Consistent Mesh Parameterizations and Its Application in Mesh Morphing

Jin-Bey Yu∗ Jung-Hong Chuang †

Department of Computer Science and Information Engineering
National Chiao Tung University
Hsinchu, Taiwan, Republic of China

Abstract
The correspondences establishment for a set of models is a versatile
algorithm in computer graphics and geometry processing, which in
general counts on a lot of specification by users to build a common
dissection with a set of feature points between the input models.
We propose a novel method to compute the consistent parame-
terizations for multiple models with a little user-control. Our pa-
rameterization scheme performs a geometric-stretch optimization Figure 1: Example of consistent mesh parameterizations.
to optimize the parameterization of each patch. The alignment of
feature correspondences for the result of parameterizations is also
considered using a foldover-free warping algorithm. Uniform or
adaptive remeshing can be performed to yield a set of semi-regular
meshes. Moveover, geometry images can be generated easily with A single model can be remeshed to yield a (semi-)regular one.
the parameterizations to store the geometric information in a sim- Can we extend the idea to multiple models with the same connec-
ple grid structure that benefits a number of applications, including tivity? The answer is not true unless they have a common dissection
normal mapping, texture mapping, rendering, level-of-detail, com- in which each patch of one model corresponds exactly to one patch
pression, and morphing. We also demonstrate the mesh morphing in all other models. In consequence, they will possess a same base
application between two or more objects in both spatial and wavelet domain and their parameterizations will be consistent. The diffi-
domain based on the correspondence established by a simple com- culty is that the common dissections are not easily found, because
mon dissection and remeshing. a good dissection of one model might be bad for another model.
Keywords: mesh parameterization, mesh simplification, geomet- Previous works [14, 25, 35, 36, 42] leave this problem to users, that
ric stretch, remeshing, geometry images, mesh morphing is, the user is required to specify a common dissection and many
corresponding feature points manually. Besides, many applications
such as metamorphosis (or morphing) and DGP (Digital Geometry
Processing) applications [36], which require the establishment of
1 Introduction correspondences among multiple models, benefit from consistent
parameterizations.
Many applications in visualization and computer graphics require
models of 3D surface geometry. The most commonly used rep- In this paper, we propose a novel method to compute the consis-
resentation is the triangle mesh. Such meshes can be the result tent parameterizations for multiple models with a little user-control
of careful design by using modeling software, or may come as an (Figure 1). The parameterization scheme combines the properties
output of a scanning device associated with reconstruction. As 3D of geometry images with semi-regular remeshing. After users sim-
surface geometry becomes a popular media, more and more meshes ply specify the initial patches, i.e. the base domain, on the model,
are available, coming from a variety of sources including 3D scan- our parameterization scheme will perform a geometric-stretch opti-
ners and modeling software. Although these meshes capture ge- mization to optimize the parameterization of each patch. The opti-
ometry accurately, their sampling quality is usually far from ideal mization algorithm borrows techniques from Sander et al. [38, 39].
for subsequent applications. For instance, these complex irregu- We prevent parametric “foldover” and penalize undersampling by
lar meshes are not appropriate for rendering on low-end computers, using a coarse-to-fine optimization strategy with geometric-stretch
texture mapping, morphing between different models. metric. This parameterization scheme can be extended to multiple
Remeshing, i.e. modifying the sampling and connectivity of a models and guarantees to yield consistent parameterizations. Extra
geometry to generate a new mesh which approximates the original feature correspondences can also be specified by users. The align-
mesh, is therefore a fundamental step for efficient mesh processing. ment of feature correspondences during parameterization process is
The state-of-the-art algorithms of remeshing involve initially dis- also considered by using a foldover-free warping [12].
secting the original irregular meshes into a set of topological disk- Once the consistent parameterizations are obtained, a number of
like patches, called base domain. These patches are later parame- applications can operate directly on the base domain. We perform
terized to compute a bijection between the 3D surface and the 2D the remeshing process and obtain a set of semi-regular meshes. The
parametric domain. remeshing process can be either uniform or adaptive. Based on the
Parameterizing a mesh amounts to computing a correspondence parameterization of the patches, normal maps, which capture geom-
between a 3D surface mesh and an 2D planar mesh through etry details, can also be easily resampled. This improves the real-
a piecewise linear mapping. In practice, this piecewise linear time rendering quality of semi-regular meshes in coarse levels. A
mapping is simply defined by assigning each mesh vertex a 2D multiresolution mesh morphing both in spatial and wavelet domains
coordinates (u, v) referring to its position on the planar domain. is also demonstrated as an application of our proposed approach.
Such an one-to-one mapping provides a flat parametric domain, al- Besides, we also exploit the idea of geometry images to store our
lowing users to perform any complex operation directly on the flat remesh data. The geometry images introduced by Gu et al. [15] sup-
domain rather than on the curved surface. Many parameterization port surfaces of arbitrary genus by allowing an arbitrary surface cut.
techniques are proposed in recent decade. However, the cut topology constrains level-of-detail mip-mapping.
In contrast, our reconstruction bases on a simple base domain, and
the topology of our polyhedral cuts is fixed and simple. The cuts
∗ e-mail: [email protected] give rise to symmetry rules at the image boundaries, which enable
† Contact author. e-mail: [email protected] morphing and better level-of-detail control.
2 Related Work view can be found in [1]. Particularly, Alexa pointed out in [1] that
the remeshing approach is appealing for morphing applications, be-
Parameterization. Several schemes have been proposed to flat- cause it allows to scale the size of the representation mesh. On the
ten a surface region to establish a parameterization over the last contrary, the conventional merging approach generates a more com-
ten years in computer graphics. Almost all techniques explicitly plicated intermediate representation.
aim at producing least-distorted parameterizations, and vary only
by an objective function (distortion considering) and the minimiza- Kanai et al. [24] uses a single patch and the patch will be param-
tion processes used. The main distinction between the functions eterized by harmonic mapping. The corresponding patch will be
is how they measure the distance of the parameterization from an aligned and merged to generate a merged mesh as intermediate rep-
isometry (a mapping preserving lengths and angles). resentation. In their recent works [25], the user first defines a set
of corresponding features vertices and applied their approximate
Maillot et al. [34] base their metric on spring-like energies, that is, shortest path algorithm [23] to find an initial dissection.
edge springs of nonzero rest length where rest length corresponds
to edge length on the surface. Eck et al. [8] proposed the discrete The work of Zöckler et al. [42] also requires users to provide the ini-
harmonic map, which assigns non-uniformly spring weights to the tial dissection as a high-level correspondence. The corresponding
mesh edges. patches will be applied foldover-free warping [12] in the parametric
domain for detailed level feature correspondences.
Floater [10] proposed specific weights based on the barycentric
maps [40] to improve the area deformations and shape-preserving Lee et al. [28] utilized their MAPS to the mesh morphing applica-
of the mapping. It guarantees embedding for convex boundaries tion. Their MAPS scheme supports feature point and line tagging
and the parameterization can be found by solving a linear system. such that the constructed base domains will preserve these features.
Floater [11] proposed another weighting algorithm. The weights However, the corresponding features points manually specified by
are derived from the mean value theorem for harmonic functions. users on both models may lie in different positions in both base
In addition, it are faster to compute than the shape-preserving pa- domain.
rameterizations [10] and have the theoretical advantage. Michikawa et al. [35] proposed a multiresolution interpolation
Hormann and Greiner [20] proposed the MIPS parameterization, meshes representation for mesh morphing. An interface is designed
which roughly attempts to preserve the ratio of singular values over for users to define a common patch layout on both source and tar-
the parameterization with a hierarchical solver [21]. get meshes. Each patch is then parameterized and a surface fitting
is performed to produce a semi-regular mesh for the intermediate
Sander et al. [39] proposed another non-linear energy for texture mesh. The method can be extended to multi-target morphing.
stretch distortion. Their stretch metric minimizes undersampling by
integrating the sum of squared singular values over the map. Sander
et al. [38] showed the stretch metric is related to SAE (Signal- 3 Consistent Mesh Parameterizations
Approximation Error) — the difference between a signal defined
on the surface and its reconstruction. 3.1 Approach overview
Several researchers introduced methods which satisfy positional Our strategy for consistent mesh parameterizations is as follows.
constraints approximately or “softly”. Lévy and Mallet [31] com- For each of the given meshes, we first consistently partition the
bine orthogonality and iso-parametric terms in their metric and al- meshes into a set of minimum disk-like patches such that the
lowed user-defined matching of iso-parametric curves. Lévy [30] patches from different meshes are in one-to-one correspondence.
suggested a method to incorporate soft positional constraints into For meshes of genus 0, the consistent partition is derived by con-
the parameterization problem. necting four user-specified partition points, which are in correspon-
Eckstein et al. [9] introduced a method that enforces “hard” dence and will serve as patch’s corners. The point connection can
(exact) positional constraints by deforming an existing embedding be done by applying the Dijkstra’s shortest path algorithm [7] based
while adding a number of Steiner vertices. on geodesic distance. Note that, additional feature points in cor-
respondence can be assigned by users for facilitating the vertex-
Remeshing. Lounsbery [32], Lounsbery et al. [33] have devel- correspondence problem commonly found in mesh morphing.
oped a technique that introduces multiresolution analysis to a re- For each patch, we derive a base mesh by applying a sequence
stricted class of meshes with subdivision connectivity. Initial dis- of half-edge collapsing operations. During the half-edge collaps-
section is crucial for remeshing, because it defines the base do- ing, all feature points are retained. On the base mesh, we com-
main. Eck et al. [8] use a Voronoi-based partition to dissect the pute an initial embedding of the base mesh by using the mean
input mesh. value coordinates and optimize the resulting embedding using a
geometric-stretch metric. To achieve the consistent mesh param-
MAPS [29] and Normal Meshes [16] employe a mesh simplifica- eterization among given meshes, all feature points in correspon-
tion algorithm to yield an initial dissection. The whole algorithm dence are aligned by using a foldover-free warping algorithm. Hav-
for finding an initial dissection is also automatic and more practical ing the initial embedding of the base mesh with the aligned feature
than the method in [8]. Based on the normal mesh algorithm, Praun points and the refinement sequence, we perform a coarse-to-fine
et al. [36] proposed an algorithm for finding the common initial parameterization to embed all other vertices and then optimize its
dissection for multiple models. geometric-stretch metric.
Given a number of meshes with corresponding partition points
More recently, Gu et al. [15] introduced Geometry Images. The and feature points specified by users, the steps of the proposed con-
process involves cutting the surface into a disk using a network of sistent mesh parameterization can be summarized as follows:
cut paths, and then parameterizes the resulting disk onto a square
domain [38, 39]. Using this parameterization, the surface geometry 1. Consistently partition the meshes into a set of minimum disk-
is resampled onto the pixels of an image. like patches by connecting partition points.
Alliez et al. [3] proposed an interactive sampling technique. A mesh 2. For patches in correspondence, we perform a consistent patch
is decomposed into a set of maps inserted in a pipeline of signal pro- parameterization
cessing algorithms. The output of this pipeline is a density map, in-
teractively resampled using an error diffusion technique commonly (a) Derive a base mesh for each patch by applying a se-
used for gray level image halftoning. quence of half-edge collapsing operations, compute an
initial embedding for the base mesh and perform geo-
Morphing. The correspondence problem in the mesh morphing metric stretch optimization.
is naturally related to parameterizations. The survey of 3D morph- (b) Align feature points in correspondence on the initial
ing can be referred to [27], and mesh morphing, an extensive re- embeddings of all the patches.
 (c) Perform a coarse-to-fine parameterization for each 3.3.2 Initial embedding of base meshes
patch with geometric-stretch optimization.
Mean value coordinates. An initial embedding for the base
mesh is derived by using mean value coordinates parameterization
Floater [11]. The patch’s parameterization builds a bijective map-
3.2 Consistent mesh partition ping u between a 3D mesh and a 2D parametric domain. With mean
value coordinates approach, we first fix the parameters for boundary
Genus-zero meshes. We partition a genus zero mesh into two vertices of the quadrilateral patch on a unit square and then derive
quadrilateral disk-like patches by requiring users to specify four the function u that satisfies the following equations for each of the
partition points on the surface and then performing Dijkstra’s short- interior vertices vi , i = 1, 2, ..., n:
est path algorithm [7] based on geodesic distance to trace from one
partition point to another and form a closed loop. Now with this
closed curves, the original mesh is dissected into two patches and
u(vi ) = ∑ λi,k u(vk ), (1)
k∈V (i)
the respected faces are well marshaled. Note that, to consistently
parameterize the given meshes, the partition points must be in cor-
respondence.
∑ λi,k = 1, (2)
k∈V (i)

where V (i) is the 1-ring neighborhood of the vertex vi , λi,k ≥ 0,

Extending to genus-n meshes. For a genus-n mesh, the for each i and k, and n is the number of the vertices of the patch.
number of patches must at least be 2(n + 1). In the case of genus- Equation (1) expresses that vi as a convex combination of its neigh-
one meshes, we specify four partition points to locate the corners boring points. To derive the mapping u, we first compute the λi,k ,
of the four patches and eight auxiliary points to guide the tracing of k ∈ V (i), for each i, that satisfies the properties of barycentric coor-
patch’s boundaries. dinates. Then, we solve the linear system of Equation (1) for u(vi ).
In the simplest case k = 3, the weights λi,1 , λi,2 , λi,3 are uniquely
determined by (1) and (2) alone; they are the barycentric coordi-
3.3 Consistent patch parameterization nates of vi with respect to the triangle [vi1 , vi2 , vi3 ], and they are pos-
itive. For general k, these weights λi,k can be derived from an ap-
With consistent mesh partition, each given mesh is partitioned into a plication of the mean value theorem for harmonic functions, which
set of patches, with the property that each patch has one-to-one cor- suggests calling them mean value coordinates. As shown in Figure
respondence to a patch in each of other meshes. Consistent patch 2, the derivation of λi,k for vi is as follows:
parameterization parameterizes those corresponding patches such
that all corresponding feature points are aligned to identical param- wi,k tan(αk−1 /2) + tan(αk /2)
eter coordinates. Furthermore, stretch minimization over the pa- λi,k = , wi,k = , (3)
rameterization is also addressed. ∑l∈V (i) wi,l vik − vi 
Although the mesh parameterization and remeshing ensure the
one-to-one vertex correspondence required for mesh morphing. where αk , the angle at vi in the triangle [vi , vik , vik+1 ], is defined
The methods for parameterization and remeshing, however, deter- cyclically. The coordinates approximate harmonic maps by piece-
mine how well the original shapes are approximated. Minimizing wise linear maps over triangulations, in such a way that injectivity
the stretch norm over the parameterization is critical in this respect, is preserved.
but, on a uni-resolution parameterization, it is slow and often con-
verges to bad local minima. Here, a strategy similar to that in Hor-
mann et al. [21], Sander et al. [38, 39] is taken to obtain better 0 00
1 11
vik+1

11 1
00 0 1
0
γ
performance in both speed and quality. vi k
α
0α γ β
1
For each patch, we start the parameterization by deriving a bash k
mesh and its associated embedding. All feature points are retained

0
1
k−1
in the base mesh, and those feature points in correspondence are k

0
1 0
1
aligned to have identical parameter coordinates by a fold-free mesh
β k−1

0
1
warping. The patch parameterization is finally derived by a coarse- k−1 vik
to-fine parameterization with geometric-stretch optimization. vik−1

Figure 2: Star-shaped polygon.

3.3.1 Base mesh derivation
The base mesh for a patch is the coarsest mesh derived after apply- The main advantage of mean value coordinates is that the λi,k
ing the progressive mesh (PM) simplification [18] to the patch. are always positive, because they are all derived from barycentric
coordinates. This guarantees the mapping is bijective and a sparse
linear system derived from Equation (1) can be solved robustly by
The PM is derived by simplifying the patch’s M i+1
the iterative biconjugate gradient method. In addition, the mean
mesh using a sequence of half-edge collapsing op- value coordinates are faster to compute than the shape-preserving
erations. The quadric error metric [13] that ap- v2
coordinates [10], and have the theoretical advantages.
proximates both visual and geometric error is used.
The half-edge collapsing operation (v1 , v2 ) → v1 v1
affects the neighborhood of v2 as shown on the edge collapse Stretch optimization. To optimize the parameterization, we
right figure and leaves the position and attributes of apply a global optimization similar to that in Sander et al. [39].
v1 unchanged. Compared to the full-edge collaps- Mi
To minimize geometric-stretch metrics L2 (M) and L∞ (M), we be-
ing, half-edge collapsing avoids writing to the ver- gin with the mean value coordinates parameterization, and perform
tex buffer during run-time LOD switches, and, in several optimization iterations. Within each iteration, the optimiza-
consequence, parametric coordinates of every ver- tion procedure sweeps all vertices in the base mesh and optimize
v1
tex are the same at all levels. each one in turn. More precisely, they place all vertices in a priority
During simplification, we disallow an edge collapsing of queue ordered by the amount of geometric stretch in their neigh-
(v1 , v2 ) → v1 if v2 is a patch corner (to preserve corners), or if v2 borhood. The process stops when the largest change is below a
is on a patch boundary and edge (v1 , v2 ) is not on a patch boundary threshold (e.g. 10− 3). For each vertex, we minimize the geometric-
(to preserve boundary straightness), or if v2 is a feature point (to stretch metric by repeatedly updating its parametric coordinates us-
retain feature points in the initial embedding). ing bracketed parabolic minimization (instead of the binary line
 (a) Initial embedding of the (b) Coarse-to-fine param. (c) Coarse-to-fine param. (d) Final parameterization.
base mesh (100 faces) (1402 faces) (2704 faces) (4006 faces)

Figure 3: Several coarse hierarchy levels of a triangulated data set (top) and the corresponding optimized parameterizations (bottom).

search used in [37]) along randomly search directions in the (s,t) 3.4.1 Feature adjustment
parametric domain.
Given patches Si , i = 1, ..., m, and each associated set of feature
points { f1i , ..., fei }, where e is the number of feature points on Si .
3.3.3 Feature point alignment We compute the averaged feature position fk for each set of corre-
To achieve the consistent mesh parameterization among multiple sponding feature points fk1 , fk2 , ..., fkm , k = 1, ..., e, as the destination
patches, parameter coordinates of corresponding feature points are of the alignment for each fki , i = 1, ..., m; i.e.,
aligned to the same position on the initial embedding by using a
foldover-free warping algorithm, which will be detailed in Sec- 1 m i
tion 3.4. fk = ∑ fk ,
n i=1
k = 1, ..., e.

3.3.4 Coarse-to-fine parameterization

3.4.2 Parameterization warping
Having derived the initial embedding with feature correspondences
aligned, we now describe the coarse-to-fine parameterization algo- Patch triangulation For each S i , we first construct a warp mesh
rithm which embeds all other vertices and, in the meantime, opti- by a 2D Delaunay triangulation which takes four corners (four seed
mizes geometric-stretch metric. points), and f1i ,..., and f ei as input. All other points will be mar-
shaled into their respected enclosing triangles and their Barycentric
coordinates are also computed. The objective of the parameteriza-
Vertex insertion and optimization. For each vertex-split- tion warping for Si is to move feature point fki to fk for all i and k
refinement operation in the PM sequence, we place the new vertex
at the centroid of the kernel of its neighborhood polygons to main- and recompute parametric coordinates for all other points such that
tain an embedding (i.e. avoid flipped or degenerate triangles). Note the mapping is still bijective.
that if the mapping is an embedding prior to the vertex insertion, the
kernel can not be empty. After inserting a new vertex, we exploit Triangulation over time During warping the warp mesh, the
the local vertex optimization to optimize the stretch of the vertex movement of the feature points may results in triangle degeneration
and its neighborhood one at a time. The vertex optimization always and, in the consequence, folding over. We call such case an event.
decreases the geometric-stretch metric, and since it is positive, it Before moving the feature point fki , we detect if there will be an
must converge to a local minimum. Note that feature points are re- event by a binary search between its current position and the des-
tained in the optimization process to preserve the feature alignment. tination. If an intermediate position of the event is found, we first
Figure 3 shows some hierarchy levels of the coarse-to-fine parame- alter the local triangulation and recompute all Barycentric coordi-
terization and their corresponding optimized parameterizations.
nates affected by this alteration, and then move fki to the position
of the event. Then all parameters are recomputed by using the new
3.4 Feature point alignment barycentric coordinates. If no event is found, simply move the fea-
ture point to the destination and recompute all parameters. Note
To achieve the consistent mesh parameterizations, a foldover-free that the influence range of the warping is not global and only pa-
warping in the parametric domain is applied to the initial embed- rameters marshaled inside this range are affected.
ding such that corresponding feature points are aligned to the same The triangulation over time process is applied to all feature point,
position. An approach similar to image foldover-free warping pro- one at a time. The processing order shall affect the final distribution
posed [12] is used. of the non-feature points, the mapping is, however, bijective.
 (a) Base domain and feature points

(b) Initial embedding

(c) After feature alignment

(d) After coarse-to-fine parameterization

(e) With checkered texture mapping

Figure 4: The steps of consistent mesh parameterizations. (The green marks are “feature points”)
3.5 Examples Adaptive remeshing will have more samples in the area of high cur-
vature.
Figure 4 shows the steps of our consistent mesh parameterization. For a given input mesh M , a base domain consisting of some
The checkerboard texture is used to indicate the quality of the pa- quad-faces corresponding to patches are constructed. For a given
rameterization. Note that four different colors represent the parts in quad-face q, we find a best-fitting plane g, and measure the min-
correspondence between two models. imum Euclid distance between the plane g and each vertex in the
patch Pq associated with quad-face q. Let |v − p| be the minimum
Eucilid distance for each v ∈ q and each p ∈ g. We define a error
4 Remeshing function E(q) for each quad-face q of the remesh as the maximum
of such distances, i.e.,
With derived consistent parameterization for each corresponding
patches, the original models can be approximated by either uni- E(q) = max d(v) (4)
v∈Pq
form remeshing or adaptive remeshing. Remeshing is a process that
samples in the parameterized domain and derives inverse mapped
3D vertices. The face containing the inverse mapped vertex is first The error function can be normalized by the diagonal length of the
derived by a point location algorithm. The inverse mapped vertex bounding box of the input mesh, denoted as B(M ); i.e.,
is them computed using the barycentric coordinates.
Uniform remeshing does the sampling based on regular subdivi- E(q)
EN (q) = (5)
sion on the square domain, and in consequence, the total number of B(M )
samples is (2n + 1)2 for level n. For multiple models, the remeshed
models of the same level will automatically have the identical con- We begin the remeshing process with base domain mesh and con-
nectivity. Note that, with the parameterization, all required geom- struct a quadtree root for each quad-face of base domain mesh.
etry information can be stored in a 2D array or a quad-tree data Then we evaluate the error of quad-faces in each quadtree based
structure depending on the applications. Figure 5 shows the result on the error function EN (q). If the error of a quad-face q, EN (q), is
of uniform remeshing derived from the parametrizations in Figure exceeding a pre-defined error bound ε , the quad-face is further re-
4. fined and its geometry information are resampled. In other words,
we take the quad-face to next finer level and four new children will
be attached into the quadtree. The process is performed recursively
and the adaptive remesh is constructed. However, there will be lots
of T-vertices, which appear along the boundaries between quad-
faces of different levels. We first force the level difference between
neighboring quad-faces to be at most one by deliberately refining
(a) Pig: M1 , 8 faces (b) Triceratops: M1 , 8 faces the quad-face of coarser level. Then perform adaptive subdivision
to quad-face of coarser level. To perform adaptive remeshing con-
sistently on multiple models, we simply take the maximum of the
error of each corresponding quad-faces in the multiple models as
follows:

EC = max (EN (qi )) (6)

1≤i≤n

where n is the number of models. This scheme of adaptive remeshes

will result in meshes with the same connectivity. Figure 6 shows the
(c) Pig: M3 , 128 faces (d) Triceratops: M3 , 128 faces results of the adaptive remeshing.

(e) Pig: M5 , 2048 faces (f) Triceratops: M5 , 2048 faces

Figure 6: Adaptive remeshing of pig and triceratops models, 17998
faces (ε = 0.002).

Although adaptive remeshing has done a great job for reducing

the total number of faces under the geometry criterion, it still is an
approximation of the original mesh. Much small detailed geometric
features might still not well reconstructed. We can employ normal
mapping technique for further compensation on this issue. With
the modern graphics hardware, which support “multi-texturing” and
(g) Pig: M7 , 32768 faces (h) Triceratops: M7 , 32768 faces “pixel shader” (or “register combiners” in OpenGL), the per-pixel
lighting can be performed in real-time [26]. Figure 7 shows the
result with and without normal mapping.
Figure 5: The uniform remeshing derived from the parameteriza-
tions in Figure 4.
5 Results and Analysis
Uniform remeshing has the drawback that in order to resolve a
small local feature on the original mesh, one may need to subdivide We have implemented consistent mesh parametrization and remesh-
to a very fine level. This total number of faces will be quadrupled. ing as described above. The applications was written in C++ using
 (a) Original model: 69642 faces (b) M5 : 2048 faces (c) M7 : 32768 faces

(d) Normal Map (e) M5 with normal mapping (f) M7 with normal mapping

Figure 7: Normal mapping example.

standard computational geometry data structures, for example, half- level remesh can still be morphed with high visual quality. It’s a
edge data structures. The operating system is Microsoft Windows pity that normal map interpolation can not be easily formulated for
2000. The graphics rendering is based on OpenGL with nVidia ex- hardware acceleration such as register combiners or pixel shaders.
tensions. Because linear interpolated normal vector will not have the unit
The remeshing results are evaluated by IRI-CNR Metro tool [6]. length, the further normalization is needed. Figure 8 shows the
The Metro tool compares two triangular meshes which describe the spatial morphing sequence from a pig model to a triceratops model.
same surface. The tool accepts input meshes in the Open Inven-
tor, PLY and SMF format. Error can be measured in terms of the
symmetric mean distance, the symmetric RMS distance and the vol-
ume difference for closed surfaces. After got the symmetric RMS 6.2 Scheduled interpolation in wavelet domain
distance, we follow the evaluate function in [15]. Accuracy is mea-
sured as Peak Signal to Noise Ratio PSNR = 20 log10 (peak/d), The interpolation path is also a major issue in the context of mor-
where peak is the diagonal length of the bounding box and d is the phing. Besides linear interpolation, spline interpolation can also be
symmetric RMS Hausdorff error (geometric distance) between the employed easily. We also can perform the interpolation in another
original mesh and the remesh. Roughly speaking, the PSNR about domain instead of spatial domain. Hughes [22] apply Fourier trans-
70dB is considered to be a nice approximation. form to volume data and perform interpolation for volume morph-
Table 1 shows calculation time, which is performed on a 1.5GHz ing in frequency domain while He et al. [17] performed similiar
AMD Athlon XP PC. Table 2 shows the result of uniform and adap- process in wavelet domain. This idea is impossible for irregular
tive remeshing. Both single object adaptive remeshing and multiple meshes, but is now possible for regular and semi-regular meshes.
object adaptive remeshing are shown. Table 3 shows the PSNR and We use the bi-orthogonal lifting wavelet proposed by Bertram et al.
Table 4 shows the symmetric Hausdorff distance, i.e. maximum er- [4]. Their wavelet is much similiar to Catmull-Clark [5] subdivi-
ror. sion wavelet, and has small support. Therefore, both forward and
inverse wavelet transform can be computed quickly. The generated
wavelet coefficients can be manipulated in such a way that the in-
verse wavelet transform will reconstruct a morphed model. We ap-
6 Mesh Morphing ply different schedules to different resolutions in wavelet domain,
i.e. assign different starting interpolation time and speed for differ-
6.1 Morphing in spatial domain ent resolution in wavelet domain. We design the schedule as shown
in Figure 9. The high frequency part of source object is disppearing
The meshes constructed from the remeshing process for multiple quickly, and the low frequency part is interpolated gradually. The
models have the same connectivity. We can simply linearly inter- high frequency par of target object is then add back quickly to the
polate their positions. A sequence of intermediate morphed mod- end of morphing.
els will be generated, and they still have multiresolution structures.
Note that our method can handle multi-target morphing as well.
Extra feature correspondences are also crucial to the quality of the
morphing. 6.3 Multi-target morphing
The morphing speed is fast at coarser level, but the visual quality
is poor. As previously mentioned, by using consistently adaptive As shown in Figure 10, with the consistent parameterizations and
remeshing, we can construct adaptive remeshes for multiple models remeshing, we can produce any morphing sequence among these
and morph them. By using normal mapping technique, a coarse models.
(a) Morphing from a pig to a triceratops (b) Close-up view (c) With normal maps

Figure 8: Spatial morphing examples.

 Figure 9: Morphing of scheduled interpolation in wavelet domain.

7 Conclusion and Future Work [5] E. Catmull and J. Clark. Recursively generated B-spline sur-
faces on arbitrary topological meshes. Computer-Aided De-
We have proposed a novel method to compute consistent parame- sign, 10:350–355, Sept. 1978.
terizations for multiple models with a few seed points specified by
users. Extra feature points in correspondences can also be specified [6] P. Cignoni, C. Rocchini, and R. Scopigno. Metro:
for semantics and aligned during the parameterization procedure. Measuring error on simplified surfaces. Com-
Based on the result of parameterizations, uniform and adaptive puter Graphics Forum, 17(2):167–174, 1998. URL
remeshing can be performed to yield a set of semi-regular meshes. http://vcg.iei.pi.cnr.it/metro.html.
Moveover, geometric details can easily be resampled and stored as
normal maps to improve the visual effects using the modern graph- [7] T. H. Cormen, C. E. Leiserson, and R. L. Rivest. Introduction
ics hardware. In addition, the geometry images generated using the to Algorithms. MIT Press, Cambridge, Mass., 1990.
same parameterizations have a simple grid structure that benefits
a number of applications, including normal mapping, texture map- [8] M. Eck, T. DeRose, T. Duchamp, H. Hoppe, M. Lounsbery,
ping, rendering, level-of-detail, compression, and morphing. We and W. Stuetzle. Multiresolution analysis of arbitrary meshes.
have demonstrated the 3D mesh morphing application by the result- In SIGGRAPH 95 Conference Proceedings, pages 173–182,
ing remeshing. Besides interpolation in spatial domain, scheduled Aug. 1995.
interpolation in wavelet domain is also demonstrated.
[9] I. Eckstein, V. Surazhsky, and C. Gotsman. Texture map-
There are a number of areas for future work: ping with hard constraints. Computer Graphics Forum, 20
(3), 2001. ISSN 1067-7055.
• Higher genus models: We seek more convenient methods to
handle the layout of the domain for higher genus models. [10] M. S. Floater. Parametrization and smooth approximation of
surface triangulations. Computer Aided Geometric Design, 14
• Improvement on parameterizations: Models with some (3):231–250, 1997.
protrusions are difficult to get good parameterizations. Be-
sides, anisotropy parameterization along these protrusions [11] M. S. Floater. Mean value coordinates. Computer Aided Ge-
also becomes unavoidable. ometric Design, 20:19–27, 2003.
• Feature-sensitive remeshing: Remeshing schemes used here [12] K. Fujimura and M. Makarov. Foldover-free image warping.
are not effective in capturing sharp features. Other resampling Graphical models and image processing: GMIP, 60(2):100–
schemes may improve the geometric quality to represent sharp 111, Mar. 1998.
features accurately [2, 3, 19, 41].
[13] M. Garland and P. S. Heckbert. Surface simplification using
• Robust to multiple models: An improvement or a control- quadric error metrics. pages 209–216, Aug. 1997.
lable scheme on feature alignment is significant for the corre-
spondence establishment among models. [14] A. D. Gregory, A. State, M. C. Lin, D. Manocha, and M. A.
Livingston. Feature-based surface decomposition for corre-
spondence and morphing between polyhedra. In Computer
References Animation’98, June 1998.
[1] M. Alexa. Mesh morphing. In EUROGRAPHICS’01 State of [15] X. Gu, S. J. Gortler, and H. Hoppe. Geometry images. In
The Art Report, 2001. SIGGRAPH 2002 Conference Proceedings, pages 335–361,
2002.
[2] P. Alliez, É. C. de Verdière, O. Devillers, and M. Isenburg.
Isotropic surface remeshing. In Proceedings of Shape Model- [16] I. Guskov, K. Vidimče, W. Sweldens, and P. Schröder. Nor-
ing International, 2003. mal meshes. In Proceedings of the Computer Graphics Con-
ference 2000 (SIGGRAPH-00), pages 95–102, July 23–28
[3] P. Alliez, M. Meyer, and M. Desbrun. Interactive geometry 2000.
remeshing. ACM Transactions on Graphics. Special issue for
SIGGRAPH conference, 21, 3:347–354, 2002. [17] T. He, S. Wang, and A. Kaufman. Wavelet-based volume mor-
phing. In Proceedings of the Conference on Visualization,
[4] M. Bertram, M. A. Duchaineau, B. Hamann, and K. I. Joy. pages 85–92, Oct. 1994.
Bicubic subdivision-surface wavelets for large-scale isosur-
face representation and visualization. In IEEE Visualization [18] H. Hoppe. Progressive meshes. In SIGGRAPH 96 Conference
’00 (VIS ’00), pages 389–396, Oct. 2000. Proceedings, pages 99–108, Aug. 1996.
[19] H. Hoppe, T. DeRose, T. Duchamp, J. McDonald, and [37] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flan-
W. Stuetzle. Mesh optimization. 27:19–26, Aug. 1993. nery. Numerical Recipes in C: The Art of Scientific Comput-
ing (2nd ed.). Cambridge University Press, Cambridge, 1992.
[20] K. Hormann and G. Greiner. MIPS: An efficient global ISBN 0-521-43108-5.
parametrization method. In Curve and Surface Design: Saint-
Malo 1999, pages 153–162. Vanderbilt University Press, [38] P. V. Sander, S. J. Gortler, J. Snyder, and H. Hoppe.
2000. Signal-specialized parametrization. In Proceedings of the
13th Eurographics Workshop on Rendering (RENDERING
[21] K. Hormann, G. Greiner, and S. Campagna. Hierarchical TECHNIQUES-02), pages 87–98, June 26–28 2002.
parametrization of triangulated surfaces. In Proceedings of Vi-
sion, Modeling, and Visualization 1999, pages 219–226, Nov. [39] P. V. Sander, J. Snyder, S. J. Gortler, and H. Hoppe. Texture
1999. mapping progressive meshes. In SIGGRAPH 2001 Confer-
ence Proceedings, pages 409–416, 2001.
[22] J. F. Hughes. Scheduled Fourier volume morphing. In Com-
puter Graphics (SIGGRAPH ’92 Proceedings), volume 26, [40] W. T. Tutte. How to draw a graph. Proc. London Mathemati-
pages 43–46, July 1992. cal Society, 13:743–768, 1963.

[23] T. Kanai and H. Suzuki. Approximate shortest path on poly- [41] J. Vorsatz, C. Rössl, L. P. Kobbelt, and H.-P. Seidel. Feature
hedral surface and its applications. Computer-Aided Design, sensitive remeshing. Computer Graphics Forum, 20(3), 2001.
33(11):801–811, September 2001.
[42] M. Zöckler, D. Stalling, and H.-C. Hege. Fast and intuitive
[24] T. Kanai, H. Suzuki, and F. Kimura. Three-dimensional geo- generation of geometric shape transitions. The Visual Com-
metric metamorphosis based on harmonic maps. The Visual puter, 16(5):241–253, 2000.
Computer, 14(4):166–176, 1998.
[25] T. Kanai, H. Suzuki, and F. Kimura. Metamorphosis of arbi-
trary triangular meshes. IEEE Computer Graphics and Appli-
cations, 20(2):62–75, Mar./Apr. 2000.
[26] M. J. Kilgard. A practical and robust bump-mapping tech-
nique for today’s gpus. In Game Developer’s Conference
2000 Course ”Adavanced OpenGL Game Development”,
2000.
[27] F. Lazarus and A. Verroust. Three-dimensional metamorpho-
sis: a survey. The Visual Computer, 14:373–389, 1998.
[28] A. Lee, D. Dobkin, W. Sweldens, and P. Schröder. Multires-
olution mesh morphing. In Proceedings of the Conference on
Computer Graphics (Siggraph99), pages 343–350, Aug.8–13
1999.
[29] A. W. F. Lee, W. Sweldens, P. Schröder, L. Cowsar, and
D. Dobkin. MAPS: Multiresolution adaptive parameteriza-
tion of surfaces. In SIGGRAPH 98 Conference Proceedings,
pages 95–104, July 1998.
[30] B. Lévy. Constrained texture mapping for polygonal meshes.
In SIGGRAPH 2001 Conference Proceedings, pages 417–
424, 2001.
[31] B. Lévy and J. Mallet. Non-distorted texture mapping for
sheared triangulated meshes. In SIGGRAPH 98 Conference
Proceedings, pages 343–352, July 1998.
[32] M. Lounsbery. Multiresolution Analysis for Surfaces of Arbi-
trary Topological Type. PhD thesis, Department of Computer
Science and Engineering, University of Washington, 1994.
[33] M. Lounsbery, T. D. DeRose, and J. Warren. Multiresolu-
tion analysis for surfaces of arbitrary topological type. ACM
Transactions on Graphics, 16(1):34–73, Jan. 1997.
[34] J. Maillot, H. Yahia, and A. Verroust. Interactive texture map-
ping. In Computer Graphics (SIGGRAPH ’93 Proceedings),
volume 27, pages 27–34, Aug. 1993.
[35] T. Michikawa, T. Kanai, M. Fujita, and H. Chiyokura. Mul-
tiresolution interpolation meshes. In Proceedings of the ninth
Pacific Conference on Computer Graphics and Applications
(PACIFIC GRAPHICS-01), pages 60–69, Oct. 16–18 2001.
[36] E. Praun, W. Sweldens, and P. Schröder. Consistent mesh
parameterizations. In SIGGRAPH 2001 Conference Proceed-
ings, pages 179–184, 2001.
 (a) The target models with extra feature correspondences.

(b) All morphing sequence.

Figure 10: Another example of multi-target morphing.

 time(sec)
Model face
parametrization remeshing
venus+isis 10000+10000 35.69 2.38
horse+female 15000+10000 51.21 2.63
horse+triceratops 15000+5660 43.61 2.47
horse+pig 15000+7164 45.63 2.51
pig+triceratops 7164+5660 30.03 2.41
bunny 69642 100.14 2.15

Table 1: Statistics for computational time.

The parametrization applies coaese-to-fine parametrization.
The remeshing scheme is uniform and the maximum remeshing level is 8.

face
Model uniform remeshing adaptive
original
M1 M2 M3 M4 M5 M6 M7 M8 SA CA
venus 9647
10000 13570
isis 7470
pig 7164 8 32 128 512 2048 8192 32768 131072 14693
17998
triceratops 5660 10475
bunny 69642 14006

Table 2: Statistics for various faces.

SA means “separately adaptive remshing with ε = 0.002;
CA means “consistently adaptive remeshing with ε = 0.002.

PSNR (dB)
Model uniform remeshing adaptive
M1 M2 M3 M4 M5 M6 M7 M8 SA CA
venus 23.71 31.33 39.91 48.69 56.40 64.41 73.53 82.77 65.76 67.03
isis 27.12 30.15 37.48 47.18 55.83 65.06 74.46 84.00 66.23 68.00
pig 24.45 29.74 33.63 40.78 47.14 53.21 59.52 64.27 53.36 53.39
triceratops 24.72 29.86 34.35 42.93 48.89 54.20 61.07 68.69 55.56 55.62
bunny 23.34 29.24 36.87 44.24 52.31 60.79 70.96 80.33 65.82

Table 3: Statistics for PSNR.

symmetric Hausdorff distance(%)

Model uniform remeshing adaptive
M1 M2 M3 M4 M5 M6 M7 M8 SA CA
venus 9.99 6.58 4.46 2.50 1.05 0.63 0.42 0.16 0.27 0.27
isis 9.99 9.99 7.71 2.35 1.37 0.62 0.42 0.20 0.24 0.25
pig 9.99 9.99 9.34 4.05 2.89 1.51 1.09 0.80 0.93 0.91
triceratops 9.99 9.99 9.21 3.77 2.53 1.37 0.82 0.61 0.72 0.72
bunny 10.00 9.99 6.68 3.65 2.35 1.28 0.49 0.21 0.35

Table 4: Statistics for symmetric Hausdorff distance with respect to the bounding box diagonal.