As Conformal As Possible Discrete Volumetric Mapping
As Conformal As Possible Discrete Volumetric Mapping
As Conformal As Possible Discrete Volumetric Mapping
R
t
D
T
t,i
D
t,j
R
T
t
9V
t
,
where T v
i
,v
j
are the tetrahedra sharing edge fv
i
,v
j
g and T v
i
,v
i
n
t,i
n
t,j
9V
t
:
We then obtain a symmetric linear system that can be solved
with a Conjugate Gradient method with a Jacobi preconditioner,
keeping surface vertices constant. The number of iterations is
usually around 150 for each mesh with a threshold of 10
8
.
Step 3local rotations and coefcients. Using the harmonic map
as a rst approximation, we can compute the local rotation of each
vertex. Using those rotations, we compute the K
ij
coefcients.
Step 4as-conformal-as-possible map. The last step is to com-
pute the volumetric map with our operator by solving Eq. (3)
using a Conjugate Gradient method with a Jacobi preconditioner.
The number of iterations is also usually around 150 for each mesh
with a threshold of 10
8
.
We found that iteratively updating local rotations and the
mapping only leads to epsilon improvements, which is why we
only evaluate local rotations once.
4. Results
This section presents the results of the volumetric maps
obtained using our metric. We also compare with alternative
metrics and discuss the main differences between these metrics.
Note that our aim is to introduce a fast and simple metric that
optimizes conformality. For this reason, we focus our attention on
the behavior of each metric instead of the nal map as a whole.
The method is implemented in Cin the Graphite software [31]
and all tests were conducted on an Intel Xenon E5520 processor
without any GPU acceleration technique. Computation times of the
interior (i.e., without surface map) are shown in Table 1.
The volumetric parameterization times are approximately
multiplied by a factor between three and four using our operator
compared to the Laplace operator alone. It is important to note
that in general, most of the complete parameterization process
is completely absorbed by the surface parameterization as it
uses a Gradient Descent method, which is known to be slow.
A multi-grid method could be used for this part, but this is left for
future work. In fact, the volumetric parameterization times
represent generally much less than 1% of that of the surface,
which thus far outweighs the increase in time of performing as-
conformal-as-possible mapping. Considering the minimal cost
and the fact that results are always at least somewhat better, it
is worth taking the extra step toward conformality.
Table 1
Quantitative results for different meshes. Timings are in seconds and the optimal minimal value for both energy measures is 2.
Mesh 9V9 9T 9 Codomain Harmonic ACAP
Time E
angle
E
volume
Time E
angle
E
volume
Sphere 10k 53k Box 0.18 2.246 2.096 0.89 2.226 2.060
Torus 15k 59k Polycube 0.12 2.424 2.392 0.58 2.400 2.331
Squirrel 18k 73k Polycube 0.15 3.118 3.453 0.84 2.638 3.113
Bust 18k 77k Polycube 0.19 2.829 3.526 0.87 2.660 3.447
Foot 20k 89k Polycube 0.19 2.441 2.455 0.98 2.372 2.384
Pensatore 30k 150k Box 0.47 2.554 2.488 2.03 2.441 2.288
Lion Vase 40k 163k Polycube 0.44 3.324 3.229 2.09 2.869 3.216
Grog 46k 197k Polycube 0.51 3.521 3.600 2.30 3.174 3.461
Bumpy Sphere 66k 307k Sphere 2.08 2.003 2.011 5.45 2.003 2.007
Box 2.01 2.283 2.126 6.53 2.252 2.077
Alien Shark 185k 850k Polycube 5.16 2.697 2.783 17.48 2.538 2.568
Human Elder 188k 877k Polycube 5.68 2.598 2.417 30.59 2.524 2.298
Bimba Con Nastrino 203k 969k Polycube 6.36 2.825 5.912 23.14 2.486 5.443
Chinese Lion 273k 1203k Polycube 9.37 2.964 3.105 34.71 2.636 3.079
Kitten 342k 1968k Polycube 23.72 2.668 4.073 74.57 2.530 3.655
Human Torso 389k 1863k Polycube 16.33 3.056 3.910 53.91 2.571 3.569
Mudface 1183k 5569k Polycube 48.12 2.634 3.178 154.52 2.469 2.802
.
Fig. 4. Results of (left) harmonic maps and (right) as-conformal-as-possible maps.
G.-P. Paille, P. Poulin / Computers & Graphics 36 (2012) 427433 430
From a qualitative point of view, we can note some visual
indicators that show improvements. While harmonic maps have
very good behavior in the center of the object, quality at the
borders is often poor. It is shown in the bumpy sphere of Fig. 1
and it is also noticeable in Fig. 4. As a compromise, improvements
at the borders come at the price of slightly reducing the quality of
the interior, but nonetheless the overall quality is improved and
the interior has still a good behavior.
To quantitatively measure improvements of the mapping, we
use an angle and a volume energy measure based on equations
found in the book by Botsch et al. [3] and summarized by
Solomon et al. [32]. Adapted to volumes, we obtain
E
angle
1
V
X
t AT
V
t
s
t,min
s
t,max
s
t,max
s
t,min
,
E
volume
1
V
X
t AT
V
t
s
t,1
s
t,2
s
t,3
1
s
t,1
s
t,2
s
t,3
,
where V is the total volume of the mesh, s
t,i
are the singular
values of the Jacobian matrix of the tetrahedra transformation,
and s
t,min
and s
t,max
are respectively the minimum and maximum
singular values.
These measures can be numerically unstable near singulari-
ties. For this reason, we decided that energy values higher than a
xed threshold should be ignored, due to the bias that they
introduce. Even though ignoring values also introduces a bias, the
result is still much more representative of the total energy than
incorporating unstable data. These extremal values represent
generally less than 0.3% of the total volume, which renders the
bias negligible when computing global map energies.
From these measures, we can see an improvement on both
angle and volume global energies for all tested meshes, as
summarized in Table 1. Although not shown in the table, energy
variance is also signicantly decreased on both measures. Fig. 5
shows a color-coded energy distribution of a planar cut through
the Chinese Lion model along with a histogram of this distribu-
tion for the entire tetrahedral mesh.
Note that bijectivity is not guaranteed for large deformations. In
fact, linear constraints do not generally lead to barrier metrics,
which means that applying enough force on an element will
eventually invert it. For this reason, codomains should be chosen
carefully. While polycubes in this paper are built manually, Gregson
et al. [20] provide an automatic geometry-friendly polycube con-
struction method that could be used to minimize element inversion.
4.1. Alternative metrics
Other metrics can also optimize conformality in higher dimen-
sions. Because the Jacobian matrix of a conformal map should be a
scaled rotation, the following energy functional could be used
E
t
1
2
Z
t
JJsRJ
2
dV
t
,
where s 9J9
1=3
.
2
E
angle
6
Fig. 5. Angle energy measure for the Chinese Lion mapped to a polycube. A planar cut with color-coded energy and the associated deformation distribution for (left)
the harmonic map and (right) the as-conformal-as-possible map. Note that 86% of the volume has an energy value below 3 for our map compared to 76% for the harmonic
map. (c) Harmonic, (d) ACAP. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)
G.-P. Paille, P. Poulin / Computers & Graphics 36 (2012) 427433 431
We found that the results with this metric are better than
harmonic maps, but not better than ACAP maps, as shown in
Table 2. An average of four iterations is needed to converge for
tested meshes.
The metric developed by Gregson et al. [20] could also be
considered
~ v
i
1
N
X
N
j 1
~ v
j
1
N
X
N
j 1
R
i
R
j
2
v
i
v
j
,
where N 9Nv
i
9 and ~ v
i
is the mapped vertex position.
Unfortunately, this operator is highly dependent on the local
topology of the vertex and less on the geometry of the volume, in
the same vein as massspring systems.
Li et al. [26] developed a metric trying to improve orthogon-
ality, but it has a preferred direction. While it is appropriate for
their application, it is an unwanted behavior for general volu-
metric mappings.
Cage-based deformations could also be considered as potential
mapping metrics. The surface map acting as the cage deformation,
interior points are moved accordingly to the formulation of the
method. Green coordinates [27], being not interpolatory, lead to
highly distorted elements near boundaries since vertices can
stand outside the cage after deformation. Harmonic maps being
uniquely dened by their boundary values, Harmonic coordinates
[33] were not considered here. Finally, mean value coordinates
(MVC) [34] gave slightly better results than harmonic maps
regarding energy measures on the tested meshes, as shown in
Table 2. However, the computation time is prohibitive for large
meshes and the negative value property of MVC can lead to high
distortion in concave parts.
4.2. Uniform scaling
We can also introduce a parameter to operator D to add a degree
of freedom to the results. A logical parameterization would be to
separate the uniform scaling constraint from the orthogonality
constraint using a weighting factor oA0; 1, noting o 1o.
Explaining only for the discrete case, we redene the matrix of Eq.
(2) to be
D
i,o
2
0 on
i,y
on
i,z
0 on
i,z
on
i,y
on
i,x
0 on
i,z
on
i,z
0 on
i,x
on
i,x
on
i,y
0
on
i,y
on
i,x
0
2
6
6
6
6
6
6
6
6
6
4
3
7
7
7
7
7
7
7
7
7
5
,
while all other equations remain the same.
From experiments, we found that oA0:3,0:7 gives best
results, o0:5 being the original operator. Extremal values of
o must be avoided because singularities of the surface map may
cause degeneracies in the volumetric map. Fig. 6 shows mappings
computed for different values of o.
5. Conclusion and future work
We introduced an operator that minimizes a 3D conformal
energy similar to the CauchyRiemann equations in 2D. We
showed that this operator can be parameterized such that uni-
form scaling and orthogonality constraints can be weighted as
desired. In fact, any system of constraints using only rst-order
derivatives can be used with the same derivations. The general-
ization in any dimension is also trivial.
The derivations are based on a mathematically sound
approach. Although the results might not always visually appear
signicantly improved, they are consistently better than not using
our metric, and at a cost of less than 1% of the total computations.
Like most other methods, we rely on a conformal surface map.
Our polycube map helps to obtain better results, but as shown in
Fig. 7, simple polycubes prove sufcient to achieve reasonable
surface map.
Several methods could benet from our work. Indeed, most
papers that use volumetric harmonic maps as a central technique
can be adapted without much difculty to our maps. This is due
to the fact that the core concepts of the mapping remain
unchanged, while adding more generality and exibility to these
methods.
Our operator shares the same limitations than the Laplace
operator, i.e., that the one-to-one mapping and the convergence
are not guaranteed for meshes that lead to ill-conditioned linear
systems.
In the future, we will look for a rotational-invariant operator,
hopefully linear, thus eliminating the need for harmonic map
Table 2
Angle energy measures of alternative metrics for three meshes. Similar results
were obtained with all tested meshes and the volume measure E
volume
. Compared
metrics are the scaled rotated Jacobian matrix (SRJ), the metric fromGregson et al. [20],
the metric from Li et al. [26], and mean value coordinates [34].
Mesh SRJ [GSZ11] [LLT11] MVC
Sphere 2.242 2.258 2.234 2.245
Pensatore 2.493 2.560 2.638 2.442
Alien Shark 2.637 2.584 2.746 2.585
Fig. 6. Uniform scale optimization for the sphere parameterized to a box:
(a) o0:3; (b) o0:45; (c) o0:6; and o0:7.
Fig. 7. As-conformal-as-possible maps of (a) Foot, (b) Bust, (c) Pensatore, and (d) Human Torso. (e) Codomains of each object.
G.-P. Paille, P. Poulin / Computers & Graphics 36 (2012) 427433 432
pre-processing. We also aim at letting the boundary evolve on the
codomain surface as the interior is optimized. Such method
would give more room for optimization and could lead to a much
better energy minimization. However, this generalization is not
straightforward and would need nonlinear optimization.
As parameterization can be seen as a deformation of a shape,
we believe that the presented method can have several applica-
tions in animation, notably for cage-based, skeleton-based, and
point-based deformations.
Acknowledgments
The models are courtesy of Nick Zuccarello (Alien Shark,
Human Elder, Mudface) and Aim@Shape shape repository. This
work was supported in parts by grants from FQRNT and NSERC.
Appendix A. Supplementary material
Supplementary data associated with this article can be found in
the online version of http://dx.doi.org/10.1016/j.cag.2012.03.014.
References
[1] Hormann K, Polthier K, Sheffer A. Mesh parameterization: theory and
practice. In: SIGGRAPH Asia course notes; 2008.
[2] Martin T, Cohen E, Kirby RM. Volumetric parameterization and trivariate
B-spline tting using harmonic functions. Comput Aided Geom Des
2009;26(6):64864.
[3] Botsch M, Kobbelt L, Pauly M, Alliez P, Le vy B. Polygon mesh processing. AK
Peters; 2010.
[4] Floater MS, Hormann K. Surface parameterization: a tutorial and survey. In:
Advances in multiresolution for geometric modeling. Springer; 2005. p. 15786.
[5] Sheffer A, Praun E, Rose K. Mesh parameterization methods and their
applications. Found Trends Comput Graph Vis 2006;2:10571.
[6] Hormann K, Le vy B, Sheffer A. Mesh parameterization: theory and practice.
In: ACM SIGGRAPH course notes; 2007.
[7] Wang Y, Gu XD, Yau S-T. Volumetric harmonic map. Commun Inf Syst
2003;3(3):192202.
[8] Wang Y, Gu XD, Chan T, Thompson P, Yau S-T. Volumetric harmonic brain
mapping. In: IEEE international symposium on biomedical imaging: nano to
macro, 2004, vol. 2; 2004. p. 12758.
[9] Li X, Guo X, Wang H, He Y, Gu XD, Qin H. Harmonic volumetric mapping for
solid modeling applications. In: ACM symposium on solid and physical
modeling, SPM 07; 2007. p. 10920.
[10] Li X, Guo X, Wang H, He Y, Gu XD, Qin H. Meshless harmonic volumetric
mapping using fundamental solution methods. IEEE Trans Autom Sci Eng
2009;6(3):40922.
[11] Li X, Xu H, Wan S, Yin Z, Yu W. Feature-aligned harmonic volumetric
mapping using MFS. Comput Graph 2010;34(3):24251 [Shape modeling
international (SMI) conf; 2010].
[12] Li X. Nondegeneracy of harmonic volumetric parameterization on star-
shaped domains. Technical report. Electrical and Computer Engineering
Department, Louisiana State University; 2010.
[13] Xia J, He Y, Han S, Fu C-W, Luo F, Gu XD. Parameterization of star-shaped
volumes using Greens functions. In: Advances in geometric modeling and
processing, Lecture notes in computer science, vol. 6130. Springer; 2010. p.
21935.
[14] Xia J, He Y, Yin X, Han S, Gu XD. Direct-product volumetric parameterization
of handlebodies via harmonic elds. In: Shape modeling international conf
(SMI); 2010. p. 312.
[15] Tarini M, Hormann K, Cignoni P, Montani C. Polycube-maps. In: ACM
SIGGRAPH 2004 papers; 2004. p. 85360.
[16] Martin T, Cohen E. Volumetric parameterization of complex objects by
respecting multiple materials. Comput Graph 2010;34(3):18797.
[17] Li B, Li X, Wang K, Qin H. Generalized polycube trivariate splines. In: Shape
modeling international conf (SMI); 2010. p. 2615.
[18] Wang J, Zhao S, Tong X, Lin S, Lin Z, Dong Y, et al. Modeling and rendering of
heterogeneous translucent materials using the diffusion equation. ACM Trans
Graph 2008;27(1):9:118.
[19] Yin X, Jin M, Luo F, Gu XD. Discrete curvature ows for surfaces and
3-manifolds. In: Emerging trends in visual computing. Springer-Verlag;
2009. p. 3874.
[20] Gregson J, Sheffer A, Zhang E. All-hex mesh generation via volumetric
polycube deformation. Comput Graph Forum 2011;30(5):140716.
[21] Nieser M, Reitebuch U, Polthier K. CubeCoverparameterization of 3D
volumes. Comput Graph Forum 2011;30(5):1397406.
[22] K alberer F, Nieser M, Polthier K. QuadCoversurface parameterization using
branched coverings. Comput Graph Forum 2007;26(3):37584.
[23] Le vy B, Liu Y. Lp Centroidal Voronoi Tessellation and its applications. ACM
Trans Graph 2010;29(4) 119:111.
[24] Han S, Xia J, He Y. Hexahedral shell mesh construction via volumetric
polycube map. In: ACM symposium on solid and physical modeling, SPM
10; 2010. p. 12736.
[25] Zeng W, Marino J, Kaufman A, Gu XD. Volumetric colon wall unfolding using
harmonic differentials. Comput Graph 2011;35(3):72632 [Shape modeling
international (SMI) conf; 2011].
[26] Li M-F, Liao S-H, Tong R-F. Facial hexahedral mesh transferring by volumetric
mapping based on harmonic elds. Comput Graph 2011;35(1):928.
[27] Lipman Y, Levin D, Cohen-Or D. Green coordinates. ACM Trans Graph
2008;27:78:110.
[28] M uller M, Dorsey J, McMillan L, Jagnow R, Cutler B. Stable real-time
deformations. In: ACM symposium on computer animation, SCA 02; 2002.
p. 4954.
[29] Lang S. Algebra, graduate texts in mathematics. Springer; 2002.
[30] Le vy B, Petitjean S, Ray N, Maillot J. Least squares conformal maps for
automatic texture atlas generation. In: ACM SIGGRAPH 02; 2002. p. 36271.
[31] Graphite software /http://alice.loria.frS (February 2012).
[32] Solomon J, Ben-Chen M, Butscher A, Guibas L. As-killing-as-possible vector
elds for planar deformation. Comput Graph Forum 2011;30(5):154352.
[33] Joshi P, Meyer M, DeRose T, Green B, Sanocki T. Harmonic coordinates for
character articulation. In: ACM SIGGRAPH07. ACM; 2007.
[34] Ju T, Schaefer S, Warren J. Mean value coordinates for closed triangular
meshes. In: ACM SIGGRAPH05. ACM; 2005. p. 5616.
G.-P. Paille, P. Poulin / Computers & Graphics 36 (2012) 427433 433