Eurographics Symposium on Geometry Processing (2003)

L. Kobbelt, P. Schröder, H. Hoppe (Editors)

Multi-Chart Geometry Images

P. V. Sander1, Z. J. Wood2, S. J. Gortler1, J. Snyder3, and H. Hoppe3

1 2 3
Harvard University, Caltech, Microsoft Research

Abstract
We introduce multi-chart geometry images, a new representation for arbitrary surfaces. It is created by resampling
a surface onto a regular 2D grid. Whereas the original scheme of Gu et al. maps the entire surface onto a single
square, we use an atlas construction to map the surface piecewise onto charts of arbitrary shape. We demonstrate
that this added flexibility reduces parametrization distortion and thus provides greater geometric fidelity, particu-
larly for shapes with long extremities, high genus, or disconnected components. Traditional atlas constructions
suffer from discontinuous reconstruction across chart boundaries, which in our context create unacceptable surface
cracks. Our solution is a novel zippering algorithm that creates a watertight surface. In addition, we present a
new atlas chartification scheme based on clustering optimization.
Categories and Subject Descriptors (according to ACM CCS): I.3.5 [Computer Graphics]: Surface representations.

1. Introduction
Regular remeshing is the process whereby an irregular mesh
is approximated by a mesh with (semi-)regular connectivity
[Eck et al. 1995]. The simplicity of a regularly remeshed
representation has many benefits. In particular it eliminates
the indirection and storage of vertex indices and texture
coordinates. This will allow graphics hardware to perform
rendering more efficiently, by removing random memory
accesses and thus improving memory access performance.
Geometry images. The most extreme such method, which
creates the most regular remeshed representation, is the
geometry image (GIM) introduced by Gu et al. [2002]. Their
construction converts the surface into a topological disk using
a network of cuts and parametrizes the resulting disk onto a
square domain. Using this parametrization, the surface
geometry is resampled onto the pixels of an image. As an
added benefit, techniques such as image compression can be
directly applied to the remesh. chartified orig. mesh image 400x160 remesh
PSNR=71.7 dB
However, this extreme approach of mapping an entire surface Figure 1: Example of a multi-chart geometry image.
to a single square has limitations. Models having disconnect-
ed components require a separate geometry image per editing [Zorin et al. 1997] and compression [Khodakovsky et
component, and complicated shapes with many extremities or al. 2000].
topological handles have distorted parametrizations.
This more flexible representation still has constraints that can
Semi-regular remeshing. A less extreme approach is to negatively impact parametrization efficiency. In particular,
create a remesh with the connectivity of a subdivided base each chart (each surface region associated with a triangle of
mesh. Examples of this approach include the methods of Eck the base mesh) is effectively parametrized onto an equilateral
et al. [1995], Lee et al. [1998; 2000], Kobbelt et al. [1999], triangle, which is then evenly sampled. Charts with non-
Guskov et al. [2000], and Wood et al. [2000]. For these triangular shapes are thus distorted by the parametrization;
representations, special kinds of continuous multiresolution charts that are long and skinny are invariably sampled aniso-
basis functions can be derived that allow multiresolution tropically. In addition, all charts, regardless of their size or
© The Eurographics Association 2003.
 Sander et al. / Multi-Chart Geometry Images
information content, must be allotted the same number of 2. Previous work
samples. 1 Atlas parametrizations. Some atlas schemes map individual
Multi-chart geometry images. We describe a new atlas- triangles or pairs of triangles separately into a texture [e.g.
based parametrization to define multi-chart geometry images Cignoni et al. 1998, Carr and Hart 2002]. In this section, we
(MCGIMs). Motivated by the atlas approach for texture review more general chart-based atlas constructions.
mapping (e.g. Maillot et al. [1993]), we partition the surface Maillot et al. [1993] partition a mesh into charts based on
into a geometrically natural set of charts, parametrize the bucketing of face normals. Their parametrization method
charts individually onto irregular polygons, and pack the optimizes edge springs of non-zero rest length.
polygons into a geometry image (Figure 1). Such an atlas
parameterization reduces distortion because the smaller charts Piponi and Borshukov [2000] manually cut a subdivision
are easier to flatten and because their parametrized bounda- surface using a network of edges. They parametrize the
ries can assume more natural shapes. Low-distortion resulting single chart using a “pelting” analogy, by stretching
parametrizations distribute samples more uniformly over the out the surface boundary using a collection of springs.
surface and therefore better capture surface signals. Each Sander et al. [2001] partition a mesh using greedy face
chart can be allotted an appropriate number of “defined” clustering (also done independently by Garland et al. [2001]).
samples in the geometry image, separated by “undefined” They parametrize the resulting charts onto convex polygons
samples. Our representation can be viewed as piecewise using geometric stretch. The charts are packed into a square
regular since it is composed of sub-regions of a regular GIM. using a greedy algorithm based on bounding rectangles.
MCGIMs retain the key advantage of the original GIMs – Lévy et al. [2002] align chart boundaries with high-curvature
rendering involves a simple scan of the samples in stored features of the surface. After locating a set of seed faces
order. And, MCGIMs overcome the distortion present in farthest from sharp features, they grow charts about these
GIMs, at the small expense of assigning some samples special seeds, and merge some resulting charts. They parametrize
“undefined” values. each chart using least-squares conformal maps with free
A serious drawback of a general atlas parametrization is that boundaries. They use a Tetris-like packing algorithm that
its reconstructed signal is discontinuous across chart bounda- searches for best fit over the horizon of pieces packed so far.
ries. Because irregular chart outlines do not align with the Sorkine et al. [2002] grow charts while simultaneously
sampling grid, boundaries between neighboring charts are parametrizing them. The chart growth stops when a distortion
generally sampled differently. For geometry images, such bound is reached or if self-overlap is detected, and a new
signal discontinuities create unacceptable cracks in the chart is started. Their parametrization uses a stretch-based
surface. Even a few erroneous pixels become glaring arti- metric that penalizes both undersampling and oversampling.
facts. To prevent cracks, we present a novel zippering
Sheffer and Hart [2002] cut a surface into a single chart by
scheme that unifies boundaries of the discretized MCGIM
cutting through high-distortion, less-visible surface regions.
charts to create a continuous (watertight) model.
None of these atlas constructions address the problem of
Our main contribution is the MCGIM representation obtained
inter-chart cracks when resampling geometry, as our zipper-
through this zippering scheme. To create accurate MCGIMs,
ing scheme does. Our chartification method is less greedy
we also introduce several improvements to existing atlas
than previous methods, thus yielding better results.
parametrization methods. We develop a new atlas chartifica-
tion scheme, based on general clustering optimization Model decomposition. Shlafman et al. [2002] also develop a
inspired by the work of Lloyd [1957] and Max [1960]. Our clustering-based approach to mesh partitioning. Because
scheme creates compact, roughly flat charts whose bounda- their application is morphing and not parametrization, their
ries align with the model’s creases. We apply the mesh clustering distance metric does not account for overall chart
optimization framework of Hoppe et al. [1993] to refit the planarity.
MCGIM samples to the original surface, thereby improving Zippering. Turk and Levoy [1994] reconstruct watertight
its accuracy considerably. And, we extend the “Tetris” surfaces from scanned height meshes by zippering mesh
packing algorithm of Lévy et al. [2002] to optimize both chart boundaries together. Because the desired surface is unknown,
rotations and overall domain dimensions. With these im- zippering is a challenging operation. In contrast, our zipper-
provements, we demonstrate that MCGIMs outperform both ing task (Section 4.5) is given a reference surface as input,
single-chart geometry images and semi-regular remeshes on and so can be made simple and robust.
example models.
Several algorithms consider the problem of reconstructing
surfaces from contours [e.g. Fuchs et al. 1977]. The simplest
case they consider is that of building a surface ribbon that
spans two parallel contour polygons. This is quite similar to
our zippering algorithm. The difference is that we seek to
unify the vertices across the gap rather than to construct a
1
This assumes that one wants a remesh free of T-vertices, while ribbon of new triangles between them.
avoiding the indirection-based representations that would
result from adaptive subdivision refinement.

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images
3. MCGIM representation local edge dihedral angles, then identify seed faces farthest
A multi-chart geometry image (MCGIM) is a rectangular 2D from sharp features, and finally grow charts about these
grid whose entries store 3D points sampled from a given seeds. In a sense, our new method can be seen as a generali-
surface. Since the MCGIM consists of multiple, arbitrarily zation of this approach, but uses a more general optimization
shaped charts packed together into a grid, there are some and does not require feature detection.
wasted samples that lie outside the charts. We refer to these Algorithm overview. Our algorithm is inspired by Lloyd-
as undefined samples. The remaining are called defined Max quantization, which partition sets of elements (e.g.
samples, and can be further distinguished into boundary colors), into a fixed number of clusters (e.g. quantized
samples (with at least one undefined sample among its four colors). The basic strategy is to iteratively optimize the
immediate neighbors) and interior samples. The defined clusters using two steps:
region of the image is specified either using a bitmask or by
• Partition the elements according to current cluster models.
storing NaN (Not-a-Number) into the coordinate values of
undefined samples. • Refit a model to each cluster.
In order to reconstruct a surface from the MCGIM, we must In our context, elements are mesh faces, clusters are charts,
specify how polygon faces are formed. For each 2-by-2 quad and the model is a seed face for each chart. We call the first
of samples, we examine how many samples are defined. If step “chart growth” and the second step “seed computation”.
fewer than three are defined, no polygon is formed. If exactly We now describe these in detail. Later, we present bootstrap-
three samples are defined, one triangle is created. And if all ping and termination conditions.
four samples are defined, two triangles are created, with the Phase 1: Chart growth. Given n seeds representing the n
diagonal chosen to be the shorter one in 3D. Applying this charts, assign each face to a chart.
method to all quads in the image creates a triangle mesh.
In this phase we partition the elements by growing all the
Although some triangles are in fact degenerate due to our
charts simultaneously using a Dijkstra search algorithm on
zippering, such degenerate triangles can still be fed to the
the dual graph of the mesh. This graph connects each pair of
graphics pipeline since they have no effect on rendering.
adjacent mesh faces by an edge. We assign edge costs that
Our goal is MCGIM construction is to create a watertight encourage charts to be planar and to have compact bounda-
mesh, i.e. a 2-dimensional manifold where every edge is ries. Whereas Shlafman et al. [2002] only consider local
adjacent to exactly two faces and every vertex is adjacent to dihedral angles, our metric considers distance to a global
exactly one ring of faces. Note that MCGIMs can represent chart normal. Specifically, the edge cost between a face 𝐹 on
meshes with several connected components, as well as a chart 𝐶 and its adjacent face 𝐹 ′ that is a candidate to join 𝐶
meshes with boundaries (where some edges are adjacent to is a measure of geometric distance between the two faces, and
only one face). difference in normal between 𝐹 ′ and the chart normal 𝑁𝐶 :
For rendering, normals are computed at each sample point cost(𝐹, 𝐹 ′ ) = �𝜆 − (𝑁𝐶 ⋅ 𝑁𝐹′ )� |𝑃𝐹′ − 𝑃𝐹 | ,
using the normalized cross product of central differences. where 𝑁𝐶 is the normal of the chart containing 𝐹 (the average
When an adjacent sample is undefined, it is replaced by the normal of all faces already in the cluster), and 𝑃𝐹′ and 𝑃𝐹 are
current sample in the central difference. (Opposite neighbors the centroid points of 𝐹 ′ and 𝐹. The parameter 𝜆 regulates
cannot both be undefined if the mesh is manifold.) the relative weight between the normal and geometric dis-
4. MCGIM creation tance terms. For all of our experiments, we set 𝜆 to 1. Phase 1
terminates when all faces have been assigned to charts.
4.1 Mesh chartification
Phase 2: Seed computation. Given a chartified mesh, update
The first step in producing a multi-chart geometry image is to the seed for each chart.
partition the mesh into charts. Charts that have compact
In this phase, we update the seed of each chart to be the “most
boundaries and are mostly flat yield better parametrization
interior” face within the chart. To find this face, we again
and packing results, thereby reducing the error of the geome-
perform a Dijkstra search on the dual graph of the mesh, but
try image.
in the opposite direction. We start the search from all faces
We first tried the chart merging approach of Sander et al. adjacent to a chart boundary. The edge cost is simply the
[2001]. They apply a greedy face clustering algorithm. geodesic distance
Because the resulting charts have irregular boundaries, they cost(𝐹, 𝐹 ′ ) = |𝑃𝐹′ − 𝑃𝐹 | .
subsequently straighten the boundaries. While the method
The last face reached within each chart becomes its new seed.
works well when the desired number of charts is high, for a
smaller number of charts the boundaries do not usually follow Convergence. Phases 1 and 2 are repeated until the new
the “creases” of the model. For instance, after the straighten- seeds are identical to those of the previous iteration. Occa-
ing step, the gargoyle wings of Figure 2a are each composed sionally, this discrete optimization fails to converge because
of a single chart. the seeds begin to cycle between nearby faces (defining
nearly identical chartifications). We detect these cycles by
This same observation motivated Lévy et al. [2002] to
checking the set of seeds against the ones obtained in previ-
develop an algorithm that aligns chart boundaries with high-
ous iterations. In practice, if a cycle is detected, any
curvature features. They first estimate sharp features using
chartification in the cycle is acceptable.
© The Eurographics Association 2003.
 Sander et al. / Multi-Chart Geometry Images
4.2 Chart parametrization
To parametrize the charts, we minimize the 𝐿2 geometric
stretch metric of Sander et al. [2001]. This same metric is
used by Gu et al. [2002]. It penalizes undersampling, and
results in samples that are uniformly distributed over the
surface. Sander et al. [2002] shows that the stretch metric is a
predictor of the reconstruction error under the assumption of
locally constant reconstruction. Instead of fixing the bounda-
ry to a square, we allow it to be free during the optimization
[Sander et al. 2002].
(a) Sander et al. [2001]; 𝐿2 stretch efficiency: 80%
We employ the hierarchical parametrization algorithm from
Sander et al. [2002] to solve the minimization problem. This
hierarchical algorithm finds better minima, and is orders of
magnitude faster than uniresolution methods. Once all the
charts are parametrized, we scale them relative to each other
based on their stretch [Sander et al. 2002].
4.3 Chart packing
The algorithm in the previous section produces a set of charts
parametrized on a continuous domain. The next step is to
(b) our method; 𝐿2 stretch efficiency: 99% discretize the charts and pack them into an image.
Figure 2: Comparison of the chartification algorithm Packing a set of 2D shapes is a provably hard problem
from Sander et al. [2001] with our new approach on a [Milenkovic 1998], thus motivating heuristic approaches.
12-chart gargoyle. (Garland et al. [2001] would pro- Our approach is a simple extension of the “Tetris” packing
duce results comparable to (a).) algorithm by Lévy et al. [2002]. Their scheme iteratively
adds charts to a square image while keeping track of a
Bootstrapping. In order to bootstrap the chartification horizon that bounds all the charts packed so far. For each
process, we first assign a random face to be the first seed. chart to insert, they locate the horizontal position that mini-
Then we run successive iterations of phases 1 and 2, with one mizes the wasted vertical space between the current horizon
small modification: at the end of phase 1, if the desired and the new chart.
number of seeds has not been reached, we add a new seed.
The new seed is set to be the last face that was assigned to a Our method considers 16 orientations for each chart. Also,
chart during phase 1. This tends to place the new seed far we let the domain be an arbitrary rectangle, rather than a
away from the other seeds. Once the desired number of seeds fixed-sized square, since current graphics hardware can
has been reached, we repeat phases 1 and 2 until convergence support textures with arbitrary dimensions. The user specifies
without adding new seeds. a desired upper-bound on the total number of MCGIM
samples, and we optimize for the best rectangle within this
Chart topology constraint. To ensure that all charts are area bound. Figure 3 demonstrates the improvement.
topological disks, we disallow the formation of annuli and
boundaryless (closed) charts. This is a simple local check on Chart scales. A scaling factor determines how the continu-
the chart boundary performed during chart growth. There are ous charts map to the discrete image grid. We initially set this
situations where a face cannot be assigned to any chart due to scaling factor assuming 100% packing efficiency, and then
the above constraint. When that happens, we add as a new successively reduce it by 1% until a successful packing is
seed the first face that failed this constraint during the chart achieved. For each scaling factor, the charts are rasterized
growth. For instance, one cannot chartify a sphere with one into a buffer to obtain their discrete shapes (see Section 4.4).
chart. So even if one chart is requested, more seeds are Domain dimensions. For a given chart scale, two nested
inserted automatically. loops consider MCGIM widths and heights whose product is
Result example. As shown in Figure 2b, with our new less than the specified upper-bound. We limit the aspect ratio
method the front and back of the ears and wings are each to be less than 4.
assigned their own chart. Also note that the chartification is Chart placement. We place
much more symmetric. We parametrized both models and each chart in the position and
calculated their stretch efficiencies, which is a measure of orientation that minimizes
distortion of the parametrization. While the method of Sander “wasted space” (inset in red). In
et al. [2001] has a stretch efficiency of just 80% (due to high Lévy et al. [2002], the empty
stretch on the gargoyle wings), the new method has the near space is the region between the
optimal stretch efficiency of 99%. More chartification discretized upper horizon of the
examples are shown throughout the paper. atlas (black) and the discretized
lower horizon of the inserted chart (green). We also add any

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images
wide rectangle centered along the boundary edge, and scan-
convert it into the texture domain. During this boundary
rasterization, undefined samples covered by the boundary-
polygon are labeled as defined and filled in with the closest
geometric point on the boundary. Previously defined samples
are not overwritten.
Non-manifold dilation. Even with boundary rasterization, a
discrete chart can have a non-manifold reconstruction at
boundary samples (Figure 6). For instance, a single line
peninsula of defined samples produces no triangles. Recon-
(a) chartified original mesh (b) Lévy et al. [2002] structions can also erroneously produce two triangles
(50 charts) (packing efficiency = 58.0%) touching at a single vertex. To correct these non-manifold
neighborhoods, we extend (dilate) the chart to include addi-
tional samples. For each non-manifold sample 𝑝, we make it
manifold by making “defined” a minimal number of its
undefined neighbors. We assign these new samples the same
position as 𝑝. We iterate until all samples are manifold.

(c) our method (packing efficiency = 75.6%) There are cases when our algorithm can fail, particularly
when the requested geometry image area is small. Some
Figure 3: Our improvement on the packing of Lévy et al. discretized charts could become annuli for coarse samplings.
[2002]. For practical image resolutions this is not a problem.

unused space within the intersection of the upper and lower

envelope of the new chart (red region in the upper-left), since
a different rotation could avoid that wasted space.

4.4 Geometry image sampling

Given desired scale and orientations of the charts within the
geometry image, we discretize the surface charts into the
image. The image pixels will receive (𝑥, 𝑦, 𝑧) positions
sampled from the surface (Figure 4).
Figure 4: Discretization of the bunny.
The set of defined image samples implicitly create a triangu-
lation, using the rules described earlier in Section 3. We
require that this triangulation maintain the same topology as
the original surface. Satisfying this requirement involves two
sub-problems:
(1) Creating discretized charts that are topological disks, since
the surface charts are known to be disks.
(2) Connecting the charts together without cracks. The
explicit chart boundaries on the original mesh will help in this
zippering process, as explained in the next section.
We address the first sub-problem as follows. (The second
problem is the subject of the next section.) Figure 5: Rasterizing the boundary of the chart ensures
Interior rasterization. We begin by rasterizing each chart’s that it is connected. (Samples in red are added by
3D triangles into the 2D image domain using ordinary scan boundary rasterization.)
conversion. During rasterization, the value stored in the
image domain sample is the corresponding (𝑥, 𝑦, 𝑧) location
from the 3D triangle. The unwritten samples remain unde-
fined. This simple procedure does not guarantee that the
rasterized charts will have disk topology. For example,
undersampling a narrow chart “peninsula” leads to broken
charts (see Figure 5a).
Boundary rasterization. To prevent charts from breaking
into disconnected components, we rasterize the chart’s continuous param. discretized charts after dilation
boundary edges to add connecting samples (Figure 5b). In Figure 6: The discretized charts may have non-manifold
particular, for each boundary edge, we construct a 1-pixel structures. We repair these by local chart dilation.

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images
4.5 Geometry image zippering not affect the manifold property of the mesh and can be
To prevent cracks in the reconstructed mesh, we “zipper” removed during reconstruction or harmlessly rendered as
boundary samples along discrete boundaries of charts that is. Even with texture mapping, degenerate faces do not
adjoin on the 3D surface. The overall process is illustrated in introduce artifacts, since the texture samples are unified
Figure 7. just like the geometric ones.
Cut-Nodes. The zippering algorithm first identifies all the
cut-nodes in the mesh. A cut-node is a mesh vertex with
more than two adjacent charts; in other words, a chart corner.
For every chart adjacent to the chart corner, the cut-node must
be mapped to a unique boundary cut-node sample in the
image domain. We choose the closest boundary sample in
terms of 3D distance. In addition, cut-node placement is
constrained to preserve the clockwise-order of the cut-paths.
In practice the GIM resolution is large enough to map each
cut-node to a distinct cut-node sample. Cut-node samples
corresponding to the same cut-node are all assigned its 3D
geometric position.
(a) Domain: (b) Before zippering;
Cut-Paths. Cut-nodes form the end points of a set of edges boundary samples in black and surface cut-path in red
comprising the cut-path or boundary between two adjacent cut-node samples in pink
charts. Every cut-path in the mesh maps to two discrete cut-
paths (sets of boundary samples) in the image domain. We
unify boundary samples along these pairs of discrete cut-
paths. See Figure 7 for an example of the discrete cut-paths.
Snap. For each boundary sample in the discrete cut-path, we
snap (overwrite) its position to the closest point along the
continuous cut-path. This narrows the gap between the charts
but the mesh is still not watertight because of sampling
differences (Figure 7c).
Unify. To seal the boundaries completely, we then unify (c) Boundary samples snapped (d) Boundary samples unified
boundary samples. Two discrete cut-paths corresponding to to the cut-paths to form watertight mesh
the same continuous chart boundary generally contain Figure 7: Illustration of the zippering process.
different numbers of samples. So, simple 1-to-1 unification
along the two cut-paths is insufficient: more than two samples
must occasionally be unified.
A simple greedy algorithm produces high quality zipperings.
It performs a one-pass, lock-step traversal of the two discrete
cut-paths. Figure 8 shows two steps in this traversal. We (a) (b) (c)
begin the zipper at a cut-node sample (see Figure 8a), which Figure 8: Three stages of the unification algorithm.
has already been unified as described above. During the Green bubbles represent unified clusters.
traversal, we can advance the zipper along either of the
discrete cut-paths (see Figure 8b); in this case the newly
visited sample is assigned the geometric position of the 4.6 Geometry image optimization
current unified cluster. Alternatively, we can advance the
Having created a zippered, watertight MCGIM, we optimize
zipper along both of the discrete cut-paths (see Figure 8c); in
its defined samples to improve geometric fidelity to the
this case we create a new cluster of samples that is assigned
original mesh. As in feature-sensitive remeshing [Vorsatz et
the geometric position of one of these samples. From these
al. 2001], we seek to more accurately represent sharp features
three alternatives, we select the zipper advancement that
such as creases and corners. In addition, we want to smooth
alters the updated sample by the least amount.
and regularize the sample positions across the zippered chart
After unification, all samples along a cut-path are boundaries.
unified to at least one sample on the other cut-path. The We extend the mesh optimization technique of Hoppe et al.
mesh reconstructed from the discrete samples is water- [1993]. In that work, both the connectivity and vertex
tight. See Figure 7 for a close-up image of the unified positions of the approximating mesh are optimized to fit the
cut-paths in the bunny reconstruction. When two adja- reference surface. The fit accuracy is measured by sampling
cent boundary samples in the same cut-path are unified a dense set of points from the reference surface and minimiz-
together, the reconstruction forms a degenerate triangle, ing the squared distanced of those points from the
which is in fact just a line. These degenerate faces do approximating mesh.

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images
We express geometric accuracy as Peak Signal to Noise Ratio
PSNR = 20 log10 (𝑝𝑝𝑝𝑝/𝑑) , where 𝑝𝑝𝑝𝑝 is the bounding
box diagonal and 𝑑 is the symmetric rms Hausdorff error
(geometric distance) between the original mesh and the
approximation. We similarly measure PSNR for a normal
signal derived from the geometry (Section 3). A large normal
error suggests parametrization artifacts like anisotropy.
Geometry image comparison. We created MCGIMs for a
variety of models and compared them with GIMs from Gu et
al. [2002]. To produce a fair comparison between single-
chart and multi-chart geometry images, we also applied our
geometry image optimization (Section 4.6) to the prior GIM
results.
As shown in Table 1, we obtain significantly better PSNR
results using MCGIMs for the models in Figure 14. For the
same memory cost (i.e., including storage cost for our unde-
fined samples), the geometry PSNR for our horse and feline
meshes was more than 10dB higher than those of Gu et al. To
make a comparison for equivalent rendering cost, we created
new MCGIMs with the same number of defined vertices as
Figure 9: Comparison of two multi-chart geometry GIMs. The additional samples yield a further improvement
images before and after mesh optimization. of about 2dB. For the Buddha model of Figure 1, the opti-
In our setting, the mesh connectivity is completely deter- mized GIM has a PSNR of 65.5dB, whereas our equivalent-
mined by the grid structure and the vertex positions. Recall memory-cost MCGIM has a PSNR of 71.7 dB.
from Section 3 that each quad of the grid is split into two
triangle along its shorter diagonal. Thus, we only need to
optimize vertex positions, noting that the grid diagonals may models gargoyle horse dragon feline
change as vertices are moved. We optimize using a non- genus 0 0 1 2
linear conjugate-gradient solver, which converges to reasona-
257x257 GIM from Gu et al. [2002]
bly good answers even though the fit measure lacks
continuity due to diagonal flipping. stretch efficiency 67.8% 32.4% 42.2% 33.3%
# defined vertices 66,049 66,049 66,049 66,049
To keep the mesh from folding over during the fit, we add a
mesh smoothness term to the energy function being mini- # unique vertices 65,537 65,537 65,537 65,537
mized. This term is set to the 𝐿𝑝 norm of the dihedral angles geometry PSNR 79.4 73.4 70.2 68.0
of the mesh edges. Setting 𝑝=4 has produced good results. normal PSNR 30.3 19.4 20.1 17.6
Figure 9 shows examples of MCGIM models with sharp 257x257 GIM from Gu et al. + geometry image optimization
features that are improved using geometry image optimiza-
geometry PSNR 83.5 71.9 73.6 68.1
tion. The results are surprisingly good considering the
restricted mesh connectivities. normal PSNR 31.5 18.9 19.6 17.0
MCGIM (equivalent memory cost as Gu et al.)
5. Results
# charts 15 25 40 50
We have run our completely automatic pipeline on a large stretch efficiency 98.7% 99.2% 92.7% 99.1%
number of models. The process takes approximately 1-2
packing efficiency 72.7% 75.6% 73.1% 75.6%
hours per model. Chartification takes no more than 5 minutes,
parameterization no more than 15 minutes, sampling and dimensions 466x138 281x228 369x174 478x133
zippering less than a minute, and mesh optimization about 10 rectangle area 64,308 64,068 64,206 63,574
minutes. The remainder of the time is needed for the discrete # defined vertices 46,724 48,389 46,913 48,038
optimization used for packing. One could easily substitute # unique vertices 41,961 41,857 38,516 35,956
other packing schemes – packing is not one of our main geometry PSNR 83.8 84.6 77.9 79.5
contributions.
normal PSNR 33.7 30.8 22.9 28.7
We use two cost metrics to compare our results with other
MCGIM (equivalent rendering cost as Gu et al.)
approaches. Memory cost is the total number of samples in
the domain grid (i.e., domain rectangle area). Rendering cost # unique vertices 58,769 55,329 52,059 55,329
is the number of samples that must be processed (i.e., defined geometry PSNR 85.2 85.3 79.4 82.5
samples). Packing efficiency is the ratio of defined samples normal PSNR 35.5 30.6 23.1 30.2
to total number of samples. Thus, Table 1: Geometry image comparison results.
rendering cost = memory cost * packing efficiency .

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images
Texture-mapping using the implicit MCGIM parametrization
is demonstrated in Figure 10.
Unlike GIMs, MCGIMs can also represent meshes with
multiple components. The teapot in Figure 11 is an example
with four components. Note the excellent chartification and
packing.
Semi-regular remeshing comparison. We also made (a) chartified mesh (b) reconstructed mesh
comparisons with semi-regular remeshes created by Guskov (4 components; 30 charts) (PSNR=86.2db)
et al. [2000] and Lee et al. [1998] (Figure 12). We created
MCGIMs with rendering cost (i.e., number of vertices) no
larger than those of the semi-regular meshes. As shown in
Table 2, our PSNR results are approximately 3db higher than
those of the semi-regular meshes, representing a 30% reduc-
tion in Hausdorff error.

semi-regular meshes our MCGIM approach (c) geometry image charts

# unique geom. # unique geom. (dimensions=388x168; packing efficiency=81.3%)
vertices PSNR vertices PSNR Figure 11: As opposed to the GIM framework, MCGIMs
feline 258,046 84.4 256,862 88.3 can represent meshes with multiple components, such as
horse 112,642 87.8 99,851 90.2 this teapot.
Table 2: Semi-regular remeshing comparison results.

Original irregular mesh

MCGIM (3,800 faces) same coarse mesh,

(44x86 geometry image) normal-mapped

Semi-regular remeshing

(medium frequency) (near Nyquist rate)

texture-mapped using a sampled, projected stripe texture
Figure 10: An MCGIM texture-mapped with three
352x688 images. Texture coordinates are implicit, and
each triangle is exactly associated with 9x9/2 texture
samples. The resulting renderings are nearly seamless. Multi-chart geometry image
The problem is that some zippered triangles are longer
than interior ones, resulting in local undersampling. Figure 12: Horse close-up showing improved sampling
uniformity of MCGIM over semi-regular remeshing.

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images
CIGNONI, P., MONTANI, C., ROCCHINI, C., AND SCOPIGNO, R. 1998. A
79 general method for recovering attribute values on simplified meshes.
IEEE Visualization 1998, pp. 59-66.
78 ECK, M., DEROSE, T., DUCHAMP, T., HOPPE, H., LOUNSBERY, M., AND
STUETZLE, W. 1995. Multiresolution analysis of arbitrary meshes.
77 SIGGRAPH 1995, pp. 173-182.
FUCHS, H., KEDEM, Z., AND USELTON, S. 1977. Optimal surface
76 reconstruction from planar contours. Comm. ACM 20.
PSNR

GARLAND, M., WILLMOTT, A., AND HECKBERT, P. S. 2001. Hierarchical

75
face clustering on polygonal surfaces. 2001 ACM Symposium on
Interactive 3D Graphics, pp. 49-58.
GU, X., GORTLER, S. J., AND HOPPE, H. 2002. Geometry images.
74
SIGGRAPH 2002, pp. 355-361.
GUSKOV, I., VIDIMCE, K., SWELDENS, W., AND SCHRÖDER, P. 2000.
73
Normal meshes. SIGGRAPH 2000, pp. 95-102.
HOPPE, H., DEROSE, T., DUCHAMP, T., MCDONALD, J., AND STUETZLE,
72 W. 1993. Mesh optimization. SIGGRAPH 1993, pp. 19-26.
1 10 100 1000
KOBBELT, L., VORSATZ, J., LABSIK, U., AND SEIDEL, H.-P. 1999. A
number of charts shrink wrapping approach to remeshing polygonal surfaces. Eu-
rographics 1999, pp. 119-130.
Figure 13: MCGIM accuracy as a function of the num-
KHODAKOVSKY, A., SCHRÖDER, P., AND SWELDENS, W. 2000. Progres-
ber of charts on the horse model, for 1282 samples. sive geometry compression. SIGGRAPH 2000, pp. 271-278.
Selecting the number of charts. As shown by the graph in LEE, A., MORETON, H., HOPPE, H. 2000. Displaced subdivision surfaces.
SIGGRAPH 2000, pp. 85-94.
Figure 13, MCGIM accuracy exhibits a rough “hill” shape
LEE, A., SWELDENS, W., SCHRÖDER, P., COWSAR, L., AND DOBKIN, D.
with increasing number of charts. Noise in the graph data 1998. MAPS: Multiresolution adaptive parameterization of surfaces.
results from optimization procedures involved in chartifica- SIGGRAPH 1998, pp. 95-104.
tion, parameterization, zippering, and packing. The graph’s LÉVY, B., PETITJEAN, S., RAY, N., AND MAILLOT, J. 2002. Least squares
shape comes from two countervailing trends: MCGIMs with conformal maps for automatic texture atlas generation. SIGGRAPH
fewer charts use less gutter space and unify fewer samples; 2002, pp. 362-371.
MCGIMs with more charts produce less parametric distortion LLOYD, S. P. 1957. Least squares quantization in PCM. Unpublished,
reprinted as IEEE Trans. Inform. Theory, IT-28(2):127-135, 1982.
and more efficient packings. An optimal number of charts
MAILLOT, J., YAHIA, H., AND VERROUST, A. 1993. Interactive texture
thus exists for a given model, but good accuracy is obtained mapping. SIGGRAPH 1993, pp. 27-34.
over a broad range. MAX, J. 1960. Quantizing for minimum distortion. IEEE Trans.
Inform. Theory, IT-6(1): pages 7-12, March 1960.
6. Summary and future work MILENKOVIC, V. 1998. Rotational polygon containment and minimum
We introduced multi-chart geometry images, a new surface enclosure. Annual Symposium on Comp. geometry, ACM.
representation that extends the geometry images of Gu et al. PIPONI, D. AND BORSHUKOV, G. D. 2000. Seamless texture mapping of
subdivision surfaces by model pelting and texture blending. SIG-
[2002]. MCGIMs have the same regular grid structure of GRAPH 2000, pp. 471-478.
GIMs, except that some of their image samples are undefined.
SANDER, P. V., GORTLER, S. J., SNYDER, J., AND HOPPE, H. 2002. Signal-
This permits partitioning of the model into charts, and so specialized parametrization. Eurographics Rendering Workshop
obtains the flexibility needed to parametrize meshes of 2002.
arbitrary complexity, high genus, and multiple components SANDER, P. V., SNYDER, J., GORTLER, S. J., AND HOPPE, H. 2001.
with less distortion. Texture mapping progressive meshes. SIGGRAPH 2001, pp. 409-
416.
We showed that MCGIMs significantly exceed the geometric SHEFFER, A. AND HART, J. C. 2002. Seamster: Inconspicuous Low-
accuracy of GIMs for equivalent memory cost. Obviously, Distortion Texture Seam Layout. IEEE Visualization 2002.
this advantage increases if we neglect the cost of undefined SHLAFMAN, S., TAL, A., AND KATZ, S. 2002. Metamorphosis of polyhe-
samples, which are ignored in reconstruction and rendering. dral surfaces using decomposition. Eurographics 2002, pp. 219-228.
We also showed that MCGIMs yield better geometric approx- SORKINE, O., COHEN-OR, D., GOLDENTHAL, R., AND LISCHINSKI, D.
imations than those of the semi-regular remeshing of Guskov 2002. Bounded-distortion piecewise mesh parameterization. IEEE
Visualization 2002.
et al. [2000] and Lee et al. [1998].
TURK, G. AND LEVOY, M. 1994. Zippered polygon meshes from range
Several areas of future work remain, especially MCGIM images. SIGGRAPH 1994, pp. 311-318.
level-of-detail control and compression. We are also interest- VORSATZ, J., RÖSSL, C., KOBBELT, L. P., AND SEIDEL, H. 2001. Feature-
ed in rendering MCGIMs using graphics hardware, and sensitive remeshing. Computer Graphics Forum, 20(3):393--401,
2001.
exploiting their simplicity in future architectural designs.
WOOD, Z. J., DESBRUN, M., SCHRÖDER, P., AND BREEN, D. 2000. Semi-
regular mesh extraction from volumes. IEEE Visualization 2000, pp.
References 275-282.
CARR, N., AND HART, J. 2002. Meshed atlases for real-time procedural ZORIN, D., SCHRÖDER, P., AND SWELDENS, W. 1997. Interactive
solid texturing. ACM Transactions on Graphics. 21 (2), pp. 106- multiresolution mesh editing. SIGGRAPH 1997, pp. 259-268.
131.

© The Eurographics Association 2003.

 Sander et al. / Multi-Chart Geometry Images

genus 0; 15 charts genus 0; 25 charts genus 1; 40 charts genus 2; 50 charts

packing efficiency=72.7% packing efficiency=75.6% packing efficiency=73.1% packing efficiency=75.6%

PSNR=83.8 dB PSNR=84.6 dB PSNR=75.6 dB PSNR=79.5 dB

Original irregular meshes
[Gu 2002] geom. images
Multi-chart geom. images

Figure 14: Multi-chart geometry image results. Black squares indicate close-up views used in comparisons below.

© The Eurographics Association 2003.