Journal of Computational and Applied Mathematics 329 (2018) 294–306

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

A novel local/global approach to spherical parameterization✩

Zhao Wang a,c , Zhongxuan Luo b, *, Jielin Zhang a,d , Emil Saucan e
a
School of Mathematical Sciences, Dalian University of Technology, Dalian, China
b
School of Software, Dalian University of Technology, Dalian, China
c
College of Science, North China University of Science and Technology, Tangshan, China
d
School of Mathematics, Jilin University, Changchun, China
e
Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany

article info a b s t r a c t
Article history: This paper proposes a novel local/global spherical parameterization for the genus-zero
Received 5 September 2016 triangular mesh, which naturally extends the planar approach to the spherical case. In
Received in revised form 28 April 2017 our method, we derive two fitting matrices (conformal and isometric) in 3D space. By
optimizing the so-called spring energy, the spherical results are achieved by solving a
Keywords:
nonlinear system with spherical constraints. Intuitively, it represents the stitching together
Geometric modeling
of the 1-ring patches to form a unit sphere. Moreover, the derivation of the 3D fitting
Spherical parameterization
Stretch operator matrices can also be applied to planar triangles directly, so that we can obtain a class
of novel planar approaches (conformal, isometric, authalic) to the problem of flattening
triangular meshes. In order to enhance robustness of the proposed spherical method, a
stretch operator is introduced for dealing with high-curvature models. Numerical results
demonstrate that our method is simple, efficient and convergent, and it outperforms
several state-of-the-art methods in terms of trading-off the distortions of angle, area and
stretch. Furthermore, it achieves better visualization in texture mapping.
© 2017 Published by Elsevier B.V.

1. Introduction

Recently, spherical parameterization has become a fundamental research topic in computer graphics, and can be widely
used in texture mapping, surface remeshing and mesh morphing. It is equivalent to embedding the connectivity graph of
the original mesh on the sphere, and assigning a 3D position to each of the mesh vertices. In this process, a good spherical
result should have no overlapping of the resulting triangles, and preserve the geometric properties of the original mesh as
much as possible.
In this paper, we focus on the spherical parameterization, and propose a novel local/global spherical parameterization.
Our work is mainly inspired by the planar methods [1–4] and the spherical methods [5,6]. Our methods are based on
the optimization of the spring energy [7], and stitching together the 1-ring patches to form a unit sphere. It is a natural
extension of the planar local/global approach to the spherical case. In order to enhance robustness of our spherical method, an
r-adaptive stretch operator [8] is introduced for dealing with high-curvature models. Consequently, our method has a good
trade-off between the distortions of angle, area and stretch, so that it achieves a better visualization in texture mapping than
those by other methods, see Fig. 1.

✩ The project is supported by National Natural Science Foundation of China (Grant Nos. 61432003, 61572105, 11171052, 61328206)

* Corresponding author.
E-mail address: zxluo@dlut.edu.cn (Z. Luo).

http://dx.doi.org/10.1016/j.cam.2017.05.011
0377-0427/© 2017 Published by Elsevier B.V.
 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306 295

a b c d e

Fig. 1. Texture mapping of different spherical parameterizations. (a) Convex [5]; (b) Conformal [9]; (c) ARAP; (d) ASAP++; (e) ARAP++.

The main contributions of this paper are as follows,

• We devise two novel fitting matrices (isometric and conformal) between two triangles in 3D space, so that the planar
local/global approach [2] can be extended to the spherical case naturally. Moreover, our ARAP further simplifies the
original ARAP [6] by stitching together 3D triangle instead of individual tetrahedron, and the distortions of our ARAP
is better than original ARAP (Sections 4.1–4.3).
• The derivation of 3D fitting matrices can also be applied to the planar triangles, so that we can obtain novel planar
approaches (conformal, isometric, authalic) by substituting the planar fitting matrices into our method (Appendix).
• To deal with high-curvature models, we introduce an r-adaptive stretch operator to our spherical method. This
operator enhances the adaptiveness of the algorithm, and eliminates the influence of overlapping and flipping
(Section 4.4).

2. Related works

Above all, we refer readers to [10] for surveys regarding fundamental theories and methods about spherical parameteri-
zation. Below, we briefly review the major techniques that have close relationships to our work.
Progressive spherical parameterization [11–13] consists of two stages. First, the original mesh is simplified by vertices
removal until only one tetrahedron remains. Second, the removed vertices are inserted back one by one to the progressive
meshes. The essence of these methods is to perform optimization of the global results during the coarse-to-fine processing.
In their spherical convex combination approaches, Gotsman et al. [5] gave a nice relationship between spectral graph
theory and spherical parameterization. Saba et al. [14] provided an efficient method to reduce the number of unknowns, and
obtained the solution of the large non-linear system in [5]. Gu and Yau [15,16] presented the computing conformal structures
based on Hodge theory. Gu et al. [9] found an unique mapping between any two genus-zero manifolds by minimizing the
harmonic energy of the map. Zayer et al. [17] proposed a curvilinear spherical parameterization. Athanasiadis et al. [18]
presented a feature preserving spherical parameterization based on geometrically constrained optimization.
Some methods focus on the metric optimization. Sheffer et al. [19] extended the planar ABF [20] method to the spherical
case. Friedel et al. [21] improved the discrete energy function which blends angle and area measures to reduce global
distortion. Chen et al. [22] presented a framework for spherical parameterization and registration using Euclidean Ricci
flow. Wang et al. [6] proposed a two-step iterative algorithm based on optimizing the discrete Dirichlet energy [23] in 3D
space.
More, there are a variety of other methods. Kent et al. [24] inflated the model with air like balloon. Spherical relaxed
methods [25,26] fixed some vertices on the sphere and relaxed the left vertices with Laplace operator. Haker et al. [27],
Isenburg et al. [28], Kharevych et al. [29], and Hu et al. [30] cut the mesh into two sections, and got the spherical results by
using the stereo mapping from plane to sphere. Zou et al. [31] introduced Lie advection to parameterization. Zhao et al. [32]
presented an authalic method based on the optimal mass transport technique.

3. Preliminaries

In this section, we review two planar approaches and some distorted metrics that are closely related to our work.

3.1. The planar ARAP and ARAP++ approaches

In Fig. 2, S is a 3D mesh, xi and xj are the internal nodes of S; S ∗ is the corresponding mesh, ui and uj are the internal nodes
of S ∗ . Inspired by [33], the local Dirichlet energy [34] over the 1-ring neighborhood is defined as
∑
E(i)Dirichlet = cot(θi,j )∥ui − uj ∥2 .
j∈N(i)
296 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306

Fig. 2. The local 1-ring neighborhood. (a) The 3D 1-ring; (b) The 2D 1-ring.

By optimizing the global Dirichlet energy of the whole mesh, we can obtain the planar ARAP approach [2],
∑( ∑(
cot(θi,j ) + cot(θj,i ) (ui − uj ) = cot(θi,j )L′(i,j) + cot(θj,i )L′(j,i) (xi − xj ),
) )
(1)
j∈N(i) j∈N(i)

where θi,j is the angle opposite the oriented edge (i, j), and L′(i,j) is a 2 × 2 matrix.
Similarly, the local spring energy [7] over the 1-ring neighborhood is defined as
∑
E(i)spring = ωi,j ∥ui − uj ∥2 .
j∈N(i)

Following [1,3,4], by optimizing the global spring energy of the whole mesh, we can obtain another planar approach that
named ARAP++,
∑ ∑ ωi,j (
ωi,j (ui − uj ) = L′(i,j) + L′(j,i) (xi − xj ).
)
2 (2)
j∈N(i) j∈N(i)

There are five classic types of weights ωi,j that can be applied to Eq. (2), namely Uniform [35], Shape-preserving [36], Mean-
value [37], Cotan [38–40], and Intrinsic [33] weights. Just like various spring coefficients corresponding to different spring
energies, we can regard the Dirichlet energy as a special spring energy. Therefore, ARAP is a special case of ARAP++.

3.2. Distortion metrics

Let θ and T denote the angle and facet of S, φ and T ∗ denote the angle and facet of S ∗ , respectively. F denotes the face
number of S. Then, the distortions (angle, area) between S and S ∗ can be defined as
F 3
1 ∑∑⏐
⏐θj,i − φj,i ⏐ ,
⏐
Dangle =
3F
j=1 i=1

⏐
F ⏐
⏐
∑ ⏐ area(Tj ) area(Tj∗ ) ⏐⏐
Darea = ⏐ ∑F − ∑F
∗ ⏐
⏐.
⏐ i=1 area(Ti ) i=1 area(Ti )
j=1

4. Spherical ARAP++ and ASAP++ approaches

In this section, we describe the novel spherical ARAP++ and ASAP++ approaches. Our work is mainly inspired by [2–6].
The algorithm also consists of two main steps: a local phase and a global phase. They are the linear iterative schemes, which
require an initial value to start it off. As shown in Fig. 3, the initial value is obtained by the spherical convex combination
approach [5].
First, we introduce a fitting matrix between two 3D triangles. Given two 3D triangles 11 : 1p1 p2 p3 and 12 : 1q1 q2 q3 , by
using the fitting matrix L, 1p1 p2 p3 can rotate, translate and scale to obtain a transformed triangle that is close to 1q1 q2 q3 ,
where L is a 3 × 3 matrix, the coordinates of pi and qi (i = 1, 2, 3) are 3 × 1 matrices. We discuss two cases (isometric and
conformal) of the fitting matrices in the following.

4.1. The isometric matrix

In 1p1 p2 p3 and 1q1 q2 q3 , we set

e1 = p1 − p2 , e2 = p2 − p3 , e3 = p3 − p1 ,
{
e′1 = q1 − q2 , e′2 = q2 − q3 , e′3 = q3 − q1 ,
 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306 297

Fig. 3. Local processing. (a) Original model; (b) Initial value.

and the isometric matrix: LI = R. Then, the minimal distance between the transformed triangle and 1q1 q2 q3 is gained by
3 3
∑ ∑
d(11 , 12 ) = ∥e′i − Rei ∥2 = (e′i − Rei )T (e′i − Rei )
i=1 i=1
3
∑ T T
= (e′i e′i − 2e′i Rei + ei T ei ),
i=1

where R R = I. Following [1], this is solved as R = Vi UiT , where Ui , Vi can be obtained by the signed SVD factorization of the
T

cross-covariance matrix
3
∑ T
S= ei e′i = Ui Σi ViT .
i=1

4.2. The conformal matrix

After obtaining the rigid rotation matrix R, we can scale the edges of the transformed triangle to approach 1q1 q2 q3 with
the scale factor a. Set the conformal matrix LC = a · R. Then, the minimal distance between two triangles is gained by
3 3
∑ ∑
d(11 , 12 ) = ∥e′i − a · Rei ∥2 = (e′i − a · Rei )T (e′i − a · Rei )
i=1 i=1
3
∑ T T
= (e′i e′i − 2a · e′i Rei + a2 · e i T e i )
i=1 ( 3 )
3 3
∑ ∑ T ∑ T
= a2 · ei T ei − 2a · Tr R ei e′i + e′i e′i .
i=1 i=1 i=1

In consequence, the minimum of d(11 , 12 ) is obtained in the conformal case, then

( 3
)/ 3
∑ ∑
′T
a = Tr R ei ei ei T ei ,
i=1 i=1
∑3
i=1 ei ei > 0. The derivation of the 3D fitting matrices (isometric and conformal) can also be applied to planar
T
where
triangles, and the planar fitting matrices are obtained. When substituting them to Eqs. (1) and (2), we can obtain the novel
planar methods. The flattened results are equivalent to [2,41]. Appendix gives the computational details of the planar authalic
matrix.

4.3. Algorithm

Based on the planar ARAP++ Eq. (2) and the isometric matrix in 3D, the spherical ARAP++ can be obtained under the
spherical constraints. Inspired by [5], we introduce an auxiliary variable αi to every vertex pi , which can ensure the whole
mesh map on the unit sphere precisely, and achieve valid results as much as possible. Therefore, the planar ARAP++ can be
298 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306

Fig. 4. The spherical results by stitching the individual triangles under ARAP (a) and 1-ring patches under ARAP++ (b).

extended to the spherical case naturally. In order to simplify Eq. (2), we set
−
→ ∑ ωi,j
Bi = (L(i,j,j+1) + L(i,j,j−1) )(pi − pj ),
2
j∈N(i)

where L has been derived in Sections 4.1–4.2. The spherical ARAP++ can be represented as
qi ∥ = 1, (i = 1, . . . , n),
⎧
⎨∥∑
−
→ (3)
⎩ ωi,j (αi · qi − qj ) = Bi .
j∈N(i)

When L is the conformal matrix in Eq. (3), we name it spherical ASAP++ approach. Similarly, the planar ARAP Eq. (1) can also
be extended to the spherical case in the same way, and we name it spherical ARAP approach. By comparison of spherical
ARAP and ARAP++, ARAP++ can obtain the spherical results by stitching together the 1-ring patches instead of the individual
triangles of ARAP, see Fig. 4. By the analysis of energy optimization in Section 3.1, spherical ARAP can also be regarded as a
special case of spherical ARAP++. Eq. (3) is a nonlinear system which contains 4n quadratic equations on the 3n positions of
the qi and the n auxiliary variables αi . To simplify the formula, we set
−
→
qi = (xi , yi , zi )T , Bi = (ai , bi , ci )T ,

thus Eq. (3) can be represented as

xi + y2i + zi2 = 1,
⎧ 2
⎪
⎪
⎪ ∑
ωi,j (αi xi − xj ) = ai ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨j ∈ N(i)
∑
ωi,j (αi yi − yj ) = bi , (4)
⎪
⎪
⎪
⎪j∈N(i)
⎪
⎪ ∑
ωi,j (αi zi − zj ) = ci .
⎪
⎪
⎪
⎪
⎩
j∈N(i)

According to the Newton method, Eq. (4) can be transformed into the following linear system:
⎧ (k) (k+1)
⎪
⎪xi xi + y(k)
i yi
(k+1)
+ zi(k) zi(k+1) = 1,
⎪ ∑
(k) (k+1) (k) (k+1)
α α ωi,j x(k +1)
= αi(k) x(k)
i + ai ,
(k)
⎪
x + x −
⎪
⎪
⎪
⎪ i i i i j
⎪
⎪
⎨ j∈N(i)

(5)
(k) (k+1) (k) (k+1)
∑
⎪yi αi + αi yi − ωi,j y(k
j
+1)
= αi(k) y(k)
i + bi ,
(k)
⎪
⎪
⎪
⎪ j∈N(i)
⎪ ∑
(k) (k+1) (k) (k+1)
z α + αi zi ωi,j zj(k+1) = αi(k) zi(k) + ci(k) ,
⎪
⎪ −
⎩i i
⎪
⎪
j∈N(i)

which can be solved within a small number of iterations. The initial values are
(0)
qi = (x(0) (0) (0)
i , yi , zi ), αi(0) = 1.
−
→(k) (k) (k) (k) (k) (k)
The vectors Bi = (ai , bi , ci ) are obtained from the original nodes pi and the spherical nodes qi .
 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306 299

Fig. 5. The influences of initial values (Hierarchical [13] (above) and Convex [5] (below)) to spherical ARAP++ during 1–3 iterations. (a) Original model; (b)
Initial guess; (c) 1 iteration; (d) 2 iteration; (e) 3 iteration.

Fig. 6. Numerical tests for convergence. (a) ASAP++; (b) ARAP++.

In our method, the final results always rely on the initial values, so that there might be various iterative results according
to different initial values. In Fig. 5, the initial values are provided by [13] (above) and [5] (below), the following three spherical
results on the right are produced by spherical ARAP++ during 3 iterations of Eq. (5), where ωi,j represents Mean-value weight,
and L is the isometric matrix. In other words, the local phase relies on a good initialization, while the global phase is affected
by different initializations, please refer to [42].
Empirically, our methods [ASAP++,∑ ARAP++] appear to converge with the increase of iterations. To all the considered
n
models, the global energy W (S ∗ ) = i=1 E(i) always decreases at each step, and finally stabilizes at a minimum. As a
termination of our algorithm, the numerical results suggest that the maximum iterations are less than 5, see Fig. 6.

4.4. The spherical stretch operator

Inspired by [4], we can select an r-adaptive stretch operator [8] to improve Eq. (3), which can reduce the area and stretch
distortions for high-curvature models. The novel stretch operator is suitable for spherical parameterization, and it is more
pretty than the previous one [43]. The stretch operator is an area-based error monitor, which can reduce flipping and
overlapping of the spherical ARAP++. The expression are as follows,

⎨e(i,j,j−1) = A(qi , qj , qj−1 )/A(pi , pj , pj−1 ),

⎧
e(i,j,j+1) = A(qi , qj , qj+1 )/A(pi , pj , pj+1 ),
e(i,j) = (e(i,j,j−1) + e(i,j,j+1) )/2,
⎩ Area

where A(pi , pj , pk ) denotes the area of 1pi pj pk . We can improve spherical ARAP++ by substituting the stretch operator into
Eq. (3), and obtain the following formulas:
qi ∥ = 1, (i = 1, . . . , n),
⎧
⎨∥∑
θ −
→ (6)
⎩ (eArea
(i,j) ) ωi,j (αi · qi − qj ) = Bi .
j∈N(i)
300 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306

a b c d

Fig. 7. The spherical stretch operator with different values of the operator θ . (a) Original model; (b) θ = 0; (c) θ = −0.3; (d) θ = −0.5.

Eq. (6) can also be solved by the Newton method, which requires only one step of iteration. In order to control the stretch
distortion, the exponents θ can be adjusted according to different models. Fig. 7 shows the spherical results with different
θ , where ωi,j represents Mean-value weight, and L is the isometric matrix.

4.5. An overview of our algorithm

We outline in Algorithm 1, the method detailed in the preceding sections.

Algorithm 1 Spherical ASAP++/ARAP++
1: Input: A 3D mesh S
2: Output: A spherical mesh S ∗
3: S ∗ = S0∗ ; // Initial phase.
4: loop
5: // Local phase
6: for i = 1 : n do
7: △pi pj pj+1 ← RingPatches(pi );
8: △qi qj qj+1 ← RingPatches(qi );
9: ωi,j ← ComputeWeights(RingPatches(pi ));
10: L(i,j,j+1) ← FittingMatrix(△pi pj pj+1 , △qi qj qj+1 );
11: end for
12: // Global phase
13: S1∗ ← ComputePara(ω, L);
14: if d(S1∗ − S ∗ ) > ε , then
15: S0∗ = S1∗ ;
16: else
17: S ∗ = S1∗ ;
18: break;
19: end if
20: end loop

5. Results and comparisons

All the experiments are tested under MATLAB on a i7 Eight-core, 3.6 GHz CPU computer with 32G RAM. To confirm the
effectiveness of our method, we carry on numerical experiments on several typical models.

5.1. The comparison of five classic weights

In this subsection, we compare the spherical results for different types of the weights ωi,j in Eq. (3), more precisely ωi,j
are Uniform, Shape-preserving, Mean-value, Cotan and Intrinsic. In Fig. 8, for the models of Face and David, the ASAP++
and ARAP++ are used, respectively. For contrast, all displayed results are iterated 3 times with same initial values. Table 1
illustrates the angle and area distortions corresponding to the five class weights, showing that Mean-value weight performs
better than others in the angle and area distortions. Therefore, we can ascertain that the Mean-value weight is the best choice
in the following experiments.

5.2. Comparisons of original ARAP and our ARAP

In this subsection, we compare the distortion measures (angle, area), running time and texture mapping of original
ARAP [6] and our ARAP, where our ARAP have been mentioned in Section 4.3. The two methods try to achieve the similar goal,
but have different geometric meanings. They obtain the spherical results by stitching together the individual tetrahedron
(original ARAP) and 3D triangle (our ARAP), respectively. In comparison, we adopt two typical models (Hand, Max-Planck),
 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306 301

Fig. 8. The comparison of five classic weights: Face (ASAP++), David (ARAP++). (a) Original model; (b) Uniform; (c) Shape-preserving; (d) Mean-value;
(e) Cotan; (f) Intrinsic.

Table 1
Comparison of five classic weights on two standard models.
Method Face (V:2892, F:5780) David (V:24382, F:48760)
Angle Area Angle Area
Uniform 0.111 0.341 0.177 0.219
Shape-preserving 0.102 0.328 0.176 0.209
Mean-value 0.091 0.319 0.159 0.199
Cotan 0.097 0.323 0.166 0.208
Intrinsic 0.098 0.325 0.176 0.218

Table 2
Comparison of distortion measures (angle, area) and running time between
original ARAP and our ARAP.
Method n Hand(V:1197, F:2390) Max-Planck(V:7684, F:15364)
Angle Area Time Angle Area Time
1 0.131 0.459 0.86 0.126 0.337 16.7
Original ARAP 2 0.105 0.411 1.85 0.094 0.325 33.5
3 0.103 0.402 2.79 0.073 0.282 50.1
1 0.125 0.429 0.54 0.118 0.313 14.1
Our ARAP 2 0.101 0.389 1.02 0.089 0.282 28.5
3 0.099 0.369 1.56 0.062 0.267 42.8

and computed the spherical results during 3 iterations. Table 2 reveals that our ARAP performs slightly better than original
ARAP in angle and area distortions. Compared with original ARAP, our method needs only one step computation in every
iteration while the original ARAP needs two steps computations as well, thus our ARAP has a lower computational complexity
and running time. As shown in Fig. 9, the textured models produced by our ARAP are more uniform and better visualization
than original ARAP. Therefore, our ARAP is superior to original ARAP base on the above analysis and comparisons.

5.3. Comparisons of spherical results and texture mappings of several popular methods

In this subsection, we perform some comparisons of spherical results and texture mappings on standard models. Our
methods [ARAP, ASAP++, ARAP++] are compared with two state-of-the-art parameterizations, namely Convex [5] and
Conformal [9], where ARAP employs Eq. (1) with the isometric matrix in 3D, ASAP++ and ARAP++ employ Eq. (3), and
ωi,j represents Mean-value weight. Fig. 10 shows the spherical results and texture mappings of the genus-zero models
(Gargoyle, Man, Lion). Table 3 illustrates the distortions (angle, area) and running time of five methods. It reveals that the
angle distortion of ASAP++ is better than other methods, but the area distortion is a little larger. ARAP++ produces the best
results in area distortion, but it renders a slightly larger angle distortion than the other methods. The running time of ARAP,
ASAP++ and ARAP++ is at the same level. They are slightly slower than Convex method, and are faster than the Conformal
method. Convex, Conformal and ASAP++ are conformal approaches which well preserve the angle of checkers. But these
methods obtain inesthetic view results of texture mapping since the checkers are non-uniform size. ARAP and ARAP++
provide good trade-off between the overall angle and area distortions, so that they achieve better visual results than the
above methods from the point of uniform checkers.
302 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306

Fig. 9. Comparison of spherical parameterization (Original ARAP: the first and third rows of (a)–(c); Our ARAP: the second and fourth rows of (a)–(c)) and
texture mapping of two methods. (a) 1 iteration; (b) 2 iteration; (c) 3 iteration; (d) Original ARAP; (e) Our ARAP.

Table 3
The comparison of five methods on three standard models.
Method Gargoyle(V:9001, F:17998) Man(V:6282, F:12560) Lion(V:7302, F:14600)
Angle Area Time Angle Area Time Angle Area Time
Convex 0.127 1.109 48.9 0.101 0.748 26.9 0.131 0.693 29.3
Conformal 0.113 1.075 98.7 0.074 0.709 42.1 0.105 0.622 66.7
ARAP 0.135 0.982 50.3 0.166 0.441 27.8 0.160 0.555 31.0
ASAP++ 0.110 1.062 51.0 0.072 0.722 27.9 0.104 0.652 31.1
ARAP++ 0.127 0.980 50.1 0.158 0.392 27.7 0.154 0.542 30.9

5.4. Comparisons of several popular methods on high-curvature models

In this subsection, we perform some comparisons of spherical results and texture mappings on high-curvature models
(Bird, Homer, Hippo). Our methods [ASAP++, ARAP++] are compared with two classic parameterizations, namely Convex [5]
and Hierarchical [13]. ARAP++ employs Eq. (6), which adjusts the exponent θ to obtain the valid results. In order to perform
the texture mapping of a high-curvature model, we have first got a quadrilateral remeshing of the original model according to
the spherical result. The details of this remeshing method can be found in [44]. For contrast, the remeshed models are given
by corresponding sets of vertices, facets, and topological connections. During the uniform subdivision of the base mesh, we
apply the same number of umbrella operators [44] on the remeshed models. Fig. 11 shows the spherical results and texture
mappings of four methods. Table 4 illustrates the angle, area and stretch (L2 ) [45] distortions about the spherical results. In
comparisons, the Hierarchical method achieves the best results in the area and stretch distortions. ASAP++ is the best one
that performs in the angle distortion. ARAP++ provides a good performance in trading-off the distortions of angle, area and
stretch, so that it is the best choice of the four methods that can lead to a uniform result in texture mapping.

6. Conclusion and outlook

Spherical parameterization is one of the fundamental operations in digital geometry processing. After more than a decade
of researching, many spherical methods have been achieved by the scholars, but there are still many complicated problems
that need to be resolved. In this paper, we presented a novel spherical parameterization which extends the planar ARAP
approach to the spherical case. To deal with high-curvature models, we introduce the stretch operator that has been proved
superior for trading-off the angle, area and stretch distortions, as well as attaining better visualization in texture mapping.
 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306 303

Fig. 10. The spherical results and texture mappings of five methods. (a) Convex; (b) Conformal; (c) ARAP; (d) ASAP++; (e) ARAP++.

Table 4
The comparison of four methods on three high-curvature models.
Method Bird(V:5852, F:11700) Homer(V:5103, F:10202) Hippo(V:9574, F:19144)
Angle Area Stretch Angle Area Stretch Angle Area Stretch
Convex 0.102 1.049 401.1 0.109 0.868 102.7 0.164 0.985 87.2
Hierarchical 0.499 0.333 89.5 0.414 0.227 78.3 0.530 0.295 34.6
ASAP++ 0.082 1.076 433.8 0.107 0.919 117.3 0.159 1.099 108.9
ARAP++ 0.230 0.764 225.2 0.208 0.632 96.1 0.321 0.694 62.2

Our future work will focus on deducing an authalic matrix in spherical ARAP++, and finding a formula that can compute
the exponent of stretch operator adaptively. Another natural direction of study would be to extend our method to the volume
parameterization. These researches will undoubtedly shed some new light on the improvement of computer graphics.
304 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306

Fig. 11. The spherical results and texture mappings on high-curvature models. (a) Convex; (b) Hierarchical; (c) ASAP++; (d) ARAP++.

Acknowledgments

The authors express their thanks to the reviewers for their valuable comments and suggestions that significantly
contributed to the improvement of this paper. Special thanks to Dr. Zhaoliang Meng and Dr. Wanfeng Qi for their many
helpful insights on this topic. Thanks also to Prof. Xianfeng Gu and Prof. Xin Li for providing results of the algorithms [9,13].

Appendix. The planar authalic matrix

According to the derivation in Section 4, we can simply obtain the planar isometric matrix (L′I = R′ ) and the planar
conformal matrix (L′C = a · R′ ), where R′ is a 2 × 2 matrix. Set e′i = (ui , vi )T , R′ ei = (mi , ni )T , then we can also deduce the
 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306 305

Fig. 12. The parameterizations (angle, area) and texture mappings of the novel planar method with three fitting matrices. (a) Conformal (0.109, 0.539); (b)
Isometric (0.133, 0.254); (c) Authalic (0.275, 0.179).

[ ]
b 0
planar authalic matrix (L′A = 0 1/b
· R′ ) as
3  [ ] 2
∑ b 0
d(11 , 12 )
 ′ ′ 

= ei − 0 1/b R ei 

i=1
3 ( [ ] )T ( [ ] )
∑ b 0 b 0
= e′i − R′ ei e′i − R′ ei
0 1/b 0 1/b
i=1
3
∑
(u2i + vi2 ) − 2(ui mi b + vi ni /b) + (m2i b2 + n2i /b2 )
( )
=
i=1
3
∑
(ui − mi b)2 + (vi − ni /b)2 .
( )
=
i=1

Then, we can compute the derivative of d(11 , 12 ) with respect to b to achieve the minimal distance, and arrive at
( 3
) ( 3
) ( 3
) 3
∑ ∑ ∑ ∑
m2i 4
b − mi ui 3
b + ni vi b− n2i = 0.
i=1 i=1 i=1 i=1

The quartic equation can be solved by Ferrari formula, and we can find a positive real root to minimize the object
function d(11 , 12 ). When the three fitting matrices are substituted to the planar method Eq. (2), we achieve the results
of parameterizations and texture mappings in Fig. 12, where ωi,j represents Mean-value weight.

References

[1] O. Sorkine, M. Alexa, As-rigid-as-possible surface modeling, in: Proceedings of Eurographics Symposium on Geometry Processing, 2007, pp. 109–116.
[2] L. Liu, L. Zhang, Y. Xu, C. Gotsman, A local/global approach to mesh parameterization, Comput. Graph. Forum 27 (5) (2008) 495–504. http://dx.doi.org/
10.1111/j.1467-8659.2008.01290.x.
[3] S. Bouaziz, M. Deuss, Y. Schwartzburg, T. Weise, M. Pauly, Shape-Up: Shaping discrete geometry with projections, Comput. Graph. Forum 31 (5) (2012)
1657–1667. http://dx.doi.org/10.1111/j.1467-8659.2012.03171.x.
[4] Z. Wang, Z. Luo, J. Zhang, E. Saucan, ARAP++: An extension of the local/global approach to mesh parameterization, Front. Inf. Technol. Electron. Eng.
17 (6) (2016) 501–515. http://dx.doi.org/10.1631/FITEE.1500184.
[5] C. Gotsman, X. Gu, A. Sheffer, Fundamentals of spherical parameterization for 3D meshes, ACM Trans. Graph. 22 (3) (2003) 358–363. http://dx.doi.org/
10.1145/1201775.882276.
306 Z. Wang et al. / Journal of Computational and Applied Mathematics 329 (2018) 294–306

[6] C. Wang, Z. Liu, L. Liu, As-rigid-as-possible spherical parametrization, Graph. Models 76 (5) (2014) 457–467. http://dx.doi.org/10.1016/j.gmod.2014.
03.016.
[7] H. Hoppe, T. DeRose, T. Duchamp, J. McDonald, W. Stuetzle, Mesh optimization, in: Proceedings of SIGGRAPH, 1993, pp. 16–26. http://dx.doi.org/
10.1145/166117.166119.
[8] R. Zayer, C. Rössl, H. Seidel, R-Adaptive parameterization of surfaces, Technical Report, 2004.
[9] X. Gu, Y. Wang, T. Chan, P. Thompson, S. Yau, Genus zero surface conformal mapping and its application to brain surface mapping, IEEE Trans. Med.
Imaging 23 (8) (2004) 949–958. http://dx.doi.org/10.1109/TMI.2004.831226.
[10] K. Hormann, B. Lévy, A. Sheffer, Mesh parameterization: theory and practice, in: Proceedings of SIGGRAPH, 2007, pp. 1–122.
[11] A. Shapiro, A. Tal, Polyhedron realization for shape transformation, Vis. Comput. 14 (8) (1998) 429–444. http://dx.doi.org/10.1007/s003710050153.
[12] E. Praun, H. Hoppe, Spherical parameterization and remeshing, ACM Trans. Graph. 22 (3) (2003) 340–349. http://dx.doi.org/10.1145/1201775.882276.
[13] S. Wan, T. Ye, M. Li, H. Zhang, X. Li, An efficient spherical mapping algorithm and its application on spherical harmonics, Sci. China Inf. Sci. 56 (9) (2013)
1–10. http://dx.doi.org/10.1007/s11432-013-4992-5.
[14] S. Saba, I. Yavneh, C. Gotsman, A. Sheffer, Practical spherical embedding of manifold triangle meshes, in: Proceedings of Shape Modeling and
Applications, 2005, pp. 256–265. http://dx.doi.org/10.1109/SMI.2005.32.
[15] X. Gu, S. Yau, Computing conformal structures of surfaces, Commun. Inf. Syst. 2 (2) (2002) 121–146. http://dx.doi.org/10.4310/CIS.2002.v2.n2.a2.
[16] X. Gu, S. Yau, Global conformal surface parameterization, in: Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Geometry Processing,
2003, pp. 127–137.
[17] R. Zayer, C. Rössl, H. Seidel, Curvilinear spherical parameterization, in: Proceedings of Shape Modeling and Applications, 2006, pp. 57–64.
http://dx.doi.org/10.1109/SMI.2006.9.
[18] T. Athanasiadis, I. Fudos, C. Nikou, V. Stamati, Feature-based 3D morphing based on geometrically constrained spherical parameterization, Comput.
Aided Geom. Design 29 (1) (2012) 2–17. http://dx.doi.org/10.1016/j.cagd.2011.09.004.
[19] A. Sheffer, C. Gotsman, N. Dyn, Robust spherical parameterization of triangular meshes, in: Proceedings of the 4th Israel-Korea Bi-National Conference
on Geometric Modeling and Computer Graphics, 2003, pp. 94–99. http://dx.doi.org/10.1007/s00607-004-0056-9.
[20] A. Sheffer, E. de Sturler, Parameterization of faceted surfaces for meshing using angle-based flattening, Eng. Comput. 17 (3) (2001) 326–337. http:
//dx.doi.org/10.1007/PL00013391.
[21] I. Friedel, P. Schröder, M. Desbrun, Unconstrained spherical parameterization, in: Proceedings of SIGGRAPH, Technical Sketches, 2005. http://dx.doi.org/
10.1145/1187112.1187274.
[22] X. Chen, H. He, G. Zou, X. Zhang, X. Gu, J. Hua, Ricci flow-based spherical parameterization and surface registration, Comput. Vis. Image Underst. 117 (9)
(2013) 1107–1118. http://dx.doi.org/10.1016/j.cviu.2013.02.010.
[23] I. Chao, U. Pinkall, P. Sanan, P. Schröder, A simple geometric model for elastic deformations, ACM Trans. Graph. 29 (4) (2010) 38. http://dx.doi.org/10.
1145/1778765.1778775.
[24] J. Kent, W. Carlson, R. Parent, Shape transformation for polyhedral objects, ACM SIGGRAPH Comput. Graph. 26 (2) (1992) 47–54.
[25] L. Kobbelt, J. Vorsatz, U. Labsik, H. Seidel, A shrink wrapping approach to remeshing polygonal surfaces, Comput. Graph. Forum 18 (3) (1999) 119–130.
[26] M. Alexa, Merging polyhedral shapes with scattered features, Vis. Comput. 16 (1) (2000) 26–37. http://dx.doi.org/10.1007/PL00007211.
[27] S. Haker, S. Angenent, A. Tannenbaum, R. Kikinis, G. Sapiro, M. Halle, Conformal surface parameterization for texture mapping, IEEE Trans. Vis. Comput.
Graphics 6 (2) (2000) 181–189. http://dx.doi.org/10.1109/2945.856998.
[28] M. Isenburg, S. Gumhold, C. Gotsman, Connectivity shapes, in: Proceedings of IEEE Visualization, 2001, pp. 135–142.
[29] L. Kharevych, B. Springborn, P. Schröder, Discrete conformal mappings via circle patterns, ACM Trans. Graph. 25 (2) (2006) 412–438. http://dx.doi.org/
10.1145/1138450.1138461.
[30] J. Hu, X. Liu, Z. Su, X. Shi, F. Liu, A spherical parameterization approach based on symmetry analysis of triangular meshes, J. Zhejiang Univ. Sci. A 10 (7)
(2009) 1009–1017. http://dx.doi.org/10.1631/jzus.A0820728.
[31] G. Zou, J. Hu, X. Gu, J. Hua, Authalic parameterization of general surfaces using Lie advection, IEEE Trans. Vis. Comput. Graphics 17 (12) (2011) 2005–
2014. http://dx.doi.org/10.1109/TVCG.2011.171.
[32] X. Zhao, Z. Su, X. Gu, A. Kaufman, J. Sun, J. Gao, F. Luo, Area-preservation mapping using optimal mass transport, IEEE Trans. Vis. Comput. Graphics
19 (12) (2013) 2838–2847. http://dx.doi.org/10.1109/TVCG.2013.135.
[33] M. Desbrun, M. Meyer, P. Allize, Intrinsic parameterization of surface meshes, Comput. Graph. Forum 21 (2) (2002) 209–218. http://dx.doi.org/10.
1111/1467-8659.00580.
[34] U. Pinkall, K. Polthier, Computing discrete minimal surface and their conjugates, Experiment. Math. 2 (1) (1993) 15–36. http://dx.doi.org/10.1080/
10586458.1993.10504266.
[35] W. Tutte, How to draw a graph, Proc. Lond. Math. Soc. 13 (3) (1963) 743–768.
[36] M. Floater, Parameterization and smooth approximation of surface triangulations, Comput. Aided Geom. Design 14 (3) (1997) 231–250. http://dx.doi.
org/10.1016/S0167-8396(96)00031-3.
[37] M. Floater, Mean value coordinates, Comput. Aided Geom. Design 20 (1) (2003) 19–27. http://dx.doi.org/10.1016/S0167-8396(03)00002-5.
[38] M. Eck, T. DeRose, T. Duchamp, H. Hoppe, M. Lounsbery, W. Stuetzle, Multiresolution analysis of arbitrary meshes, in: Proceedings of SIGGRAPH, 1995,
pp. 173–182. http://dx.doi.org/10.1145/218380.218440.
[39] A. Bobenko, B. Springborn, A discrete Laplacecbeltrami operator for simplicial surfaces, Discrete Comput. Geom. 38 (4) (2007) 740–756. http:
//dx.doi.org/10.1007/s00454-007-9006-1.
[40] Y. Liu, C. Xu, D. Fan, Y. He, Efficient construction and simplification of Delaunay meshes, ACM Trans. Graph. 34 (6) (2015) 174. http://dx.doi.org/10.
1145/2816795.2818076.
[41] Y. Yoshiyasu, W. Ma, E. Yoshida, F. Kanehiro, As-conformal-as-possible surface registration, Comput. Graph. Forum 33 (5) (2014) 257–267. http:
//dx.doi.org/10.1111/cgf.12451.
[42] Y. Liu, C. Xu, R. Yi, D. Fan, Y. He, Manifold differential evolution (MDE): a global optimization method for geodesic centroidal voronoi tessellations on
meshes, ACM Trans. Graph. 35 (6) (2016) 243. http://dx.doi.org/10.1145/2980179.2982424.
[43] S. Yoshizawa, A. Belyaev, H. Seidel, A fast and simple stretch-minimizing mesh parameterization, in: Proceedings of Shape Modeling International,
2004, pp. 200–208. http://dx.doi.org/10.1109/SMI.2004.1314507.
[44] K. Hormann, G. Greiner, Quadrilateral remeshing, in: Proceedings of Vision, Modeling and Viz, 2000, pp. 153–162.
[45] R. Sander, J. Snyder, S. Gortler, H. Hoppe, Texture mapping progressive meshes, in: Proceedings of SIGGRAPH, 2001, pp. 409–416. http://dx.doi.org/
10.1145/383259.383307.