gaze direction, which is hereby dened as the z-axis. Then iso-depth
curves have equal z-values. Iso-radius curves are contours, obtained as an intersection of the
object with a cylinder with radius r =

x
2
+y
2
, or as an intersection with a sphere with
radius r =

x
2
+y
2
+z
2
, with the z-axis parallel to the gaze direction and the y-axis parallel
to the longitudinal axis of the face. An iso-geodesic curve, or iso-geodesic, is a contour with
each part of the curve on an equal geodesic distance to a reference point, i.e. the distance of
the shortest path on the full surface between the part of the curve and the reference point. The
calculation of geodesic distances is mostly done using a polygon mesh (see section 2.3).
Examples of iso-depth, iso-radius and iso-geodesic curves are given in Fig. 3. Only points
lying on those curves are shown.
(a) (b) (c) (d)
Fig. 3. Points lying on iso-depth curves (a), on iso-radius curves obtained by intersection with
a cylinder (b) or with a sphere (c) and iso-geodesics with respect to the nose tip (d).
Prole curves on the contrary have a starting and an end point. For 3D faces, the starting point
is most frequently a point in the middle of the face, mostly the nose tip, while the end point is
often at the edge of the face. There exist an innite number of prole curves in between those
points. Figure 4(a) shows an example with each part on the curve having the same angle with
respect to a line in the xy-plane through the central point (nose tip). Figure 4(b) shows points
1
The geodesic distance is the length of the shortest path on the object surface between two points on the
object.
on lines with equal x-value, again with the z-axis parallel to the gaze direction and the y-axis
parallel to the longitudinal axis.
(a) (b)
Fig. 4. Points on prole curves with curve parts under the same angle (a) or with the same
x-value (b).
Curves are non-complete surface representations, implying that the surface is only dened on
the curves. On the one hand, this implies a loss of information, on the other hand lower stor-
age requirements. In order to construct contour curves, a reference point is needed. In 3D face
recognition mostly the nose is used, which infers the manual or automatic extraction of this
landmark. For extraction of iso-depth and cylinder based iso-radius curves and most types
of prole curves, even more information is required: the gaze direction and/or longitudinal
axis of the face. When done this, it is more easy to set correspondences between faces based
on corresponding curves.
Contour and prole curves are frequently used in 3D face recognition. Iso-geodesics are pop-
ular because of their lower sensitivity to expression variation, based on the hypothesis that
expression-induced surface variations can approximately be modeled by isometric transfor-
mations. Those transformations keep geodesic distances between every point pair on the sur-
face. Berretti et al. (2006) use the spatial relationship between the intra-subject iso-geodesics as
a subject specic shape descriptor. Feng et al. (2007) divide the iso-geodesics in segments that
form the basis of trained face signatures. Mpiperis et al. (2007) map the iso-geodesic curves to
concentric circles on a plane using a piecewise linear warping transformation. Jahanbin et al.
(2008) extract ve shape descriptors from each iso-geodesic: convexity, ratio of principal axes,
compactness, circular and elliptic variance. These features are trained with Linear Discrim-
inant Analysis (LDA). In Li et al. (2008), LDA is also used for training of texture intensities
sampled at xed angles on iso-geodesic curves. Pears & Heseltine (2006) use sphere based
iso-radius curves. Due to the innite rotational symmetry of a sphere, the representation is
invariant to pose variations. Using this representation, registration can be implemented using
a simple process of 1D correlation resulting in a registration of a comparable accuracy to ICP,
but fast, non iterative, and robust to the presence of outliers. Samir et al. (2006) compare faces
using dissimilarity measures extracted from the distances between iso-depth curves. Jahan-
bin et al. (2008) use iso-depth curves in the same way as they use iso-geodesics. Prole curves
are used by ter Haar & Veltkamp (2008), where comparison is done using the weighted dis-
tance between corresponding sample points on the curves, and in Feng et al. (2006) where the
curves are used similar as in Feng et al. (2007) (see above).
2.3 Polygon meshes
In 3D research in general, and a fortiori in 3D face recognition, the vast majority of researchers
represent 3D object surfaces as meshes. A mesh is in essence an unordered set of vertices
(points), edges (connection between two vertices) and faces (closed set of edges) that together
Surface representations for 3D face recognition 277
point wise Euclidean distances approximately equal the point wise geodesic distances
1
in
the original face representation. Three-dimensional face shapes are often modeled with a
point cloud based statistical model, mostly a PCA model as in (Al-Osaimi et al., 2009; Amberg
et al., 2008; Lu & Jain, 2006; Russ et al., 2006). In Mpiperis et al. (2008), a bilinear model,
based on point clouds, is used to seperate intra-class and inter-class variations. In Fabry et al.
(2008), point clouds are treated in an information theoretic way, leading to probability density
function as surface representation, which is discussed in more detail in section 2.6.
2.2 Contour and prole curves
Contour and prole curves are also very sparse surface representations. They can even be
sparser than point clouds, but can also be made to approximate the surface as good as wanted.
The main idea is to represent shapes by the union of curves. The curves itself can be repre-
sented by a set of connected points or as a parametric curve.
Contour curves are closed, non intersecting curves on the surface, mostly of different length.
Depending on the extraction criterion, different types of contour curves can dened. Iso-
depth curves are obtained by translating a plane through the 3D face in one direction and
considering n different intersections of the plane and the object. n is the number of contours
that form the surface representation. Mostly the plane is positioned perpendicular to and
translated along the gaze direction, which is hereby dened as the z-axis. Then iso-depth
curves have equal z-values. Iso-radius curves are contours, obtained as an intersection of the
object with a cylinder with radius r =

x
2
+y
2
, or as an intersection with a sphere with
radius r =

x
2
+y
2
+z
2
, with the z-axis parallel to the gaze direction and the y-axis parallel
to the longitudinal axis of the face. An iso-geodesic curve, or iso-geodesic, is a contour with
each part of the curve on an equal geodesic distance to a reference point, i.e. the distance of
the shortest path on the full surface between the part of the curve and the reference point. The
calculation of geodesic distances is mostly done using a polygon mesh (see section 2.3).
Examples of iso-depth, iso-radius and iso-geodesic curves are given in Fig. 3. Only points
lying on those curves are shown.
(a) (b) (c) (d)
Fig. 3. Points lying on iso-depth curves (a), on iso-radius curves obtained by intersection with
a cylinder (b) or with a sphere (c) and iso-geodesics with respect to the nose tip (d).
Prole curves on the contrary have a starting and an end point. For 3D faces, the starting point
is most frequently a point in the middle of the face, mostly the nose tip, while the end point is
often at the edge of the face. There exist an innite number of prole curves in between those
points. Figure 4(a) shows an example with each part on the curve having the same angle with
respect to a line in the xy-plane through the central point (nose tip). Figure 4(b) shows points
1
The geodesic distance is the length of the shortest path on the object surface between two points on the
object.
on lines with equal x-value, again with the z-axis parallel to the gaze direction and the y-axis
parallel to the longitudinal axis.
(a) (b)
Fig. 4. Points on prole curves with curve parts under the same angle (a) or with the same
x-value (b).
Curves are non-complete surface representations, implying that the surface is only dened on
the curves. On the one hand, this implies a loss of information, on the other hand lower stor-
age requirements. In order to construct contour curves, a reference point is needed. In 3D face
recognition mostly the nose is used, which infers the manual or automatic extraction of this
landmark. For extraction of iso-depth and cylinder based iso-radius curves and most types
of prole curves, even more information is required: the gaze direction and/or longitudinal
axis of the face. When done this, it is more easy to set correspondences between faces based
on corresponding curves.
Contour and prole curves are frequently used in 3D face recognition. Iso-geodesics are pop-
ular because of their lower sensitivity to expression variation, based on the hypothesis that
expression-induced surface variations can approximately be modeled by isometric transfor-
mations. Those transformations keep geodesic distances between every point pair on the sur-
face. Berretti et al. (2006) use the spatial relationship between the intra-subject iso-geodesics as
a subject specic shape descriptor. Feng et al. (2007) divide the iso-geodesics in segments that
form the basis of trained face signatures. Mpiperis et al. (2007) map the iso-geodesic curves to
concentric circles on a plane using a piecewise linear warping transformation. Jahanbin et al.
(2008) extract ve shape descriptors from each iso-geodesic: convexity, ratio of principal axes,
compactness, circular and elliptic variance. These features are trained with Linear Discrim-
inant Analysis (LDA). In Li et al. (2008), LDA is also used for training of texture intensities
sampled at xed angles on iso-geodesic curves. Pears & Heseltine (2006) use sphere based
iso-radius curves. Due to the innite rotational symmetry of a sphere, the representation is
invariant to pose variations. Using this representation, registration can be implemented using
a simple process of 1D correlation resulting in a registration of a comparable accuracy to ICP,
but fast, non iterative, and robust to the presence of outliers. Samir et al. (2006) compare faces
using dissimilarity measures extracted from the distances between iso-depth curves. Jahan-
bin et al. (2008) use iso-depth curves in the same way as they use iso-geodesics. Prole curves
are used by ter Haar & Veltkamp (2008), where comparison is done using the weighted dis-
tance between corresponding sample points on the curves, and in Feng et al. (2006) where the
curves are used similar as in Feng et al. (2007) (see above).
2.3 Polygon meshes
In 3D research in general, and a fortiori in 3D face recognition, the vast majority of researchers
represent 3D object surfaces as meshes. A mesh is in essence an unordered set of vertices
(points), edges (connection between two vertices) and faces (closed set of edges) that together
Face Recognition 278
represent the surface explicitly. Mostly, the faces consist of triangles, quadrilaterals or other
simple convex polygons, since this simplies rendering. Figure 5 shows the triangular mesh
corresponding with the point cloud of Fig. 2.
The problem of constructing a mesh given a point cloud is commonly called the surface re-
construction problem, although this might also incorporate reconstruction of other complete
surface representations. The most powerful algorithm to deal with mesh construction given a
point cloud is the power crust algorithm, described in Amenta et al. (2001). Other algorithms
that deal with this problem are, a.o., the algorithms of Hoppe et al. (1992), Edelsbrunner &
Mcke (1994) and Curless & Levoy (1996). A short overview of methods for triangulation can
be found in the paper of Varshosaz et al. (2005).
Fig. 5. An example of a 3D face represented as a mesh.
Probably the main benet of a polygon mesh is its ease of visualization. Many algorithms for
ray tracing, collision detection, and rigid-body dynamics are developed for polygon meshes.
Another advantage of meshes, certainly in comparison to point clouds, is the explicit knowl-
edge of connectivity, which is useful for the computation of the geodesic distance between
two points. This is particularly useful in face recognition because geodesic distances between
points on the surface are often used in 3D expression-invariant face recognition, because they
seem to vary less than Euclidean distances. The use of this was introduced by Bronstein et al.
(2003), using a fast marching method for triangulated domains (Kimmel & Sethian, 1998) for
geodesic distance calculation. Afterwards, many other researchers used the concept of in-
variance of geodesic distances during expression variations, by directly comparing point wise
distances (Gupta et al., 2007; Li &Zhang, 2007; Smeets et al., 2009) or using iso-geodesic curves
(see section 2.2). On the other hand, mesh errors like cracks, holes, T-joints, overlapping poly-
gons, dupplicated geometry, self intersections and inconsistent normal orientation can occur
as described by Veleba & Felkel (2007).
2.4 Parametric surface representations
A generic parametric form for representing 3D surfaces is a function with domain R
2
and
range R
3
(Campbell & Flynn, 2000):
S(u, v) =

x = f
1
(u, v)
y = f
2
(u, v)
z = f
3
(u, v)
(1)
where u and v are the two parametric variables.
Amongst the advantages of parametric representations in general we can count the simple but
general and complete mathematical description, the easiness to handle and the readiness of
technology to visualise these representations. General disadvantages are that only functions
can be represented, which cause problems in, for instance, the ear and nose regions.
The class of parametric surface representations is very general, but the following three deserve
particular attention.
2.4.1 Height maps
A height map is a special form of a parametric surface with x = u and y = v and is also often
referred as depth map, range image or graph surface. A height map represents the height
of points along the z-directions in a regular sampling of the x, y image axes in a matrix. An
example of a face represented as a depth map can be seen in gure 6. A big advantage of this
representation is that many 3D laser scanners produce this kind of output. Because mostly the
x and y values lay on a regular grid, the surface can be descibed by a matrix and 2D image
processing techniques can be applied on it. The most prominent disadvantage of height maps
is the limited expressional power: only what is seen when looking from one direction (with
parallel beams) can be represented.This causes problems in the representation of, for instance,
the cheeks and ears of human faces. The use of height maps in face recognition has already
been discussed by Akarun et al. (2005). One very specic example of a 3D face recognition
method using height maps is the method of Samir et al. (2006), because here height maps are
extracted fromtriangulated meshes in order to represent the surface by level curves, which are
iso-contours of the depth map. Colbry & Stockman (2007) extend the denition of depth map
leading to the canonical face depth map. This is obtained by translating a parabolic cylinder
or a quadratic, instead of a plane, along the z-direction. Because of this alternative denition,
it could also belong to section 2.4.2.
Fig. 6. An example of a height map surface representation.
Surface representations for 3D face recognition 279
represent the surface explicitly. Mostly, the faces consist of triangles, quadrilaterals or other
simple convex polygons, since this simplies rendering. Figure 5 shows the triangular mesh
corresponding with the point cloud of Fig. 2.
The problem of constructing a mesh given a point cloud is commonly called the surface re-
construction problem, although this might also incorporate reconstruction of other complete
surface representations. The most powerful algorithm to deal with mesh construction given a
point cloud is the power crust algorithm, described in Amenta et al. (2001). Other algorithms
that deal with this problem are, a.o., the algorithms of Hoppe et al. (1992), Edelsbrunner &
Mcke (1994) and Curless & Levoy (1996). A short overview of methods for triangulation can
be found in the paper of Varshosaz et al. (2005).
Fig. 5. An example of a 3D face represented as a mesh.
Probably the main benet of a polygon mesh is its ease of visualization. Many algorithms for
ray tracing, collision detection, and rigid-body dynamics are developed for polygon meshes.
Another advantage of meshes, certainly in comparison to point clouds, is the explicit knowl-
edge of connectivity, which is useful for the computation of the geodesic distance between
two points. This is particularly useful in face recognition because geodesic distances between
points on the surface are often used in 3D expression-invariant face recognition, because they
seem to vary less than Euclidean distances. The use of this was introduced by Bronstein et al.
(2003), using a fast marching method for triangulated domains (Kimmel & Sethian, 1998) for
geodesic distance calculation. Afterwards, many other researchers used the concept of in-
variance of geodesic distances during expression variations, by directly comparing point wise
distances (Gupta et al., 2007; Li &Zhang, 2007; Smeets et al., 2009) or using iso-geodesic curves
(see section 2.2). On the other hand, mesh errors like cracks, holes, T-joints, overlapping poly-
gons, dupplicated geometry, self intersections and inconsistent normal orientation can occur
as described by Veleba & Felkel (2007).
2.4 Parametric surface representations
A generic parametric form for representing 3D surfaces is a function with domain R
2
and
range R
3
(Campbell & Flynn, 2000):
S(u, v) =

x = f
1
(u, v)
y = f
2
(u, v)
z = f
3
(u, v)
(1)
where u and v are the two parametric variables.
Amongst the advantages of parametric representations in general we can count the simple but
general and complete mathematical description, the easiness to handle and the readiness of
technology to visualise these representations. General disadvantages are that only functions
can be represented, which cause problems in, for instance, the ear and nose regions.
The class of parametric surface representations is very general, but the following three deserve
particular attention.
2.4.1 Height maps
A height map is a special form of a parametric surface with x = u and y = v and is also often
referred as depth map, range image or graph surface. A height map represents the height
of points along the z-directions in a regular sampling of the x, y image axes in a matrix. An
example of a face represented as a depth map can be seen in gure 6. A big advantage of this
representation is that many 3D laser scanners produce this kind of output. Because mostly the
x and y values lay on a regular grid, the surface can be descibed by a matrix and 2D image
processing techniques can be applied on it. The most prominent disadvantage of height maps
is the limited expressional power: only what is seen when looking from one direction (with
parallel beams) can be represented.This causes problems in the representation of, for instance,
the cheeks and ears of human faces. The use of height maps in face recognition has already
been discussed by Akarun et al. (2005). One very specic example of a 3D face recognition
method using height maps is the method of Samir et al. (2006), because here height maps are
extracted fromtriangulated meshes in order to represent the surface by level curves, which are
iso-contours of the depth map. Colbry & Stockman (2007) extend the denition of depth map
leading to the canonical face depth map. This is obtained by translating a parabolic cylinder
or a quadratic, instead of a plane, along the z-direction. Because of this alternative denition,
it could also belong to section 2.4.2.
Fig. 6. An example of a height map surface representation.
Face Recognition 280
2.4.2 Geometry images
Another often used parametric surface representation is the geometry image, a regularly sam-
pled 2D grid representation, but here not representing the distance to the surface along a
viewing direction. The directions are adaptive in the sense that they are conveyed so to be
able to represent the whole surface, thus also regions that would not be representable in a
height map due to the directionality of height. Gu et al. (2002) describe an automatic sys-
tem for converting arbitrary meshes into geometry images: the basic idea is to slice open the
mesh along an appropriate selection of cut paths in order to unfold the mesh. Next the cut
surface is parametrized onto a square domain containing this opened mesh creating an n n
matrix of [x, y, z] data values. The geometry image representation has some advantages, the
biggest being that the irregular surface is represented on a completely regular grid, without
loosing information. As with the height map, this structure is easy to process, both for graph-
ics applications and for recognition applications. A big disadvantage is the computational
and technical complexity of the generation of geometry images.
Geometry images have already been used for 3D face recognition. In (Kakadiaris et al., 2007;
Perakis et al., 2009), a geometry image maps all vertices of the face model surface from R
3
to R
2
. This representation is segmented to form the Annotated Face Model (AFM) which is
rigidly aligned with the range image of a probe. Afterwards, the AFM is tted to each probe
data set using the elastically adapted deformable model framework, described by Metaxas
& Kakadiaris (2002). The deformed model is then again converted to a geometry image and
a normal map. Those images are analysed using the Haar and pyramid wavelet transform.
The distance metric that is used to compare different faces, uses the coefcients of the wavelet
transforms.
2.4.3 Splines
Spline surfaces are piecewise polynomial parametric surface representations that are very
popular in the eld of Computer Aided Design and Modelling (CAD/CAM) because of
the simplicity of their construction, including interpolation and approximation of complex
shapes, and their ease and accuracy of evaluation. The basis of splines are control points,
which are mostly lying on a regular grid.
One well-known type of spline surfaces are Bzier surfaces. These are represented as:
S(u, v) =
n

i=0
m

j=0
B
n
i
(u) B
m
j
(v) x
i,j
(2)
where B
n
i
are Bernstein polynomials. Another well-known spline surface type are nonuniform
rational B-splines (NURBS), dened as
S(u, v) =
k

i=1
N
i,n
w
i

k
j=1
N
j,n
w
j
x
i
, (3)
with k the number of control points x
i
and w
i
the corresponding weights. An example of face
representation by Bzier curves can be found in (Yang et al., 2006) and of NURBS in (Liu et al.,
2005; Yano & Harada, 2007). Although spline surfaces have been considered for representing
faces, they have not found widespread use in 3Dface recognition. Most probably because, as is
stated by Besl (1990), a.o. : it is difcult to make a surface dened on a parametric rectangle t
an arbitrary region on the surface. Also, the control points are not easily detectable. Another
disadvantage of splines is the uselessness of the spline parameters for recognition.
2.5 Spherical harmonics surface representations
Spherical harmonics are mathematical functions that can be used for representing spherical
objects (sometimes also called star-shaped objects). The surface rst has to be represented
as a function on the unit sphere: f (, ). This function can then be decomposed (spherically
expanded) as:
f (, ) =

l=0
l

m=l
a
lm
Y
m
l
(, ). (4)
The spherical harmonics Y
m
l
(, ) : |m| Nare dened on the unit sphere as:
Y
m
l
(, ) = k
l,m
P
m
l
(cos )e
im
, (5)
where [0, ], [0, 2[, k
l,m
is a constant, and P
m
l
is the associated Legendre polynomial.
The cofcients a
lm
are uniquely dened by:
a
lm
=

2
0

pi
0
f (, )Y
m
l
(, )sin()dd. (6)
While this theory is stated for spherical objects, it is important to mention that spherical har-
monics have already been used for non spherical objects by rst decomposing the object into
spherical subparts (Mousa et al., 2007) using the volumetric segmentation method proposed
by Dey et al. (2003).
Advantages of the spherical harmonics surface representation include the similarity to the
Fourier transform, which has already proven to be a very interesting technique in 1D signal
processing. Also, because of the spectral nature of this surface representation, it can lead
to large dimensionality reductions, leading to decreases in computation time and efcient
storage. Other advantages are the rotational invariance of the representation and the ability
to cope with missing data (occlusions and partial views).
The spherical harmonics surface representation has some drawbacks as well. One of them
has already been mentioned and is not insuperable: the need for spherical surfaces. Other
disadvantages include the unsuitability for intuitive editing, the non-trivial visualisation and
the global nature of the representation.
Spherical harmonics surface representations have already been used in a number of applica-
tions that bear a close relation to face recognition. Kazhdan et al. (2003) used it as a 3D shape
descriptor for use in searching a 3D model database. Mousa et al. (2007) use the spherical
harmonics for reconstruction of 3D objects from point sets, local mesh smoothing and texture
transfer. Dillenseger et al. (2006) have used it for 3D kidney modelling and registration. To
the best of our knowledge, the only use of this representation in face recognition so far is in
(Llonch et al., 2009). In this work, a similar transformation with another (overcomplete) basis
is used as a surface representation for the 3D faces in the face recognition experiments, where
this representation is also submitted to linear discriminant analysis (LDA). The performance
reported is better than PCA on depth maps, which the authors consider as baseline.
Further applications in face recognition are not yet widely explored, but we see a great poten-
tial in the method of Blow & Daniilidis (2001), who combine the spherical harmonics repre-
sentation with Gabor wavelets on the sphere. In this way, the main structure of the 3D face is
represented globally, while the (person-specic) details are modelled locally (with wavelets).
This solves the drawback of the global nature of the representation and could as such be used
for multiscale progressive 3D face recognition.
Surface representations for 3D face recognition 281
2.4.2 Geometry images
Another often used parametric surface representation is the geometry image, a regularly sam-
pled 2D grid representation, but here not representing the distance to the surface along a
viewing direction. The directions are adaptive in the sense that they are conveyed so to be
able to represent the whole surface, thus also regions that would not be representable in a
height map due to the directionality of height. Gu et al. (2002) describe an automatic sys-
tem for converting arbitrary meshes into geometry images: the basic idea is to slice open the
mesh along an appropriate selection of cut paths in order to unfold the mesh. Next the cut
surface is parametrized onto a square domain containing this opened mesh creating an n n
matrix of [x, y, z] data values. The geometry image representation has some advantages, the
biggest being that the irregular surface is represented on a completely regular grid, without
loosing information. As with the height map, this structure is easy to process, both for graph-
ics applications and for recognition applications. A big disadvantage is the computational
and technical complexity of the generation of geometry images.
Geometry images have already been used for 3D face recognition. In (Kakadiaris et al., 2007;
Perakis et al., 2009), a geometry image maps all vertices of the face model surface from R
3
to R
2
. This representation is segmented to form the Annotated Face Model (AFM) which is
rigidly aligned with the range image of a probe. Afterwards, the AFM is tted to each probe
data set using the elastically adapted deformable model framework, described by Metaxas
& Kakadiaris (2002). The deformed model is then again converted to a geometry image and
a normal map. Those images are analysed using the Haar and pyramid wavelet transform.
The distance metric that is used to compare different faces, uses the coefcients of the wavelet
transforms.
2.4.3 Splines
Spline surfaces are piecewise polynomial parametric surface representations that are very
popular in the eld of Computer Aided Design and Modelling (CAD/CAM) because of
the simplicity of their construction, including interpolation and approximation of complex
shapes, and their ease and accuracy of evaluation. The basis of splines are control points,
which are mostly lying on a regular grid.
One well-known type of spline surfaces are Bzier surfaces. These are represented as:
S(u, v) =
n

i=0
m

j=0
B
n
i
(u) B
m
j
(v) x
i,j
(2)
where B
n
i
are Bernstein polynomials. Another well-known spline surface type are nonuniform
rational B-splines (NURBS), dened as
S(u, v) =
k

i=1
N
i,n
w
i

k
j=1
N
j,n
w
j
x
i
, (3)
with k the number of control points x
i
and w
i
the corresponding weights. An example of face
representation by Bzier curves can be found in (Yang et al., 2006) and of NURBS in (Liu et al.,
2005; Yano & Harada, 2007). Although spline surfaces have been considered for representing
faces, they have not found widespread use in 3Dface recognition. Most probably because, as is
stated by Besl (1990), a.o. : it is difcult to make a surface dened on a parametric rectangle t
an arbitrary region on the surface. Also, the control points are not easily detectable. Another
disadvantage of splines is the uselessness of the spline parameters for recognition.
2.5 Spherical harmonics surface representations
Spherical harmonics are mathematical functions that can be used for representing spherical
objects (sometimes also called star-shaped objects). The surface rst has to be represented
as a function on the unit sphere: f (, ). This function can then be decomposed (spherically
expanded) as:
f (, ) =

l=0
l

m=l
a
lm
Y
m
l
(, ). (4)
The spherical harmonics Y
m
l
(, ) : |m| Nare dened on the unit sphere as:
Y
m
l
(, ) = k
l,m
P
m
l
(cos )e
im
, (5)
where [0, ], [0, 2[, k
l,m
is a constant, and P
m
l
is the associated Legendre polynomial.
The cofcients a
lm
are uniquely dened by:
a
lm
=

2
0

pi
0
f (, )Y
m
l
(, )sin()dd. (6)
While this theory is stated for spherical objects, it is important to mention that spherical har-
monics have already been used for non spherical objects by rst decomposing the object into
spherical subparts (Mousa et al., 2007) using the volumetric segmentation method proposed
by Dey et al. (2003).
Advantages of the spherical harmonics surface representation include the similarity to the
Fourier transform, which has already proven to be a very interesting technique in 1D signal
processing. Also, because of the spectral nature of this surface representation, it can lead
to large dimensionality reductions, leading to decreases in computation time and efcient
storage. Other advantages are the rotational invariance of the representation and the ability
to cope with missing data (occlusions and partial views).
The spherical harmonics surface representation has some drawbacks as well. One of them
has already been mentioned and is not insuperable: the need for spherical surfaces. Other
disadvantages include the unsuitability for intuitive editing, the non-trivial visualisation and
the global nature of the representation.
Spherical harmonics surface representations have already been used in a number of applica-
tions that bear a close relation to face recognition. Kazhdan et al. (2003) used it as a 3D shape
descriptor for use in searching a 3D model database. Mousa et al. (2007) use the spherical
harmonics for reconstruction of 3D objects from point sets, local mesh smoothing and texture
transfer. Dillenseger et al. (2006) have used it for 3D kidney modelling and registration. To
the best of our knowledge, the only use of this representation in face recognition so far is in
(Llonch et al., 2009). In this work, a similar transformation with another (overcomplete) basis
is used as a surface representation for the 3D faces in the face recognition experiments, where
this representation is also submitted to linear discriminant analysis (LDA). The performance
reported is better than PCA on depth maps, which the authors consider as baseline.
Further applications in face recognition are not yet widely explored, but we see a great poten-
tial in the method of Blow & Daniilidis (2001), who combine the spherical harmonics repre-
sentation with Gabor wavelets on the sphere. In this way, the main structure of the 3D face is
represented globally, while the (person-specic) details are modelled locally (with wavelets).
This solves the drawback of the global nature of the representation and could as such be used
for multiscale progressive 3D face recognition.
Face Recognition 282
2.6 Point cloud-based surface representations: information theoretic measures
Recently, some researchers have proposed to work directly on the point cloud, without using
any real surface representation, but instead use information theoretic measures dened di-
rectly on the raw point cloud surface representation. Some of these methods do nevertheless
implicitly use some kind of kernel surface representation, which can also be viewed as density
estimation (Silverman, 1986), and although the density estimation itself is explicit, the surface
can be thought of as implicitly present. This is also the reason why this surface reprsentation
was not included in section 2.1 (which would also make sense) but is treated as a link be-
tween the explicit and implicit surface representations. Density estimation is a fundamental
concept in statistics, the search for an estimate of the density from a given dataset. In the case
of surface representations, this dataset is a point cloud. This estimation can be nonparametric,
which can be considered as an advantage of this method. Also the generality, sound statisti-
cal base and low data requirements are advantages. Disadvantages include the difculties in
visualising the surface representation,
The most used density estimation technique is kernel density estimation KDE, introduced by
Parzen (1962). Here the density is computed as

f
h
(x) =
1
nh
n

i=1
K

x x
i
h

(7)
where K is some kernel with parameter h. A volume rendering of a KDE of a 3D face surface
can be seen in gure 7.
Fig. 7. Volume rendering of a kernel density estimation of a human face.
The information theoretic methods have already been proven to be useful in 3D registration
and recognition. Tsin & Kanade (2004) proposed the kernel correlation of two point clouds,
an entropy-related measure expressing the compatibility of two point clouds, and used this
for robust 3D registration. This measure has later been used in 3D face recognition by Fabry
et al. (2008). A related technique, which has until now not been applied to face recognition, is
found in (Wang et al., 2008). Here, groupwise registration between point clouds is performed
by minimizing the Jensen-Shannon divergence between the Gaussian mixture representations
of the point clouds.
3. Implicit surface representations
In general, implicit functions are dened as the iso-level of a scalar function : R
n
R. A
3D implicit surface S is then mathematically dened as
S = {x R
3
|(x) = }. (8)
We call this the iso-surface of the implicit function. The iso-surface at = 0 is sometimes
referred to as the zero contour or zero surface. As such, implicit surfaces are 2Dgeometric shapes
that exist in 3D space (Bloomenthal & Wyvill, 1997). The iso-surface partitions the space into
two regions: interior of the surface, and exterior of the surface. Mostly, the convention is
followed that inside the surface, the function returns negative values and outside the surface,
the function returns positive values. The inside portion is referred as

, while points with

positive values belong to the outside portion
+
. The border between the inside and the
outside is called the interface .
The simplest surfaces (spheroids, ellipsoids,. . . ) can be described by analytic functions and
are called algebraic surfaces. The surface is the set of roots of a polynomial (x) = . The
degree of the surface n is the maximum sum of powers of all terms. The general form of a
linear surface (n = 1), or plane, is
(x, y, z) = ax + by + cz d = , (9)
while the general form for a quadratic surface (n = 2) is:
(x, y, z)
= ax
2
+ bxy + cxz + dx + ey
2
+ f yz + gy + hz
2
+ iz + j
= .
(10)
Superquadrics (n > 2) provide more exibility by adding parameters to control the polyno-
mial exponent, allowing to describe more complex surfaces. Nevertheless, analytic functions
are designed to describe a surface globally by a single closed formula. In reality, it is mostly
not possible to represent a whole real-life object by an analytic function of this form.
3.1 Radial Basis Functions
Radial Basis Functions (RBFs) are another type of implicit functions that have been proven
to be a powerful tool in interpolating scattered data of all kinds, including 3D point clouds
representing 3D objects. A RBF is a function of the form
S(x) =
N

i=1

i
(x x
i
) + p(x), (11)
with
i
the RBF-coefcients, a radial basic function, x
i
the RBF centra and p(x) a polynomial of
low degree.
As can be seen from equation (11), the RBF consists of a weighted sum of radially symmet-
ric basic functions located at the RBF-centra x
i
and a low degree polynomial p. For surface
representation, the RBF-centra x
i
are simply a subset of points on the surface. Finding the
appropriate RBF-coefcients for implicitly representing a surface is done by solving:
x
i
: s(x
i
) = f
i
, (12)
Surface representations for 3D face recognition 283
2.6 Point cloud-based surface representations: information theoretic measures
Recently, some researchers have proposed to work directly on the point cloud, without using
any real surface representation, but instead use information theoretic measures dened di-
rectly on the raw point cloud surface representation. Some of these methods do nevertheless
implicitly use some kind of kernel surface representation, which can also be viewed as density
estimation (Silverman, 1986), and although the density estimation itself is explicit, the surface
can be thought of as implicitly present. This is also the reason why this surface reprsentation
was not included in section 2.1 (which would also make sense) but is treated as a link be-
tween the explicit and implicit surface representations. Density estimation is a fundamental
concept in statistics, the search for an estimate of the density from a given dataset. In the case
of surface representations, this dataset is a point cloud. This estimation can be nonparametric,
which can be considered as an advantage of this method. Also the generality, sound statisti-
cal base and low data requirements are advantages. Disadvantages include the difculties in
visualising the surface representation,
The most used density estimation technique is kernel density estimation KDE, introduced by
Parzen (1962). Here the density is computed as

f
h
(x) =
1
nh
n

i=1
K

x x
i
h

(7)
where K is some kernel with parameter h. A volume rendering of a KDE of a 3D face surface
can be seen in gure 7.
Fig. 7. Volume rendering of a kernel density estimation of a human face.
The information theoretic methods have already been proven to be useful in 3D registration
and recognition. Tsin & Kanade (2004) proposed the kernel correlation of two point clouds,
an entropy-related measure expressing the compatibility of two point clouds, and used this
for robust 3D registration. This measure has later been used in 3D face recognition by Fabry
et al. (2008). A related technique, which has until now not been applied to face recognition, is
found in (Wang et al., 2008). Here, groupwise registration between point clouds is performed
by minimizing the Jensen-Shannon divergence between the Gaussian mixture representations
of the point clouds.
3. Implicit surface representations
In general, implicit functions are dened as the iso-level of a scalar function : R
n
R. A
3D implicit surface S is then mathematically dened as
S = {x R
3
|(x) = }. (8)
We call this the iso-surface of the implicit function. The iso-surface at = 0 is sometimes
referred to as the zero contour or zero surface. As such, implicit surfaces are 2Dgeometric shapes
that exist in 3D space (Bloomenthal & Wyvill, 1997). The iso-surface partitions the space into
two regions: interior of the surface, and exterior of the surface. Mostly, the convention is
followed that inside the surface, the function returns negative values and outside the surface,
the function returns positive values. The inside portion is referred as

, while points with

positive values belong to the outside portion
+
. The border between the inside and the
outside is called the interface .
The simplest surfaces (spheroids, ellipsoids,. . . ) can be described by analytic functions and
are called algebraic surfaces. The surface is the set of roots of a polynomial (x) = . The
degree of the surface n is the maximum sum of powers of all terms. The general form of a
linear surface (n = 1), or plane, is
(x, y, z) = ax + by + cz d = , (9)
while the general form for a quadratic surface (n = 2) is:
(x, y, z)
= ax
2
+ bxy + cxz + dx + ey
2
+ f yz + gy + hz
2
+ iz + j
= .
(10)
Superquadrics (n > 2) provide more exibility by adding parameters to control the polyno-
mial exponent, allowing to describe more complex surfaces. Nevertheless, analytic functions
are designed to describe a surface globally by a single closed formula. In reality, it is mostly
not possible to represent a whole real-life object by an analytic function of this form.
3.1 Radial Basis Functions
Radial Basis Functions (RBFs) are another type of implicit functions that have been proven
to be a powerful tool in interpolating scattered data of all kinds, including 3D point clouds
representing 3D objects. A RBF is a function of the form
S(x) =
N

i=1

i
(x x
i
) + p(x), (11)
with
i
the RBF-coefcients, a radial basic function, x
i
the RBF centra and p(x) a polynomial of
low degree.
As can be seen from equation (11), the RBF consists of a weighted sum of radially symmet-
ric basic functions located at the RBF-centra x
i
and a low degree polynomial p. For surface
representation, the RBF-centra x
i
are simply a subset of points on the surface. Finding the
appropriate RBF-coefcients for implicitly representing a surface is done by solving:
x
i
: s(x
i
) = f
i
, (12)
Face Recognition 284
For a surface representation, we want the surface to be the zero-contour of the implicit surface
s(x) and hence f
i
= 0, i. To prevent the interpolation to lead to the trivial solution, s(x) = 0
everywhere, we have to add additional constraints. This is done by adding off-surface points:
points at a distance of the surface, whose implicit function value is different from zero and
mostly equal to the euclidean distance to the surface. Figure 8 gives an example of a RBF
interpolation with zero iso-surface.
Fig. 8. An example of a RBF interpolation with zero iso-surface.
A very clear introduction to the RBF-theory, and info about a fast commercial RBF-
implementation can be found in (Far, 2004). A mathematically very complete reference book
about Radial Basis Functions is (Buhmann, 2003).
The biggest advantage of radial basis function interpolation is the absence of the need for point
connectivity. Other advantages include the low input data requirements (bare point clouds),
and the possibility to insert smoothness constraints when solving for the RBF. A disadvantage
of RBFs is the computational complexity of the problem. This problem can however be alle-
viated by specic mathematical algorithms (Fast Multipole Methods (Beatson & Greengard,
1997)), or compactly supported basis functions (Walder et al., 2006). Because of this computa-
tional complexity, also the editing of the surface is not trivial.
In Claes (2007), a robust framework for both rigid and non-rigid 3D surface representation is
developed to represent faces. This application can be seen as 3D face biometrics in the wide
sense: representing and distinguishing humans by measuring their face geometry. This is used
for craniofacial reconstruction.
Thin Plate Splines, one particular kind of RBF basic function, are popular in non-rigid regis-
tration of face models. Surface registration is an important step in some model-based 3D face
recognition methods, but then the RBF is not used as the surface representation method but
merely as a preprocessing technique (Irfanoglu et al., 2004; Lu & Jain, 2005).
Another application of RBFs in face recognition can be found in (Pears, 2008), where the RBF
is sampled along concentric spheres around certain landmarks to generate features for face
recognition.
3.2 Blobby Models
The blobby model is another kind of implicit surface representation introduced by Blinn
(1982). It was originally perceived as a way to model molecular models for display, and is,
as such, tightly related to the quantum mechanical representation of an electron: a density
function of the spatial location. This way, the molecule surface can be thought of as the
iso-contour of the sum of atom contributions
D(x, y, x) =

i
b
i
exp(a
i
r
2
i
), (13)
where r
i
are distances to the atom locations. Various variants of the original blobby models
exist, which can also be called metaballs or soft objects, and instead of the exponential, one can
also use polynomials (Nishita & Nakamae, 1994) or ellipsoids (Liu et al., 2007) to represent the
blobs.
An advantage of the blobby model surface representation is the apparent possibility for huge
data reduction without loosing much detail. However, the efcient construction of blobby
models is still a problem under research (Liu et al., 2007).
Maruki Muraki (1991) used this blobby model to describe a surface originally represented by
range data with normals. He does this by solving an optimization problem with parameters
x
i
, y
i
, z
i
, a
i
, b
i
with x
i
, y
i
, z
i
the locations of the blobs and a
i
, b
i
the blob parameters. Interest-
ingly, the examples shown in this 1991 paper are representations of faces. It seems that a face
can reasonably well be represented with about 250 blobs, making this representation promis-
ing for 3D face recognition.
Nevertheless, there are not yet applications of this method in 3D face recognition. It has how-
ever been used in the related problem of natural object recognition, where 2D contours were
represented as blobby models, and these blobby models were then used for classication of
the contours (Jorda et al., 2001).
3.3 Euclidean distance functions
A special class of scalar functions are distance functions. The unsigned distance function yields
the distance from a point p to the closest point on the surface S (Jones et al., 2006):
dist
S
(p) = inf
xS
||x p||, (14)
while signed distance functions represent the same, but have a negative sign in

, inside the
object. The signed distance function is constructed by solving the Eikonal equation:
||(x, y, z)|| = 1, (15)
together with the boundary condition |
S
= 0. At any point in space, is the Euclidean
distance to the closest point on S, with a negative sign on the inside and a positive on the
outside (Sigg, 2006). The gradient is orthogonal to the iso-surface and has a unit magnitude
(Jones et al., 2006). An example of a distance function is given in gure 9. The signed dis-
tance function can also be approximated using Radial Basis Functions (Far, 2004), as shown in
gure 8.
One advantage of a surface represented by a distance function is that the surface can easily be
evolved using a level set method. In those methods, also other implicit surface representations
are possible, but distance transforms have nice numerical properties (Osher & Fedkiw, 2003).
Surface representations for 3D face recognition 285
For a surface representation, we want the surface to be the zero-contour of the implicit surface
s(x) and hence f
i
= 0, i. To prevent the interpolation to lead to the trivial solution, s(x) = 0
everywhere, we have to add additional constraints. This is done by adding off-surface points:
points at a distance of the surface, whose implicit function value is different from zero and
mostly equal to the euclidean distance to the surface. Figure 8 gives an example of a RBF
interpolation with zero iso-surface.
Fig. 8. An example of a RBF interpolation with zero iso-surface.
A very clear introduction to the RBF-theory, and info about a fast commercial RBF-
implementation can be found in (Far, 2004). A mathematically very complete reference book
about Radial Basis Functions is (Buhmann, 2003).
The biggest advantage of radial basis function interpolation is the absence of the need for point
connectivity. Other advantages include the low input data requirements (bare point clouds),
and the possibility to insert smoothness constraints when solving for the RBF. A disadvantage
of RBFs is the computational complexity of the problem. This problem can however be alle-
viated by specic mathematical algorithms (Fast Multipole Methods (Beatson & Greengard,
1997)), or compactly supported basis functions (Walder et al., 2006). Because of this computa-
tional complexity, also the editing of the surface is not trivial.
In Claes (2007), a robust framework for both rigid and non-rigid 3D surface representation is
developed to represent faces. This application can be seen as 3D face biometrics in the wide
sense: representing and distinguishing humans by measuring their face geometry. This is used
for craniofacial reconstruction.
Thin Plate Splines, one particular kind of RBF basic function, are popular in non-rigid regis-
tration of face models. Surface registration is an important step in some model-based 3D face
recognition methods, but then the RBF is not used as the surface representation method but
merely as a preprocessing technique (Irfanoglu et al., 2004; Lu & Jain, 2005).
Another application of RBFs in face recognition can be found in (Pears, 2008), where the RBF
is sampled along concentric spheres around certain landmarks to generate features for face
recognition.
3.2 Blobby Models
The blobby model is another kind of implicit surface representation introduced by Blinn
(1982). It was originally perceived as a way to model molecular models for display, and is,
as such, tightly related to the quantum mechanical representation of an electron: a density
function of the spatial location. This way, the molecule surface can be thought of as the
iso-contour of the sum of atom contributions
D(x, y, x) =

i
b
i
exp(a
i
r
2
i
), (13)
where r
i
are distances to the atom locations. Various variants of the original blobby models
exist, which can also be called metaballs or soft objects, and instead of the exponential, one can
also use polynomials (Nishita & Nakamae, 1994) or ellipsoids (Liu et al., 2007) to represent the
blobs.
An advantage of the blobby model surface representation is the apparent possibility for huge
data reduction without loosing much detail. However, the efcient construction of blobby
models is still a problem under research (Liu et al., 2007).
Maruki Muraki (1991) used this blobby model to describe a surface originally represented by
range data with normals. He does this by solving an optimization problem with parameters
x
i
, y
i
, z
i
, a
i
, b
i
with x
i
, y
i
, z
i
the locations of the blobs and a
i
, b
i
the blob parameters. Interest-
ingly, the examples shown in this 1991 paper are representations of faces. It seems that a face
can reasonably well be represented with about 250 blobs, making this representation promis-
ing for 3D face recognition.
Nevertheless, there are not yet applications of this method in 3D face recognition. It has how-
ever been used in the related problem of natural object recognition, where 2D contours were
represented as blobby models, and these blobby models were then used for classication of
the contours (Jorda et al., 2001).
3.3 Euclidean distance functions
A special class of scalar functions are distance functions. The unsigned distance function yields
the distance from a point p to the closest point on the surface S (Jones et al., 2006):
dist
S
(p) = inf
xS
||x p||, (14)
while signed distance functions represent the same, but have a negative sign in

, inside the
object. The signed distance function is constructed by solving the Eikonal equation:
||(x, y, z)|| = 1, (15)
together with the boundary condition |
S
= 0. At any point in space, is the Euclidean
distance to the closest point on S, with a negative sign on the inside and a positive on the
outside (Sigg, 2006). The gradient is orthogonal to the iso-surface and has a unit magnitude
(Jones et al., 2006). An example of a distance function is given in gure 9. The signed dis-
tance function can also be approximated using Radial Basis Functions (Far, 2004), as shown in
gure 8.
One advantage of a surface represented by a distance function is that the surface can easily be
evolved using a level set method. In those methods, also other implicit surface representations
are possible, but distance transforms have nice numerical properties (Osher & Fedkiw, 2003).
Face Recognition 286
Fig. 9. An example of a distance function.
An interesting application in face recognition (in 2D though) is given in (Akhlou & Bendada,
2008) where a distance transform is used to get an invariant representation for face recogni-
tion, using thermal face images. After extraction of the face region, a clustering technique
constructs the facial isotherm layers. Computing the medial axis in each layer provides an
image containing physiological features, called face print image. A Euclidean distance trans-
formprovides the necessary invariance in the matching process. Related to the domain of face
recognition, the signed distance function is used in craniofacial reconstruction (Vandermeulen
et al., 2006). A reference skull, represented as distance maps, is warped to all target skulls and
subsequently these warps are applied to the reference head distance map.
Signed distance maps are also interesting for aligning surfaces, as described in Hansen et al.
(2007). Symmetric registration of two surfaces, represented as signed distance maps, is done
by minimizing the energy functional:
F(p) =

xU
r
x
(
y
(W(x; p))
x
(x))
2
+

yU
r
y
(
x
(W(y; p))
y
(y))
2
,
(16)
with W(; p) the warp function, U
r
x
and U
r
y
the narrow bands around the surfaces S
x
and S
y
and the signed distance map. The width of the narrow band r should be larger than the
width of the largest structure. Hansen et al. (2007) state that the level set registration performs
slightly better than the standard ICP algorithm (Besl & McKay, 1992).
3.4 Random walk functions
This scalar surface representation gives at a point in space a value that is the average time of
a random walk to reach the surface starting from that point. This scalar function is the result
of solving the Poisson equation:
(x, y, z) = 1, (17)
again subject to the boundary condition |
S
= 0 and with =

2

x
2
+

2

y
2
+

2

z
2
. For every
internal point in the surface, the function assigns a value reecting the mean time required for
a random walk beginning at the boundaries and ending in this particular point. The level sets
of represent smoother versions of the bounding surface. A disadvantage of this function is
that a unique solution of equation (17) only exists within a closed surface. An example of a
random walk function is given in gure 10.
Fig. 10. An example of a random walk function.
To the best of our knowledge, this scalar function is not yet used in face recognition. However,
it has already been proven to be useful in 2Dobject classication which makes it for auspicious
for use in biometrics (Gorelick et al., 2006).
4. Conclusions
We can conclude that, although many representations in biometrics are based on meshes, a
number of interesting alternatives exist. We have given a systematic discussion of the differ-
ent three-dimensional surface representations that seem to be promising for use in 3D bio-
metrics and, if known, their already existing use. While we are aware of the non-exhaustive
nature of this work, we hope to have given to the face recognition and other related research
communities (computer graphics, mathematics of surfaces,. . . ) some interesting ideas.
We have paid attention to the advantages and disadvantages of the different surface represen-
tations throughout the whole text. The main advantages and disadvantages of the different
surface representations are summarized in table 1. From this we can conclude that many of
the advantages of the different surface representations have not yet been taken advantage of
in current face recognition research. This could thus be very interesting for future research.
Other interesting conclusions can be drawn from the preceeding text and table. First of all, we
see that the left branch of our taxonomy, the explicit representations, is much more frequently
used in todays face recognition research, opposed to the other branch, the implicit represen-
tations. This can be explained by the fact that explicit representations have been very much a
topic of interest in computer graphics because of hardware requirements and are as such also
the rst to be considered in 3D face recognition.
Furthermore: although the polygonal mesh is often used in 3D face recognition and certainly
has some advantages, we think it is more important to keep the other surface representations
in mind for doing face recognition research. Moreover, we already see a gain in importance
of meshfree methods in the eld of numerical analysis, where meshfree methods are used
Surface representations for 3D face recognition 287
Fig. 9. An example of a distance function.
An interesting application in face recognition (in 2D though) is given in (Akhlou & Bendada,
2008) where a distance transform is used to get an invariant representation for face recogni-
tion, using thermal face images. After extraction of the face region, a clustering technique
constructs the facial isotherm layers. Computing the medial axis in each layer provides an
image containing physiological features, called face print image. A Euclidean distance trans-
formprovides the necessary invariance in the matching process. Related to the domain of face
recognition, the signed distance function is used in craniofacial reconstruction (Vandermeulen
et al., 2006). A reference skull, represented as distance maps, is warped to all target skulls and
subsequently these warps are applied to the reference head distance map.
Signed distance maps are also interesting for aligning surfaces, as described in Hansen et al.
(2007). Symmetric registration of two surfaces, represented as signed distance maps, is done
by minimizing the energy functional:
F(p) =

xU
r
x
(
y
(W(x; p))
x
(x))
2
+

yU
r
y
(
x
(W(y; p))
y
(y))
2
,
(16)
with W(; p) the warp function, U
r
x
and U
r
y
the narrow bands around the surfaces S
x
and S
y
and the signed distance map. The width of the narrow band r should be larger than the
width of the largest structure. Hansen et al. (2007) state that the level set registration performs
slightly better than the standard ICP algorithm (Besl & McKay, 1992).
3.4 Random walk functions
This scalar surface representation gives at a point in space a value that is the average time of
a random walk to reach the surface starting from that point. This scalar function is the result
of solving the Poisson equation:
(x, y, z) = 1, (17)
again subject to the boundary condition |
S
= 0 and with =

2

x
2
+

2

y
2
+

2

z
2
. For every
internal point in the surface, the function assigns a value reecting the mean time required for
a random walk beginning at the boundaries and ending in this particular point. The level sets
of represent smoother versions of the bounding surface. A disadvantage of this function is
that a unique solution of equation (17) only exists within a closed surface. An example of a
random walk function is given in gure 10.
Fig. 10. An example of a random walk function.
To the best of our knowledge, this scalar function is not yet used in face recognition. However,
it has already been proven to be useful in 2Dobject classication which makes it for auspicious
for use in biometrics (Gorelick et al., 2006).
4. Conclusions
We can conclude that, although many representations in biometrics are based on meshes, a
number of interesting alternatives exist. We have given a systematic discussion of the differ-
ent three-dimensional surface representations that seem to be promising for use in 3D bio-
metrics and, if known, their already existing use. While we are aware of the non-exhaustive
nature of this work, we hope to have given to the face recognition and other related research
communities (computer graphics, mathematics of surfaces,. . . ) some interesting ideas.
We have paid attention to the advantages and disadvantages of the different surface represen-
tations throughout the whole text. The main advantages and disadvantages of the different
surface representations are summarized in table 1. From this we can conclude that many of
the advantages of the different surface representations have not yet been taken advantage of
in current face recognition research. This could thus be very interesting for future research.
Other interesting conclusions can be drawn from the preceeding text and table. First of all, we
see that the left branch of our taxonomy, the explicit representations, is much more frequently
used in todays face recognition research, opposed to the other branch, the implicit represen-
tations. This can be explained by the fact that explicit representations have been very much a
topic of interest in computer graphics because of hardware requirements and are as such also
the rst to be considered in 3D face recognition.
Furthermore: although the polygonal mesh is often used in 3D face recognition and certainly
has some advantages, we think it is more important to keep the other surface representations
in mind for doing face recognition research. Moreover, we already see a gain in importance
of meshfree methods in the eld of numerical analysis, where meshfree methods are used
Face Recognition 288
Accuracy
Conciseness
Acquiering
Intuitiveness
Parameterization
Computability
Displayability
M
a
i
n
a
d
v
a
n
t
a
g
e
M
a
i
n
d
i
s
a
d
v
a
n
t
a
g
e
P
r
e
v
i
o
u
s
u
s
e
i
n
f
a
c
e
r
e
c
o
g
n
i
t
i
o
n
P
o
i
n
t
C
l
o
u
d

+
+

+
+
S
i
m
p
l
i
c
i
t
y
S
p
a
r
s
e
n
e
s
s
m
a
n
y
(
s
e
e
t
e
x
t
)
C
o
n
t
o
u
r
a
n
d
p
r
o

l
e
c
u
r
v
e
s

+
+

+
I
n
v
a
r
i
a
n
c
e
S
p
a
r
s
e
n
e
s
s
m
a
n
y
(
s
e
e
t
e
x
t
)
P
o
l
y
g
o
n
a
l
M
e
s
h
+

+
M
a
n
y
t
o
o
l
s
D
i
s
c
r
e
t
e
m
a
n
y
(
s
e
e
t
e
x
t
)
P
a
r
a
m
e
t
r
i
c

H
e
i
g
h
t
m
a
p

+
+

A
q
u
i
r
i
n
g
S
e
l
f
O
c
c
l
u
s
i
o
n
m
a
n
y
(
s
e
e
t
e
x
t
)

G
e
o
m
e
t
r
y
i
m
a
g
e
+

P
a
r
a
m
e
t
e
r
i
z
a
t
i
o
n
C
o
n
s
t
r
u
c
t
i
o
n
M
e
t
a
x
a
s
&
K
a
k
a
d
i
a
r
i
s
(
2
0
0
2
)

S
p
l
i
n
e
s
+
+

+
+
+
P
a
r
a
m
e
t
e
r
i
z
a
t
i
o
n
C
o
m
p
u
t
a
b
i
l
i
t
y
n
o
t
y
e
t
.
.
.
S
p
h
e
r
i
c
a
l
h
a
r
m
o
n
i
c
s
+
+

M
u
l
t
i
r
e
s
o
l
u
t
i
o
n
G
l
o
b
a
l
l
i
t
y
L
l
o
n
c
h
e
t
a
l
.
(
2
0
0
9
)
P
o
i
n
t
-
C
l
o
u
d
b
a
s
e
d

S
o
u
n
d
t
h
e
o
r
y
C
o
m
p
u
t
a
b
i
l
i
t
y
F
a
b
r
y
e
t
a
l
.
(
2
0
0
8
)
R
a
d
i
a
l
B
a
s
i
s
F
u
n
c
t
i
o
n
s
+

C
o
m
p
l
e
t
e
n
e
s
s
C
o
m
p
u
t
a
b
i
l
i
t
y
n
o
t
y
e
t
.
.
.
B
l
o
b
b
y
M
o
d
e
l
s
+
+

C
o
m
p
r
e
s
s
i
o
n
C
o
n
s
t
r
u
c
t
i
o
n
n
o
t
y
e
t
.
.
.
D
i
s
t
a
n
c
e
F
u
n
c
t
i
o
n
s
+

E
v
o
l
u
t
i
o
n
M
e
m
o
r
y
r
e
q
.
n
o
t
y
e
t
.
.
.
R
a
n
d
o
m
w
a
l
k
f
u
n
c
t
i
o
n
+

E
v
o
l
u
t
i
o
n
M
e
m
o
r
y
r
e
q
.
n
o
t
y
e
t
.
.
.
T
a
b
l
e
1
.
S
u
m
m
a
r
y
o
f
a
d
v
a
n
t
a
g
e
s
a
n
d
d
i
s
a
d
v
a
n
t
a
g
e
s
o
f
t
h
e
d
i
s
c
u
s
s
e
d
s
u
r
f
a
c
e
r
e
p
r
e
s
e
n
t
a
t
i
o
n
s
.
in nite element modelling, for solving partial differential equations or for approximating
functions without using classic mesh discretisations. These kind of methods have many ad-
vantages: easy handling of large deformations because of the lack of connectivity between
points, good handling of topological changes, ability to include prior knowledge, support for
exible adaptable renement procedures, the ability to support multiscale,. . . (Li & Liu, 2004).
Also in computer graphics, meshfree surface representations are gaining importance, espe-
cially in the physically based deformable models (Nealen et al., 2006), but also in many other
computer graphics subelds (Dyn et al., 2008; Sukumar, 2005). In 3D face recognition, some
progressive methods make use of some of these interesting meshfree surface representations
as explained in this chapter.
5. References
Akarun, L., Gokberk, B. & Salah, A. A. (2005). 3D face recognition for biometric applications,
EUSIPCO 05: Proceedings of the 13th European Signal Processing Conference, Antalya.
Akhlou, M. & Bendada, A. H. (2008). Infrared face recognition using distance transforms,
ICIVC 2008: International Conference on Image and Vision Computing, Vol. 30 of Proceed-
ings of World Academy of Science, Engineering & Technology, Paris, France, pp. 160163.
Al-Osaimi, F., Bennamoun, M. & Mian, A. (2009). An expression deformation approach to
non-rigid 3D face recognition, International Journal of Computer Vision 81(3): 302316.
Alyz, N., Gkberk, B. & Akarun, L. (2008). A 3D face recognition system for expression and
occlusion invariance, BTAS 08: Proceedings of the IEEE Second International Conference
on Biometrics Theory, Applications and Systems, Arlington, Virginia, USA.
Amberg, B., Knothe, R. & Vetter, T. (2008). Expression invariant 3D face recognition with a
morphable model, FG 08: Proceedings of the 8th IEEE International Conference on Auto-
matic Face and Gesture Recognition, IEEE Computer Society, Amsterdam, The Nether-
lands.
Amenta, N., Choi, S. & Kolluri, R. K. (2001). The power crust, SMA 01: Proceedings of the
sixth ACM symposium on Solid modeling and applications, ACM, New York, NY, USA,
pp. 249266.
Beatson, R. & Greengard, L. (1997). A short course on fast multipole methods, Wavelets, Mul-
tilevel Methods and Elliptic PDEs, Oxford University Press, pp. 137.
Berretti, S., Bimbo, A. D. & Pala, P. (2006). Description and retrieval of 3D face models using
iso-geodesic stripes, MIR 06: Proceedings of the 8th ACM international workshop on
Multimedia information retrieval, ACM, New York, NY, USA, pp. 1322.
Besl, P. (1990). The free-form surface matching problem, Machine vision for three-dimensional
scenes pp. 2571.
Besl, P. J. & McKay, N. D. (1992). A method for registration of 3-D shapes, IEEE Transactions on
Pattern Analysis and Machine Intelligence 14(2): 239256.
Blinn, J. F. (1982). A generalization of algebraic surface drawing, ACM Trans. Graph. 1(3): 235
256.
Bloomenthal, J. & Wyvill, B. (eds) (1997). Introduction to Implicit Surfaces, Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA.
Bronstein, A. M., Bronstein, M. M. & Kimmel, R. (2003). Expression-invariant 3D face recog-
nition, in J. Kittler & M. Nixon (eds), AVBPA 03: Proceedings of the 4th International
Conference on Audio and Video-based Biometric Person Authentication, Vol. 2688 of Lec-
ture Notes in Computer Science, Springer, pp. 6269.
Surface representations for 3D face recognition 289
Accuracy
Conciseness
Acquiering
Intuitiveness
Parameterization
Computability
Displayability
M
a
i
n
a
d
v
a
n
t
a
g
e
M
a
i
n
d
i
s
a
d
v
a
n
t
a
g
e
P
r
e
v
i
o
u
s
u
s
e
i
n
f
a
c
e
r
e
c
o
g
n
i
t
i
o
n
P
o
i
n
t
C
l
o
u
d

+
+

+
+
S
i
m
p
l
i
c
i
t
y
S
p
a
r
s
e
n
e
s
s
m
a
n
y
(
s
e
e
t
e
x
t
)
C
o
n
t
o
u
r
a
n
d
p
r
o

l
e
c
u
r
v
e
s

+
+

+
I
n
v
a
r
i
a
n
c
e
S
p
a
r
s
e
n
e
s
s
m
a
n
y
(
s
e
e
t
e
x
t
)
P
o
l
y
g
o
n
a
l
M
e
s
h
+

+
M
a
n
y
t
o
o
l
s
D
i
s
c
r
e
t
e
m
a
n
y
(
s
e
e
t
e
x
t
)
P
a
r
a
m
e
t
r
i
c

H
e
i
g
h
t
m
a
p

+
+

A
q
u
i
r
i
n
g
S
e
l
f
O
c
c
l
u
s
i
o
n
m
a
n
y
(
s
e
e
t
e
x
t
)

G
e
o
m
e
t
r
y
i
m
a
g
e
+

P
a
r
a
m
e
t
e
r
i
z
a
t
i
o
n
C
o
n
s
t
r
u
c
t
i
o
n
M
e
t
a
x
a
s
&
K
a
k
a
d
i
a
r
i
s
(
2
0
0
2
)

S
p
l
i
n
e
s
+
+

+
+
+
P
a
r
a
m
e
t
e
r
i
z
a
t
i
o
n
C
o
m
p
u
t
a
b
i
l
i
t
y
n
o
t
y
e
t
.
.
.
S
p
h
e
r
i
c
a
l
h
a
r
m
o
n
i
c
s
+
+

M
u
l
t
i
r
e
s
o
l
u
t
i
o
n
G
l
o
b
a
l
l
i
t
y
L
l
o
n
c
h
e
t
a
l
.
(
2
0
0
9
)
P
o
i
n
t
-
C
l
o
u
d
b
a
s
e
d

S
o
u
n
d
t
h
e
o
r
y
C
o
m
p
u
t
a
b
i
l
i
t
y
F
a
b
r
y
e
t
a
l
.
(
2
0
0
8
)
R
a
d
i
a
l
B
a
s
i
s
F
u
n
c
t
i
o
n
s
+

C
o
m
p
l
e
t
e
n
e
s
s
C
o
m
p
u
t
a
b
i
l
i
t
y
n
o
t
y
e
t
.
.
.
B
l
o
b
b
y
M
o
d
e
l
s
+
+

C
o
m
p
r
e
s
s
i
o
n
C
o
n
s
t
r
u
c
t
i
o
n
n
o
t
y
e
t
.
.
.
D
i
s
t
a
n
c
e
F
u
n
c
t
i
o
n
s
+

E
v
o
l
u
t
i
o
n
M
e
m
o
r
y
r
e
q
.
n
o
t
y
e
t
.
.
.
R
a
n
d
o
m
w
a
l
k
f
u
n
c
t
i
o
n
+

E
v
o
l
u
t
i
o
n
M
e
m
o
r
y
r
e
q
.
n
o
t
y
e
t
.
.
.
T
a
b
l
e
1
.
S
u
m
m
a
r
y
o
f
a
d
v
a
n
t
a
g
e
s
a
n
d
d
i
s
a
d
v
a
n
t
a
g
e
s
o
f
t
h
e
d
i
s
c
u
s
s
e
d
s
u
r
f
a
c
e
r
e
p
r
e
s
e
n
t
a
t
i
o
n
s
.
in nite element modelling, for solving partial differential equations or for approximating
functions without using classic mesh discretisations. These kind of methods have many ad-
vantages: easy handling of large deformations because of the lack of connectivity between
points, good handling of topological changes, ability to include prior knowledge, support for
exible adaptable renement procedures, the ability to support multiscale,. . . (Li & Liu, 2004).
Also in computer graphics, meshfree surface representations are gaining importance, espe-
cially in the physically based deformable models (Nealen et al., 2006), but also in many other
computer graphics subelds (Dyn et al., 2008; Sukumar, 2005). In 3D face recognition, some
progressive methods make use of some of these interesting meshfree surface representations
as explained in this chapter.
5. References
Akarun, L., Gokberk, B. & Salah, A. A. (2005). 3D face recognition for biometric applications,
EUSIPCO 05: Proceedings of the 13th European Signal Processing Conference, Antalya.
Akhlou, M. & Bendada, A. H. (2008). Infrared face recognition using distance transforms,
ICIVC 2008: International Conference on Image and Vision Computing, Vol. 30 of Proceed-
ings of World Academy of Science, Engineering & Technology, Paris, France, pp. 160163.
Al-Osaimi, F., Bennamoun, M. & Mian, A. (2009). An expression deformation approach to
non-rigid 3D face recognition, International Journal of Computer Vision 81(3): 302316.
Alyz, N., Gkberk, B. & Akarun, L. (2008). A 3D face recognition system for expression and
occlusion invariance, BTAS 08: Proceedings of the IEEE Second International Conference
on Biometrics Theory, Applications and Systems, Arlington, Virginia, USA.
Amberg, B., Knothe, R. & Vetter, T. (2008). Expression invariant 3D face recognition with a
morphable model, FG 08: Proceedings of the 8th IEEE International Conference on Auto-
matic Face and Gesture Recognition, IEEE Computer Society, Amsterdam, The Nether-
lands.
Amenta, N., Choi, S. & Kolluri, R. K. (2001). The power crust, SMA 01: Proceedings of the
sixth ACM symposium on Solid modeling and applications, ACM, New York, NY, USA,
pp. 249266.
Beatson, R. & Greengard, L. (1997). A short course on fast multipole methods, Wavelets, Mul-
tilevel Methods and Elliptic PDEs, Oxford University Press, pp. 137.
Berretti, S., Bimbo, A. D. & Pala, P. (2006). Description and retrieval of 3D face models using
iso-geodesic stripes, MIR 06: Proceedings of the 8th ACM international workshop on
Multimedia information retrieval, ACM, New York, NY, USA, pp. 1322.
Besl, P. (1990). The free-form surface matching problem, Machine vision for three-dimensional
scenes pp. 2571.
Besl, P. J. & McKay, N. D. (1992). A method for registration of 3-D shapes, IEEE Transactions on
Pattern Analysis and Machine Intelligence 14(2): 239256.
Blinn, J. F. (1982). A generalization of algebraic surface drawing, ACM Trans. Graph. 1(3): 235
256.
Bloomenthal, J. & Wyvill, B. (eds) (1997). Introduction to Implicit Surfaces, Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA.
Bronstein, A. M., Bronstein, M. M. & Kimmel, R. (2003). Expression-invariant 3D face recog-
nition, in J. Kittler & M. Nixon (eds), AVBPA 03: Proceedings of the 4th International
Conference on Audio and Video-based Biometric Person Authentication, Vol. 2688 of Lec-
ture Notes in Computer Science, Springer, pp. 6269.
Face Recognition 290
Bronstein, A. M., Bronstein, M. M. & Kimmel, R. (2005). Three-dimensional face recognition,
International Journal of Computer Vision 64(1): 530.
Buhmann, M. (2003). Radial Basis Functions: Theory and Implementations, Cambridge University
Press.
Blow, T. &Daniilidis, K. (2001). Surface representations using spherical harmonics and gabor
wavelets on the sphere, Technical Report MS-CIS-01-37, University of Pennsylvania,
Department of Computer and Information Science.
Campbell, R. J. & Flynn, P. J. (2000). Survey of free-form object representation and recognition
techniques, Computer Vision and Image Understanding 81: 166210.
Chang, K. I., Bowyer, K. W. & Flynn, P. J. (2006). Multiple nose region matching for 3D face
recognition under varying facial expression, IEEE Transactions on Pattern Analysis and
Machine Intelligence 28(10): 16951700.
Claes, P. (2007). A robust statistical surface registration framework using implicit function represen-
tations - application in craniofacial reconstruction, PhD thesis, Katholieke Universiteit
Leuven, Leuven, Belgium.
URL: http://www.medicalimagingcenter.be/PhD/PeterClaes/Thesis.pdf
Colbry, D. & Stockman, G. C. (2007). Canonical face depth map: A robust 3D representation
for face verication, CVPR 07: Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, IEEE Computer Society, Minneapolis, Min-
nesota, USA.
Curless, B. &Levoy, M. (1996). Avolumetric method for building complex models fromrange
images, SIGGRAPH 96: Proceedings of the 23rd annual conference on Computer graphics
and interactive techniques, ACM, New York, NY, USA, pp. 303312.
Dey, T., Giesen, J. & Goswami, S. (2003). Shape segmentation and matching with ow dis-
cretization, Algorithms and Data Structures pp. 2536.
URL: http://www.springerlink.com/content/6qwfh2adqmm60n30
Dillenseger, J.-L., Guillaume, H. & Patard, J.-J. (2006). Spherical harmonics based intrasubject
3-d kidney modeling/registration technique applied on partial information, Biomed-
ical Engineering, IEEE Transactions on 53(11): 21852193.
Dyn, N., Iske, A. & Wendland, H. (2008). Meshfree thinning of 3d point clouds, Foundations of
Computational Mathematics 8(4): 409425.
Edelsbrunner, H. & Mcke, E. P. (1994). Three-dimensional alpha shapes, ACM Trans. Graph.
13(1): 4372.
Fabio, R. (2003). From point cloud to surface: the modeling and visualization problem,
Workshop on Visualization and Animation of Reality-based 3D Models, Tarasp-Vulpera,
Switzerland.
Fabry, T., Vandermeulen, D. & Suetens, P. (2008). 3D face recognition using point cloud ker-
nel correlation, BTAS 08: Proceedings of the IEEE Second International Conference on
Biometrics Theory, Applications and Systems, Arlington, Virginia, USA.
Faltemier, T., Bowyer, K. W. & Flynn, P. J. (2008). A region ensemble for 3-D face recognition,
IEEE Transactions on Information Forensics and Security 3(1): 6273.
Far (2004). FastRBF MATLAB Toolbox Manual.
Feng, S., Krim, H., Gu, I. & Viberg, M. (2006). 3D face recognition using afne integral invari-
ants, ICASSP 06: Proceedings of the IEEE International Conference on Acoustics, Speech
and Signal Processing, Vol. 2, Toulouse, France, pp. 189192.
Feng, S., Krim, H. & Kogan, I. A. (2007). 3D face recognition using euclidean integral in-
variants signature, SSP 07: IEEE/SP 14th Workshop on Statistical Signal Processing,
Madison, WI, USA, pp. 156160.
Gorelick, L., Galun, M. & Brandt, A. (2006). Shape representation and classication using
the poisson equation, IEEE Transactions on Pattern Analysis and Machine Intelligence
28(12): 19912005. Member-Eitan Sharon and Member-Ronen Basri.
Gu, X., Gortler, S. J. & Hoppe, H. (2002). Geometry images, ACM Trans. Graph. 21(3): 355361.
Gupta, S., Aggarwal, J. K., Markey, M. K. & Bovik, A. C. (2007). 3D face recognition founded
on the structural diversity of human faces, CVPR 07: Proceedings of the International
Conference on Computer Vision and Pattern Recognition, IEEE Computer Society, Min-
neapolis, Minnesota, USA.
Hansen, M. F., Erbou, S. G. H., Vester-Christensen, M., Larsen, R., Ersboll, B. K. & Chris-
tensen, L. B. (2007). Surface-to-surface registration using level sets, in B. K. Ersboll &
K. S. Pedersen (eds), SCIA 07: Proceedings of the 15th Scandinavian Conference on Image
Analysis, Vol. 4522 of Lecture Notes in Computer Science, Springer, pp. 780788.
Hoppe, H., DeRose, T., Duchamp, T., McDonald, J. & Stuetzle, W. (1992). Surface reconstruc-
tion fromunorganized points, SIGGRAPH 92: Proceedings of the 19th annual conference
on Computer graphics and interactive techniques, ACM, New York, NY, USA, pp. 7178.
Hubeli, A. & Gross, M. (2000). A survey of surface representations for geometric modeling,
Technical Report 335, ETH Zrich.
Irfanoglu, M., Gokberk, B. & Akarun, L. (2004). 3D shape-based face recognition using auto-
matically registered facial surfaces, Pattern Recognition, 2004. ICPR 2004. Proceedings
of the 17th International Conference on 4: 183186 Vol.4.
Jahanbin, S., Choi, H., Liu, Y. & Bovik, A. C. (2008). Three dimensional face recognition using
iso-geodesic and iso-depth curves, BTAS 08: Proceedings of the IEEE Second Interna-
tional Conference on Biometrics Theory, Applications and Systems, Arlington, Virginia,
USA.
Jones, M. W., Baerentzen, J. A. &Sramek, M. (2006). 3Ddistance elds: Asurvey of techniques
and applications, IEEE Transactions on Visualization and Computer Graphics 12(4): 581
599.
Jorda, A. R., Vanacloig, V. A. & Garca, G. A. (2001). A geometric feature set based on shape
description primitives to recognize natural objects, Proceedings of the IX Spanish Sym-
posium on Pattern Recognition and Image Analysis, Benicasim, Spain.
Kakadiaris, I. A., Passalis, G., Toderici, G., Murtuza, M. N., Lu, Y., Karampatziakis, N. &
Theoharis, T. (2007). Three-dimensional face recognition in the presence of facial
expressions: An annotated deformable model approach, IEEE Transactions on Pattern
Analysis and Machine Intelligence 29(4): 640649.
Kalaiah, A. & Varshney, A. (2003). Statistical point geometry, SGP 03: Proceedings of the 2003
Eurographics/ACM SIGGRAPH symposium on Geometry processing, Eurographics Asso-
ciation, pp. 107115.
Kazhdan, M., Funkhouser, T. & Rusinkiewicz, S. (2003). Rotation invariant spherical har-
monic representation of 3d shape descriptors, SGP 03: Proceedings of the 2003 Euro-
graphics/ACM SIGGRAPH symposium on Geometry processing, Eurographics Associa-
tion, Aire-la-Ville, Switzerland, Switzerland, pp. 156164.
Kimmel, R. & Sethian, J. A. (1998). Computing geodesic paths on manifolds, Proceedings of the
National Academy of Sciences of the United States of America 95: 84318435.
Surface representations for 3D face recognition 291
Bronstein, A. M., Bronstein, M. M. & Kimmel, R. (2005). Three-dimensional face recognition,
International Journal of Computer Vision 64(1): 530.
Buhmann, M. (2003). Radial Basis Functions: Theory and Implementations, Cambridge University
Press.
Blow, T. &Daniilidis, K. (2001). Surface representations using spherical harmonics and gabor
wavelets on the sphere, Technical Report MS-CIS-01-37, University of Pennsylvania,
Department of Computer and Information Science.
Campbell, R. J. & Flynn, P. J. (2000). Survey of free-form object representation and recognition
techniques, Computer Vision and Image Understanding 81: 166210.
Chang, K. I., Bowyer, K. W. & Flynn, P. J. (2006). Multiple nose region matching for 3D face
recognition under varying facial expression, IEEE Transactions on Pattern Analysis and
Machine Intelligence 28(10): 16951700.
Claes, P. (2007). A robust statistical surface registration framework using implicit function represen-
tations - application in craniofacial reconstruction, PhD thesis, Katholieke Universiteit
Leuven, Leuven, Belgium.
URL: http://www.medicalimagingcenter.be/PhD/PeterClaes/Thesis.pdf
Colbry, D. & Stockman, G. C. (2007). Canonical face depth map: A robust 3D representation
for face verication, CVPR 07: Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, IEEE Computer Society, Minneapolis, Min-
nesota, USA.
Curless, B. &Levoy, M. (1996). Avolumetric method for building complex models fromrange
images, SIGGRAPH 96: Proceedings of the 23rd annual conference on Computer graphics
and interactive techniques, ACM, New York, NY, USA, pp. 303312.
Dey, T., Giesen, J. & Goswami, S. (2003). Shape segmentation and matching with ow dis-
cretization, Algorithms and Data Structures pp. 2536.
URL: http://www.springerlink.com/content/6qwfh2adqmm60n30
Dillenseger, J.-L., Guillaume, H. & Patard, J.-J. (2006). Spherical harmonics based intrasubject
3-d kidney modeling/registration technique applied on partial information, Biomed-
ical Engineering, IEEE Transactions on 53(11): 21852193.
Dyn, N., Iske, A. & Wendland, H. (2008). Meshfree thinning of 3d point clouds, Foundations of
Computational Mathematics 8(4): 409425.
Edelsbrunner, H. & Mcke, E. P. (1994). Three-dimensional alpha shapes, ACM Trans. Graph.
13(1): 4372.
Fabio, R. (2003). From point cloud to surface: the modeling and visualization problem,
Workshop on Visualization and Animation of Reality-based 3D Models, Tarasp-Vulpera,
Switzerland.
Fabry, T., Vandermeulen, D. & Suetens, P. (2008). 3D face recognition using point cloud ker-
nel correlation, BTAS 08: Proceedings of the IEEE Second International Conference on
Biometrics Theory, Applications and Systems, Arlington, Virginia, USA.
Faltemier, T., Bowyer, K. W. & Flynn, P. J. (2008). A region ensemble for 3-D face recognition,
IEEE Transactions on Information Forensics and Security 3(1): 6273.
Far (2004). FastRBF MATLAB Toolbox Manual.
Feng, S., Krim, H., Gu, I. & Viberg, M. (2006). 3D face recognition using afne integral invari-
ants, ICASSP 06: Proceedings of the IEEE International Conference on Acoustics, Speech
and Signal Processing, Vol. 2, Toulouse, France, pp. 189192.
Feng, S., Krim, H. & Kogan, I. A. (2007). 3D face recognition using euclidean integral in-
variants signature, SSP 07: IEEE/SP 14th Workshop on Statistical Signal Processing,
Madison, WI, USA, pp. 156160.
Gorelick, L., Galun, M. & Brandt, A. (2006). Shape representation and classication using
the poisson equation, IEEE Transactions on Pattern Analysis and Machine Intelligence
28(12): 19912005. Member-Eitan Sharon and Member-Ronen Basri.
Gu, X., Gortler, S. J. & Hoppe, H. (2002). Geometry images, ACM Trans. Graph. 21(3): 355361.
Gupta, S., Aggarwal, J. K., Markey, M. K. & Bovik, A. C. (2007). 3D face recognition founded
on the structural diversity of human faces, CVPR 07: Proceedings of the International
Conference on Computer Vision and Pattern Recognition, IEEE Computer Society, Min-
neapolis, Minnesota, USA.
Hansen, M. F., Erbou, S. G. H., Vester-Christensen, M., Larsen, R., Ersboll, B. K. & Chris-
tensen, L. B. (2007). Surface-to-surface registration using level sets, in B. K. Ersboll &
K. S. Pedersen (eds), SCIA 07: Proceedings of the 15th Scandinavian Conference on Image
Analysis, Vol. 4522 of Lecture Notes in Computer Science, Springer, pp. 780788.
Hoppe, H., DeRose, T., Duchamp, T., McDonald, J. & Stuetzle, W. (1992). Surface reconstruc-
tion fromunorganized points, SIGGRAPH 92: Proceedings of the 19th annual conference
on Computer graphics and interactive techniques, ACM, New York, NY, USA, pp. 7178.
Hubeli, A. & Gross, M. (2000). A survey of surface representations for geometric modeling,
Technical Report 335, ETH Zrich.
Irfanoglu, M., Gokberk, B. & Akarun, L. (2004). 3D shape-based face recognition using auto-
matically registered facial surfaces, Pattern Recognition, 2004. ICPR 2004. Proceedings
of the 17th International Conference on 4: 183186 Vol.4.
Jahanbin, S., Choi, H., Liu, Y. & Bovik, A. C. (2008). Three dimensional face recognition using
iso-geodesic and iso-depth curves, BTAS 08: Proceedings of the IEEE Second Interna-
tional Conference on Biometrics Theory, Applications and Systems, Arlington, Virginia,
USA.
Jones, M. W., Baerentzen, J. A. &Sramek, M. (2006). 3Ddistance elds: Asurvey of techniques
and applications, IEEE Transactions on Visualization and Computer Graphics 12(4): 581
599.
Jorda, A. R., Vanacloig, V. A. & Garca, G. A. (2001). A geometric feature set based on shape
description primitives to recognize natural objects, Proceedings of the IX Spanish Sym-
posium on Pattern Recognition and Image Analysis, Benicasim, Spain.
Kakadiaris, I. A., Passalis, G., Toderici, G., Murtuza, M. N., Lu, Y., Karampatziakis, N. &
Theoharis, T. (2007). Three-dimensional face recognition in the presence of facial
expressions: An annotated deformable model approach, IEEE Transactions on Pattern
Analysis and Machine Intelligence 29(4): 640649.
Kalaiah, A. & Varshney, A. (2003). Statistical point geometry, SGP 03: Proceedings of the 2003
Eurographics/ACM SIGGRAPH symposium on Geometry processing, Eurographics Asso-
ciation, pp. 107115.
Kazhdan, M., Funkhouser, T. & Rusinkiewicz, S. (2003). Rotation invariant spherical har-
monic representation of 3d shape descriptors, SGP 03: Proceedings of the 2003 Euro-
graphics/ACM SIGGRAPH symposium on Geometry processing, Eurographics Associa-
tion, Aire-la-Ville, Switzerland, Switzerland, pp. 156164.
Kimmel, R. & Sethian, J. A. (1998). Computing geodesic paths on manifolds, Proceedings of the
National Academy of Sciences of the United States of America 95: 84318435.
Face Recognition 292
Levoy, M. & Whitted, T. (1985). The use of points as a display primitive, Technical report,
Department of Computer Science, University of North Carolina at Chappel Hill.
Li, L., Xu, C., Tang, W. & Zhong, C. (2008). 3d face recognition by constructing deformation
invariant image, Pattern Recognition Letters 29(10): 15961602.
Li, S. & Liu, W. (2004). Meshfree particle methods, Springer.
Li, X. & Zhang, H. (2007). Adapting geometric attributes for expression-invariant 3D face
recognition, SMI 07: Proceedings of the IEEE International Conference on Shape Modeling
and Applications, IEEE Computer Society, Washington, DC, USA, pp. 2132.
Liu, D., Shen, L. & Lam, K. (2005). Image Synthesis and Face Recognition Based on 3D Face
Model and Illumination Model, ICNC 2005, Springer, p. 7.
Liu, S., Jin, X., Wang, C. & Hui, K. (2007). Ellipsoidal-blob approximation of 3D models and
its applications, Computers & Graphics 31(2): 243251.
Llonch, R. S., Kokiopoulou, E., Tosic, I. & Frossard, P. (2009). 3D face recognition with sparse
spherical representations, Pattern Recognition .
Lu, X. & Jain, A. K. (2005). Deformation analysis for 3d face matching, WACV-MOTION
05: Proceedings of the Seventh IEEE Workshops on Application of Computer Vision
(WACV/MOTION05) - Volume 1, IEEE Computer Society, Washington, DC, USA,
pp. 99104.
Lu, X. & Jain, A. K. (2006). Deformation modeling for robust 3D face matching, CVPR 06:
Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, IEEE Computer Society, Washington, DC, USA, pp. 13771383.
Marr, D. (1982). Vision: a computational investigation into the human representation and processing
of visual information, W. H. Freeman, San Francisco.
Maurer, T., Guigonis, D., Maslov, I., Pesenti, B., Tsaregorodtsev, A., West, D. & Medioni, G.
(2005). Performance of Geometrix ActiveID
TM
3D face recognition engine on the
FRGC data, CVPR 05: Proceedings of the 2005 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR05) - Workshops, IEEE Computer Soci-
ety, Washington, DC, USA, p. 154.
Metaxas, D. N. & Kakadiaris, I. A. (2002). Elastically adaptive deformable models, IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 24(10): 13101321.
Mousa, M.-H., Chaine, R., Akkouche, S. & Galin, E. (2007). Efcient spherical harmonics rep-
resentation of 3d objects, PG 07: Proceedings of the 15th Pacic Conference on Computer
Graphics and Applications, IEEE Computer Society, Washington, DC, USA, pp. 248
255.
Mpiperis, I., Malasiotis, S. & Strintzis, M. G. (2007). 3-D face recognition with the geodesic
polar representation, IEEE Transactions on Information Forensics and Security 2(3): 537
547.
Mpiperis, I., Malassiotis, S. & Strintzis, M. G. (2008). Bilinear models for 3-D face and fa-
cial expression recognition., IEEE Transactions on Information Forensics and Security
3(3): 498511.
Muraki, S. (1991). Volumetric shape description of range data using blobby model, SIG-
GRAPH Comput. Graph. 25(4): 227235.
Nealen, A., Mueller, M., Keiser, R., Boxerman, E. & Carlson, M. (2006). Physically based
deformable models in computer graphics, Computer Graphics Forum 25(4): 809836.
URL: http://dx.doi.org/10.1111/j.1467-8659.2006.01000.x
Nishita, T. & Nakamae, E. (1994). A method for displaying metaballs by using bezier clipping,
Computer Graphics Forum 13: 271280.
Osher, S. & Fedkiw, R. (2003). Level Set Methods and Dynamic Implicit Surfaces, Vol. 153 of
Applied Mathematical Sciences, Springer-Verlag New York.
Parzen, E. (1962). On estimation of a probability density function and mode, The Annals of
Mathematical Statistics 33(3): 10651076.
URL: http://www.jstor.org/stable/2237880
Pauly, M., Keiser, R. & Gross, M. (2003). Multi-scale feature extraction on point-sampled
surfaces, Computer Graphics Forum, Vol. 22, Blackwell Publishing, Inc, pp. 281289.
Pears, N. (2008). Rbf shape histograms and their application to 3d face processing, pp. 18.
Pears, N. & Heseltine, T. (2006). Isoradius contours: New representations and techniques for
3D face registration and matching, 3DPVT06: Proceedings of the Third International
Symposium on 3D Data Processing, Visualization, and Transmission, IEEE Computer So-
ciety, Washington, DC, USA, pp. 176183.
Perakis, P., Passalis, G., Theoharis, T., Toderici, G. & Kakadiaris, I. (2009). Partial matching of
interpose 3d facial data for face recognition, BTAS 2009: Proceedings of the IEEE Third
International Conference on Biometrics: Theory, Applications and Systems, Washington
DC.
Phillips, P., Grother, P., Micheals, R., Blackburn, D., Tabassi, E. & Bone, J. (2003). FRVT 2002:
Evaluation report, Technical Report NISTIR 6965, NIST.
Rusinkiewicz, S. &Levoy, M. (2001). Efcient variants of the ICP algorithm, 3DIM 01: Proceed-
ings of the Third International Conference on 3D Digital Imaging and Modeling, pp. 145
152.
Russ, T., Boehnen, C. & Peters, T. (2006). 3D face recognition using 3D alignment for PCA,
CVPR 06: Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, IEEE Computer Society, Washington, DC, USA, pp. 1391
1398.
Samir, C., Srivastava, A. & Daoudi, M. (2006). Three-dimensional face recognition using
shapes of facial curves, IEEE Transactions on Pattern Analysis and Machine Intelligence
28(11): 18581863.
Sigg, C. (2006). Representation and Rendering of Implicit Surfaces, PhD thesis, ETH Zrich.
Silverman, B. (1986). Density estimation for statistics and data analysis, Chapman & Hall/CRC.
Smeets, D., Fabry, T., Hermans, J., Vandermeulen, D. & Suetens, P. (2009). Isometric deforma-
tion modeling using singular value decomposition for 3d expression-invariant face
recognition, Biometrics: Theory, Applications, and Systems, 2009. BTAS 09. IEEE 3rd
International Conference on, pp. 16.
Sukumar, N. (2005). Maximum entropy approximation, Bayesian Inference and Maximum En-
tropy Methods in Science and Engineering 803: 337344.
ter Haar, F. B. & Veltkamp, R. C. (2008). SHREC08 entry: 3D face recognition using facial con-
tour curves, SMI 08: Proceedings of the IEEE International Conference on Shape Modeling
and Applications, Stony Brook, NY, USA, pp. 259260.
Tsin, Y. & Kanade, T. (2004). A correlation-based approach to robust point set registration,
ECCV (3), pp. 558569.
Vandermeulen, D., Claes, P., Loeckx, D., De Greef, S., Willems, G. & Suetens, P. (2006). Com-
puterized craniofacial reconstruction using CT-derived implicit surface representa-
tions, Forensic Science International (159): S164S174.
Varshosaz, M., Helali, H. & Shojaee, D. (2005). The methods of triangulation, Map Middle East
05: Proceedings of the 1st Annual Middle East Conference and Exhibition on Geospatial
Information, Technology and Applications, Dubai, UAE.
Surface representations for 3D face recognition 293
Levoy, M. & Whitted, T. (1985). The use of points as a display primitive, Technical report,
Department of Computer Science, University of North Carolina at Chappel Hill.
Li, L., Xu, C., Tang, W. & Zhong, C. (2008). 3d face recognition by constructing deformation
invariant image, Pattern Recognition Letters 29(10): 15961602.
Li, S. & Liu, W. (2004). Meshfree particle methods, Springer.
Li, X. & Zhang, H. (2007). Adapting geometric attributes for expression-invariant 3D face
recognition, SMI 07: Proceedings of the IEEE International Conference on Shape Modeling
and Applications, IEEE Computer Society, Washington, DC, USA, pp. 2132.
Liu, D., Shen, L. & Lam, K. (2005). Image Synthesis and Face Recognition Based on 3D Face
Model and Illumination Model, ICNC 2005, Springer, p. 7.
Liu, S., Jin, X., Wang, C. & Hui, K. (2007). Ellipsoidal-blob approximation of 3D models and
its applications, Computers & Graphics 31(2): 243251.
Llonch, R. S., Kokiopoulou, E., Tosic, I. & Frossard, P. (2009). 3D face recognition with sparse
spherical representations, Pattern Recognition .
Lu, X. & Jain, A. K. (2005). Deformation analysis for 3d face matching, WACV-MOTION
05: Proceedings of the Seventh IEEE Workshops on Application of Computer Vision
(WACV/MOTION05) - Volume 1, IEEE Computer Society, Washington, DC, USA,
pp. 99104.
Lu, X. & Jain, A. K. (2006). Deformation modeling for robust 3D face matching, CVPR 06:
Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, IEEE Computer Society, Washington, DC, USA, pp. 13771383.
Marr, D. (1982). Vision: a computational investigation into the human representation and processing
of visual information, W. H. Freeman, San Francisco.
Maurer, T., Guigonis, D., Maslov, I., Pesenti, B., Tsaregorodtsev, A., West, D. & Medioni, G.
(2005). Performance of Geometrix ActiveID
TM
3D face recognition engine on the
FRGC data, CVPR 05: Proceedings of the 2005 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR05) - Workshops, IEEE Computer Soci-
ety, Washington, DC, USA, p. 154.
Metaxas, D. N. & Kakadiaris, I. A. (2002). Elastically adaptive deformable models, IEEE Trans-
actions on Pattern Analysis and Machine Intelligence 24(10): 13101321.
Mousa, M.-H., Chaine, R., Akkouche, S. & Galin, E. (2007). Efcient spherical harmonics rep-
resentation of 3d objects, PG 07: Proceedings of the 15th Pacic Conference on Computer
Graphics and Applications, IEEE Computer Society, Washington, DC, USA, pp. 248
255.
Mpiperis, I., Malasiotis, S. & Strintzis, M. G. (2007). 3-D face recognition with the geodesic
polar representation, IEEE Transactions on Information Forensics and Security 2(3): 537
547.
Mpiperis, I., Malassiotis, S. & Strintzis, M. G. (2008). Bilinear models for 3-D face and fa-
cial expression recognition., IEEE Transactions on Information Forensics and Security
3(3): 498511.
Muraki, S. (1991). Volumetric shape description of range data using blobby model, SIG-
GRAPH Comput. Graph. 25(4): 227235.
Nealen, A., Mueller, M., Keiser, R., Boxerman, E. & Carlson, M. (2006). Physically based
deformable models in computer graphics, Computer Graphics Forum 25(4): 809836.
URL: http://dx.doi.org/10.1111/j.1467-8659.2006.01000.x
Nishita, T. & Nakamae, E. (1994). A method for displaying metaballs by using bezier clipping,
Computer Graphics Forum 13: 271280.
Osher, S. & Fedkiw, R. (2003). Level Set Methods and Dynamic Implicit Surfaces, Vol. 153 of
Applied Mathematical Sciences, Springer-Verlag New York.
Parzen, E. (1962). On estimation of a probability density function and mode, The Annals of
Mathematical Statistics 33(3): 10651076.
URL: http://www.jstor.org/stable/2237880
Pauly, M., Keiser, R. & Gross, M. (2003). Multi-scale feature extraction on point-sampled
surfaces, Computer Graphics Forum, Vol. 22, Blackwell Publishing, Inc, pp. 281289.
Pears, N. (2008). Rbf shape histograms and their application to 3d face processing, pp. 18.
Pears, N. & Heseltine, T. (2006). Isoradius contours: New representations and techniques for
3D face registration and matching, 3DPVT06: Proceedings of the Third International
Symposium on 3D Data Processing, Visualization, and Transmission, IEEE Computer So-
ciety, Washington, DC, USA, pp. 176183.
Perakis, P., Passalis, G., Theoharis, T., Toderici, G. & Kakadiaris, I. (2009). Partial matching of
interpose 3d facial data for face recognition, BTAS 2009: Proceedings of the IEEE Third
International Conference on Biometrics: Theory, Applications and Systems, Washington
DC.
Phillips, P., Grother, P., Micheals, R., Blackburn, D., Tabassi, E. & Bone, J. (2003). FRVT 2002:
Evaluation report, Technical Report NISTIR 6965, NIST.
Rusinkiewicz, S. &Levoy, M. (2001). Efcient variants of the ICP algorithm, 3DIM 01: Proceed-
ings of the Third International Conference on 3D Digital Imaging and Modeling, pp. 145
152.
Russ, T., Boehnen, C. & Peters, T. (2006). 3D face recognition using 3D alignment for PCA,
CVPR 06: Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, IEEE Computer Society, Washington, DC, USA, pp. 1391
1398.
Samir, C., Srivastava, A. & Daoudi, M. (2006). Three-dimensional face recognition using
shapes of facial curves, IEEE Transactions on Pattern Analysis and Machine Intelligence
28(11): 18581863.
Sigg, C. (2006). Representation and Rendering of Implicit Surfaces, PhD thesis, ETH Zrich.
Silverman, B. (1986). Density estimation for statistics and data analysis, Chapman & Hall/CRC.
Smeets, D., Fabry, T., Hermans, J., Vandermeulen, D. & Suetens, P. (2009). Isometric deforma-
tion modeling using singular value decomposition for 3d expression-invariant face
recognition, Biometrics: Theory, Applications, and Systems, 2009. BTAS 09. IEEE 3rd
International Conference on, pp. 16.
Sukumar, N. (2005). Maximum entropy approximation, Bayesian Inference and Maximum En-
tropy Methods in Science and Engineering 803: 337344.
ter Haar, F. B. & Veltkamp, R. C. (2008). SHREC08 entry: 3D face recognition using facial con-
tour curves, SMI 08: Proceedings of the IEEE International Conference on Shape Modeling
and Applications, Stony Brook, NY, USA, pp. 259260.
Tsin, Y. & Kanade, T. (2004). A correlation-based approach to robust point set registration,
ECCV (3), pp. 558569.
Vandermeulen, D., Claes, P., Loeckx, D., De Greef, S., Willems, G. & Suetens, P. (2006). Com-
puterized craniofacial reconstruction using CT-derived implicit surface representa-
tions, Forensic Science International (159): S164S174.
Varshosaz, M., Helali, H. & Shojaee, D. (2005). The methods of triangulation, Map Middle East
05: Proceedings of the 1st Annual Middle East Conference and Exhibition on Geospatial
Information, Technology and Applications, Dubai, UAE.
Face Recognition 294
Veleba, D. &Felkel, P. (2007). Survey of errors in surface representation and their detection and
correction, WSCG 07: Proceedings of the 15th International Conference in Central Europe
on Computer Graphics, Visualization and Computer Vision, Plzen-Bory, Czech Republic.
Walder, C., Schlkopf, B. & Chapelle, O. (2006). Implicit surface modelling with a globally
regularised basis of compact support, Computer Graphics Forum 25(3): 635644.
Wang, F., Vemuri, B., Rangarajan, A. & Eisenschenk, S. (2008). Simultaneous nonrigid registra-
tion of multiple point sets and atlas construction, IEEE Transactions onPattern Analysis
and Machine Intelligence 30(11): 20112022.
Wang, Y., Pan, G., Wu, Z. & Wang, Y. (2006). Exploring facial expression effects in 3D face
recognition using partial ICP, in P. Narayanan (ed.), Computer Vision

U ACCV 2006,
Vol. 3851 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, pp. 581
590.
Yang, Y., Yong, J., Zhang, H., Paul, J. & Sun, J. (2006). Optimal parameterizations of bzier
surfaces, pp. I: 672681.
Yano, K. & Harada, K. (2007). Single-patch nurbs face, Signal-Image Technologies and Internet-
Based System, International IEEE Conference on 0: 826831.