Vertex Climax: Converting Geometry into a Non-nanifold Midsurface
Christoph Schinko
1,2
and Torsten Ullrich
1,2 a
1
Fraunhofer Austria Research GmbH, Graz, Austria
2
Technische Universit
¨
at Graz, Graz, Austria
Keywords:
Midsurface, Grid-based Algorithm, Volume Graphics, Design Automation.
Abstract:
The physical simulation of CAD models is usually performed using the finite elements method (FEM). If the
input CAD model has one dimension that is significantly smaller than its other dimensions, it is possible to
perform the physical simulation using thin shells only. While thin shells offer an enormous speed-up in any
simulation, the conversion of an arbitrary CAD model into a thin shell representation is extremely difficult
due to its non-uniqueness and its dependence on the simulation method used afterwards. The current state-of-
the-art algorithms in conversion voxelize the input geometry and remove voxels based on matched, predefined
local neighborhood configurations until only one layer of voxels remains.
In this article we discuss a new approach that can extract a midsurface of a thin solid using a kernel-based
approach: In contrast to other voxel-based thinning approaches, our algorithm applies a kernel onto a binary
grid. In the resulting density field, opposing surface-voxels are iteratively moved towards each other until a
thin representation is obtained.
1 INTRODUCTION
The analysis of thin-walled parts is a discipline of
Computer-Aided Engineering (CAE) analysis. As
physical simulations are computationally costly, they
benefit from reduced-complexity representations, so
called “Midsurfaces” (Kulkarni, 2016). The reduc-
tion in complexity is achieved by replacing such
Computer-Aided Design (CAD) models with a sim-
plified model consisting of non-manifold 2D-surfaces
lying (mostly) midway of the original walls.
Midsurfaces are part of the larger family of me-
dial objects. A common member of this family is the
medial axis.
Definition: The medial axis (also called medial sur-
face) of a closed object is the set of centers of empty
spheres which touch the surface of the object at more
than one point.
Another member of the family of medial objects
is the straight skeleton. The straight skeleton is de-
fined only for polygons and has been introduced by O.
AICHHOLZER and F. AURENHAMMER (Aichholzer
and Aurenhammer, 1996). In contrast to the medial
axis it consists only of line segments (see Figure 1).
The straight skeleton is defined by a shrinking pro-
cess of the polygon. The edges of the polygon are
a
https://orcid.org/0000-0002-7866-9762
Figure 1: The medial axis of some input geometry consists
of all points that are the center of a sphere which touches the
input at more than one point (left). The straight skeleton is
the result of a shrinking process returning a graph structure
of line segments only (right).
moved inwards in a self-parallel manner and at the
same speed. When all parts have shrunken to zero, the
process terminates. The straight skeleton now con-
sists of the paths drawn by the polygon vertices during
the shrinking process. The result can be seen as a vari-
ant of the medial axis, because, in the case of convex
polygons, both structures are identical (Demuth and
Aurenhammer, 2010).
Both, the definition of a medial axis as well as the
definition of the straight skeleton, only depend on the
input object. By contrast, the midsurface definition
not only depends on the input, but also on the output
of a subsequent analysis:
Definition: The midsurface of a closed, thin-walled
object is a set of non-manifold surfaces which re-
veal in a CAE analysis the same results (within a δ-
Schinko, C. and Ullrich, T.
Vertex Climax: Converting Geometry into a Non-nanifold Midsurface.
DOI: 10.5220/0010210401850192
In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 1: GRAPP, pages
185-192
ISBN: 978-989-758-488-6
Copyright
c
2021 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
185
tolerance) as the closed, thin-walled input model.
Therefore, any statements regarding the correct-
ness of a midsurface transformation can only be vali-
dated within a specific context. A possible difference
between the results of both transformation types is il-
lustrated in 2D in Figure 2.
Figure 2: The result of a medial axis transformation (left)
and the midsurface transformation of a thin-walled part
(grey) are usually different. As the midsurface has to re-
spect follow-up simulation usages, both results (left & right)
may be a correct midsurface depending on the application
context.
This article presents a novel voxel-based midsur-
face generation approach for thin-walled objects; the
first, preliminary sketch of this algorithm has been
presented in a poster session at the Fabrication &
Sculpting Event (FASE) at Shape Modeling Interna-
tional (SMI) 2020 (Schinko and Ullrich, 2020). In
contrast to the preliminary sketch, a detailed descrip-
tion and a thorough evaluation have been added to this
article.
2 RELATED WORK
Related work in the area of midsurface generation for
solid geometries can be divided into surface-based
approaches and grid-based (so-called thinning) ap-
proaches.
H. LOCKETT and M. GUENOV address the diffi-
culty in finding similarity measures for evaluating the
quality of midsurface models (Lockett and Guenov,
2008). Without a ground truth, the proposed meth-
ods perform a geometric and topological comparison
between the midsurface models and their solid coun-
terparts. A global measure of geometric similarity is
based on the Hausdorff distance (Ullrich et al., 2008)
and defines the proportion of solid model points that
are within a specified threshold distance of the mid-
surface model. For measuring the global topological
similarity, geometry graph attributes are used. Both
measures can be combined into an overall similarity
index to provide interesting insights.
2.1 Top-down
The group of top-down approaches are surface-
based approaches working on boundary representa-
tions (Schinko et al., 2017). A divide and conquer
strategy simplifies the transformation into midsur-
faces by splitting complex parts into smaller sub-parts
according to semantic interrelationships (Sun et al.,
2016). This kind of approach is also used by Y.
H. WOO and C. U. CHOO (Woo and Choo, 2009).
They decompose a solid model into simple volumes
and generate midsurfaces using an existing algorithm
based on face-pairs (the top and the bottom side of
a thin layer are called face-pairs or corresponding
faces (Rezayat, 1996); in a boundary representation
these faces usually don’t have any explicit correspon-
dence).
In their work, NOLAN et al. rely on feature detec-
tion to assign mixed-dimensional finite element mod-
els to three categories: complex regions, long-slender
regions, and thin-sheet regions (Nolan et al., 2014).
Relations between the original solid model and the
dimensionally reduced version are maintained.
In the area of surface-based midsurface genera-
tion, the state-of-the-art approaches for decomposi-
tion (Kulkarni et al., 2017b), identification of cor-
responding faces (Kulkarni and Kale, 2014), and a
divide and conquer strategy (Kulkarni et al., 2017a)
have been presented by Y. H. KULKARNI: “Develop-
ment of Algorithms for Generating Connected Mid-
surfaces using Feature Information in Thin-Walled
Parts” (Kulkarni, 2016).
2.2 Bottom-up
Due to the nature of our approach being grid-based,
this Section focuses on a subgroup of so-called thin-
ning approaches. The idea behind these approaches
is to iteratively shrink an object locally by offset-
ting its boundary towards its interior. This process is
stopped, when the interior vanishes and just a (usu-
ally non-manifold) surface structure remains. Like
our approach, many thinning approaches work on bi-
nary voxels, a concept attributed to the field of digital
topology. Early work in this field dates back to the late
1960s and is surveyed by T.Y KONG and A. ROSEN-
FELD (Kong and Rosenfeld, 1989). A good descrip-
tion of toplogical structures of voxels and concepts
like neighborhood, connectivity and Euler number is
given by J. TORIWAKI and T. YONEKURA (Toriwaki
and Yonekura, 2002). Their work also introduces new
local features to study the effect of deleting 1-voxels
– a fundamental process during thinning.
Thinning approaches have a long history: P.
KWOK and V. RANJAN give a good overview of the
most important thinning algorithms up to 1991 (Kwok
and Ranjan, 1991). All presented algorithms work on
voxel data and eventually result in a skeleton rather
than a surface. Another volume thinning method for
generating a skeleton that can be used in searching
for intersected cells in isosurface propagation is pre-
GRAPP 2021 - 16th International Conference on Computer Graphics Theory and Applications
186
sented by ITOH et al. (Itoh et al., 1996). Their graph-
based approach connects all extremum points of a vol-
ume to a one-cell wide skeleton and contains its topo-
logical features. S. PROHASKA and H.-C. HEGE pro-
pose a robust, noise resistant criterion for the charac-
terization of plane-like skeletons (Prohaska and Hege,
2002). Their algorithm uses the geodesic distance
along the object’s boundary and a distance map. After
obtaining a skeleton of lines and surfaces, the focus is
on obtaining a surface representation for expressive
rendering of complex structures.
In contrast to the already mentioned algorithms
that return skeletons, the work by T.-C. LEE and
R. L. KASHYAP on thinning algorithms is con-
cerned with parallel thinning for extracting medial
surfaces and medial axes of binary voxel data (Lee
and Kashyap, 1994). The algorithm uses the Euler
characteristic to preserve topological properties of bi-
nary voxel data during the thinning process. Another
thinning algorithm for extracting medial surfaces is
presented by C.-M. MA and S.-Y. WAN (Ma and
Wan, 2001). It defines two sets of voxels on the
grid: one consisting of the 26-neighborhood of all 1-
voxels and one consisting of the 6-neighborhood of
all 0-voxels. The algorithm alternatively works on
the two sets deleting all possible 1-voxels, and ter-
minates when no 1-voxels can be deleted anymore. It
is proven to preserve connectivity.
The work of FUJIMORI et al. focuses on extract-
ing separated medial surfaces (Fujimori et al., 2006).
Medial surfaces are obtained by classifying voxels
into medial surface voxels using geodesic distances
along all three axes. Connected regions are separated
by using the average thickness as an indicator (a con-
stant thickness of all parts is assumed). The separa-
tion is performed by offsetting them into the direc-
tion of the object surface. In a final step, a polygo-
nal surface is extracted using a marching cubes ap-
proach. Another thinning approach for medial sur-
face extraction is presented by K. PAL
´
AGYI and G.
N
´
EMETH (Pal
´
agyi and N
´
emeth, 2009). The proposed
algorithms rely on sufficient conditions for parallel,
topology-preserving reduction operators. A novel
thinning scheme using iteration-level endpoint check-
ing is proposed in follow-up work (N
´
emeth et al.,
2010).
2.3 Surface Reconstruction
Using a grid-based data structure for thinning, our
approach has to convert the remaining voxels rep-
resenting the midsurface into a polygonal, non-
manifold representation. An important algorithm to
extract a polygonal representation of an isosurface
within a three-dimensional, discrete scalar field is
the well-known marching cubes algorithm by W. E.
LORENSEN and H. E. CLINE (Lorensen and Cline,
1987). It is in the nature of the marching cubes al-
gorithm to produce a dense triangulation (even on
flat surfaces) and voxel artifacts. In case normal
information is available for each voxel, surface re-
construction algorithms are able to produce a much
smoother mesh (Ju et al., 2002). Another approach
is presented by KAZHDAN et al. showing that sur-
face reconstruction of Hermite data can be formu-
lated as a Poisson problem (Kazhdan et al., 2006).
The algorithm reconstructs a watertight triangulation
of the surface of a model.While all these algorithms
produce a closed surface for all connected regions
of voxels, they are not suitable for midsurface re-
construction due to its non-manifold nature (without
volume and orientation). Q. T. NGUYEN and A.
J. P. GOMES propose an algorithm for triangulating
non-manifold implicit surfaces (Nguyen and Gomes,
2016). The idea is to leverage data generated by
marching cubes and to “heal” the triangulation inside
so-called critical cubes. Additional work on polygo-
nizing non-manifold implicit surfaces is presented by
YAMAZAKI et al. (Yamzaki et al., 2002). By allow-
ing discontinuity of the field function, the proposed
method yields a non-manifold surface exhibiting fea-
tures like holes and boundaries. Another approach to
reconstruct medial surfaces from volumetric data of
thin-plate objects is presented by T. MICHIKAWA and
H. SUZUKI (Michikawa and Suzuki, 2010). Their al-
gorithm is based on distance fields of binary volumes
and uses spherical support to build correct junctions
where conventional approaches produce small cavi-
ties.
As the result of all midsurface generation algo-
rithms require manual inspection and corrections, the
manual and the automatic creation of a midsurface
still cause a similar effort.
3 VERTEX CLIMAX
Our new grid-based approach consists of four steps,
which are described shortly to give an overview, and
which are illustrated in the Figures 3 and 4. The first
step builds a binary 3D grid (see Figure 3, left) and
marks each voxel (light red) that contains some part
of the input surface geometry (dark red).
The second step interprets the binary grid as a den-
sity field with the previous markup as initial 0.0 resp.
1.0 values and applies a kernel Σ to it. In Figure 3
(right) the kernel is a simple counter (light red) that
sums the values of each 3 × 3 neighborhood N
9
. The
Vertex Climax: Converting Geometry into a Non-nanifold Midsurface
187
illustration shows the color-coded result.
Σ
number of
1-voxels in N
9
:
1234567
Figure 3: The new “vertex climax” algorithm determines
the midsurface of a geometric model. In the first step (left)
the input surface geometry is rasterized into a voxel rep-
resentation. Having rasterized the input surface geometry
in binary voxels, an accumulation kernel Σ transforms the
binary voxels into a density field (right).
In the third step (see Figure 4, left), the vertices
of the original input data are moved within the den-
sity field towards the climax (black arrows in the color
coded grid).
Finally, vertices which approach each other are
clustered and merged (see grey circles in Figure 4).
Having merged all clusters of vertices, the resulting
two-sided, incident surfaces are merged to one-sided
surfaces. The result is a midsurface.
Figure 4: The accumulated density field defines a direction:
The vertices of the input surface geometry are moved to-
wards the density climax (left). The merge step (right) is
performed by a mean-shift clustering algorithm (illustrated
by grey circles). While merging the vertices, the surfaces
(top and bottom of a geometry; dotted lines) merge as well
(black line).
3.1 Rasterization Step
The Vertex Climax algorithm can be applied to trian-
gle meshes only. This prerequisite is not a limitation
as most CAD systems offer the possibility to export
the tessellated geometry.Besides the tessellated ge-
ometry the algorithm needs two additional values: ε
and δ, which are the smallest / greatest distance be-
tween two opposite surfaces that should be merged to
one midsurface. All other subsequent parameters (see
Table 2) are set automatically and do not need to be
modified.
The first step builds a binary 3D grid and marks
each voxel that has a non-empty intersection with the
input surface geometry (Ogayar-Anguita et al., 2020).
If the grid size α, i.e. the edge length of each cube in
the voxel grid, is not specified, the algorithm uses the
default value α =
1
2
ε.
3.2 Voxel Kernel
The second step interprets the binary grid as a density
field with the values 0.0, if a voxel does not contain
any geometry, and 1.0 otherwise. The main idea of
the second step is to spread the density values using
a kernel. The overlaps of the kernels at the position
of the midsurface result in significantly higher values
compared to their neighborhood.
The vertex climax uses a uniform distribution over
a 3D sphere with radius κ. The implementation sim-
ply sums for each voxel (i, j, k) the values of all sur-
rounding voxels (r, s, t) with an Euclidean distance
less than κ between the voxels’ centers. Although
many computer vision algorithms use other metrics,
the Euclidean distance was chosen because it best cor-
responds to the approach of a manual midsurface con-
struction, which is the blueprint for this algorithm.
3.3 Vertex Climax
In the third step, the vertices of the original input tes-
sellation data are moved within the density field gen-
erated in the previous steps. The start position of a
moving vertex is the center of its corresponding voxel.
Then each vertex moves to the center of a neighboring
voxel with the highest density until
1. no improvement concerning density values is pos-
sible or
2. the length of the movement path exceeds the
greatest distance δ between two opposite surfaces
that should be merged to one midsurface.
3.4 Vertex Clustering & Topological
Cleaning
Having moved all vertices, the resulting set of vertices
is clustered using a mean-shift clustering algorithm
with a window size ω (with default value δ). The re-
sult is already a two-sided midsurface, in which each
part is represented by two incident surfaces. As a one-
sided midsurface representation is desired (the goal
is as little geometry as possible), a dense resampling
of the two-sided midsurface and a reconstruction ac-
cording to SHAWN and WATSON (Shawn and Watson,
2011) returns the desired result.
GRAPP 2021 - 16th International Conference on Computer Graphics Theory and Applications
188
4 EVALUATION
In order to evaluate the “Vertex Climax” algorithm,
we created five geometric configurations and applied
the new algorithm to them. The first data set consists
of two penetrating sheet metal parts of different size
and thickness, one of which is bent at one end. The
second data set consists of two pipes with different
diameters and wall thicknesses. The arrangement is
an acute-angled T-construction where the pipe with
the larger diameter is continuous. The third data set
consists of a widening U-shaped beam connected to a
pipe at a top outlet. The fourth configuration consists
of a bent sheet metal on which two pipes are aligned
at acute angles, i.e. the sheet metal is not penetrated
and the pipes are not continuous. The last data set is
a laser-scanned part which contains noise and which
does not meet the thin-solid requirements.
Table 1 illustrates each input data set and lists their
geometric properties; i.e. the number of triangles, the
size of the axis-aligned bounding box AABB, the
smallest distance ε and the greatest distance δ be-
tween two opposite surfaces that should be merged
to one midsurface. Please note, that the last data set
is not a thin solid. The measured values ε and δ refer
only to the parts that fulfill the thin-solid condition.
All evaluations of the “Vertex Climax” algorithm
use the heuristics of the default values (see Table 2).
No additional adjustment has been made. As a con-
sequence, having measured the geometric sizes ε and
δ, the algorithm runs fully automatically and requires
no user interaction.
4.1 Intermediate Steps
For thin-solid input geometry, vertex climax and ver-
tex clustering are the most time consuming steps (see
Table 3). When the thin-solid condition is violated
and a significant amount of grid cells is occupied, the
accumulation step contributes most to the overall run-
time of the algorithm.
Using the default values for the grid size, the re-
sulting grids vary between 0.2 × 10
6
cells and 50.7 ×
10
6
cells (for data set #2 and #3, respectively). Of all
data sets, between 0.1 × 10
6
and 12.6 × 10
6
cells are
occupied; i.e. between 19% upto 55% of all cells are
nonempty.
Figure 5 shows some intermediate results of the
rasterization and the subsequently applied kernel. In
order to have a better view on the results of the kernel,
each cell is rendered as a cube that is smaller than the
cell size; additionally, the volume is clipped in the
middle.
4.2 Geometric Distance
The vertex movement within the density field is lim-
ited by a maximum path length, which is the great-
est thickness δ. In combination with the mean-
shift-clustering, which is limited by the window size
ω, the Hausdorff distance (Ullrich et al., 2008) be-
tween the input surface geometry and the result-
ing midsurface is expected to be at the same level.
In the geometric analysis listed in Table 4, the
vertex movement including vertex clustering corre-
sponds to the one-sided distance of an actual-target-
comparison (Schinko et al., 2011). Its values (aver-
age ± standard-deviation per vertex) should be ap-
Table 1: The input data sets represent geometric configura-
tions commonly used in computer-aided geometric design.
Please note that in contrast to the data sets #1 – #4, the data
set #5 is not a thin solid. It has been included to demonstrate
limitations.
data set statistics
data set #1
# elements 84 triangles
AABB
x 149.416
y 174.837
z 155.954
ε 1.77
δ 4.84
data set #2
# elements 360 triangles
AABB
x 63.6580
y 167.720
z 172.135
ε 4.46
δ 5.22
data set #3
# elements 524 triangles
AABB
x 263.307
y 76.1719
z 212.861
ε 0.98
δ 5.65
data set #4
# elements 388 triangles
AABB
x 139.035
y 158.949
z 123.290
ε 1.51
δ 6.90
data set #5
# elements 20088 triangles
AABB
x 2.38100
y 4.04000
z 7.84600
ε 0.35
δ [ 0.60 ] (thin
solid parts only)
Vertex Climax: Converting Geometry into a Non-nanifold Midsurface
189
Table 2: The default parameter settings of the vertex climax
algorithm do not need to be modified in most application
cases.
param- default semantic
eter value
ε smallest thickness; i.e. the
smallest distance between two
opposite surfaces that should be
merged to one midsurface
δ greatest thickness; i.e. the
greatest distance between two
opposite surfaces that should be
merged to one midsurface
α
1
2
ε grid size; i.e. the edge length of
each cube in the voxel grid
κ
3
2
δ kernel radius; i.e. the support
radius of the uniform, spherical
kernel
ω δ window size of the mean shift
clustering algorithm
Table 3: The algorithm consists of four steps (rasterization,
accumulation, vertex climax and vertex clustering). For
thin-solid input data sets, vertex climax and vertex cluster-
ing are the most resource demanding steps of the algoritm.
data set #1 #2 #3 #4 #5
runtime [s]
step
#1
0.01 0.01 0.01 0.01 23.20
step
#2 0.01 0.01 0.01 0.01 2226.83
step #3
16.84 1.04 162.70 27.02 28.61
step
#4
11.09 0.83 97.64 17.01 16.38
total 27.95 1.89 260.36 44.05 2295.02
proximately in the range between ε and δ. Table 4
confirms these expectations for the data sets.
4.3 Final Midsurfaces
The geometric distance analysis can only detect gross
errors. As a consequence, the quality of the resulting
midsurfaces has been inspected and evaluated manu-
ally by several engineers with experience in construc-
tion and simulation. Figure 6 shows the input data sets
(wire frame in blue) and the corresponding midsur-
faces (in green) of the data sets #1 to #4. The results
fully meet the requirements for midsurfaces.
Nevertheless, a minor problem exist: The limited
precision is based on the grid-based approach and its
cell size; but this problem can be solved at the expense
of resources (memory and time).
The example data set #5, which is not a thin solid,
provides interesting insights into how the algorithm
works (see Figure 7). As the input data set contains
holes and is not manifold, the resulting midsurface
data set #1 data set #2
data set #3 data set #4
Figure 5: The main idea behind using a kernel is to spread
the density values. The overlaps of the kernels at the posi-
tion of the midsurface result in significantly higher values
compared to their neighborhood.
Table 4: The characteristic attributes of the input surface ge-
ometry (length of the diagonal of the axis-aligned bounding
box, smallest thickness ε, and greatest thickness δ) allow a
quantitative comparison of input surface geometry and re-
sulting midsurface using the average displacement of ver-
tices and triangles during the vertex climax and vertex clus-
tering steps. The overall comparison is quantified by the
Hausdorff distance.
data
set
#1 #2 #3 #4 #5
AABB 277.875 248.622 347.048 244.532 9.141
ε 1.770 4.460 0.980 1.510 0.350
δ 4.840 5.220 5.650 6.900 [
0.60 ]
one-sided
distance
(a
vg)
1.912 1.767 1.512 1.552 0.250
(± std)
0.793 1.253 1.182 1.615 0.087
Hausdorf
f
4.257 5.412 5.039 5.400 0.502
contains holes as well. Due to the grid-based round-
ing and the vertex clustering, the effects are intensify-
ing and the size of the holes increases. In areas which
do not meet the thin solid criteria, the density field
does not reveal a consistent directional field. There-
fore, the vertex movements within the density field
result in self-intersections, which cannot always be
cleared by the mean-shift clustering algorithm.
GRAPP 2021 - 16th International Conference on Computer Graphics Theory and Applications
190
data set #1 data set #2
data set #3 data set #4
Figure 6: This overview shows the input CAD models (wire
frame in blue) and the resulting midsurfaces calculated by
the vertex climax algorithm (in green).
density field “midsurface”
Figure 7: If the input surface geometry is not a thin solid
(data set #5), then the accumulated densities (left) do not
lead to a directional field that points towards the position of
a midsurface. As a consequence the result is undefined.
5 CONCLUSION
A midsurface consists of surface elements (2D) in or-
der to represent three-dimensional, thin solids whose
local thickness is small compared to its other dimen-
sions. This important representations allow a signif-
icant acceleration of almost all physical simulations.
Unfortunately, midsurfaces are not unique and depend
on their simulation contexts. Although the creation is
a time-consuming, manual step that requires knowl-
edge and experience, it is still the “gold-standard” to
speed up a simulation, if FEM can be replaced by
thin-shell simulations.
5.1 Contribution
The presented approach to extract a midsurface is ap-
plicable to thin-walled parts with a 4-step pipeline
consisting of rasterization of the input geometry to
obtain a grid-based representation, the generation of a
density field using a kernel, moving opposing vertices
by a density field to iteratively create a thin represen-
tation, and finally a clustering and cleaning step. The
first results of this algorithm deliver satisfying results.
5.2 Benefit
The presented approach generates a midsurface with-
out any manual interaction. Only two geometric val-
ues (the smallest and the greatest thickness) have to
be measured in advance. Once started, no user fur-
ther interaction is needed. As a consequence, the new
approach is suitable for fast feedback loops and rapid
prototyping.
5.3 Open Questions & Future Work
The vertex climax algorithm has only one minor is-
sue: the limited precision. The limited precision will
be addressed by combining the vertex climax algo-
rithm with an automatic mesh generation algorithm.
Having knowledge on the finite element simulation
performed on a midsurface allows to limit the dis-
cretization artifacts. A locally optimized cell size may
prevent additional discretization errors (on top of the
mesh generation).
In the future, developments will concentrate on
the introduction of local parameters removing the
global constants describing the smallest thickness ε,
and greatest thickness δ. The goal is to determine
these values locally and automatically. In the next
step, the influence of these parameters will be exam-
ined in more detail.
ACKNOWLEDGEMENTS
The authors would like to thank Ulrich Krispel and
Andreas Riffnaller-Schiefer for fruitful discussions;
Volker Settgast for providing the input data sets and
for his expertise in modeling and construction; Peer
Bertram and Martin Schwarz for their expert knowl-
edge in construction and simulation.
REFERENCES
Aichholzer, O. and Aurenhammer, F. (1996). Straight skele-
tons for general polygonal figures in the plane. Pro-
ceedings of the International Computing and Combi-
natorics Conference, 2:117–126.
Demuth, M. and Aurenhammer, Franz und Pinz, A. (2010).
Straight skeletons for binary shapes. Computer Vision
Vertex Climax: Converting Geometry into a Non-nanifold Midsurface
191
and Pattern Recognition Workshops (CVPRW), 8:9–
16.
Fujimori, T., Kobayashi, Y., and Suzuki, H. (2006). Sep-
arated medial surface extraction from ct data of ma-
chine parts. Geometric Modeling and Processing,
4077 (LNCS):313–324.
Itoh, T., Yamaguchi, Y., and Koyamada, K. (1996). Volume
thinning for automatic isosurface propagation. Visual-
ization, 7:303–310.
Ju, T., Losasso, F., Schaefer, S., and Waren, J. (2002). Dual
contouring of hermite data. ACM Transactions on
Graphics, 21:339–346.
Kazhdan, M., Bolitho, M., and Hoppe, H. (2006). Poisson
surface reconstruction. Symposium on Geometry Pro-
cessing, 4:61–70.
Kong, T. and Rosenfeld, A. (1989). Digital topology: Intro-
duction and survey. Computer Vision, Graphics, and
Image Processing, 48(3):357–393.
Kulkarni, Y. H. (2016). Development of algorithms for gen-
erating connected midsurfaces using feature informa-
tion in thin walled parts. PhD Thesis at Savitribai
Phule Pune University.
Kulkarni, Y. H. and Kale, M. (2014). Formulating mid-
surface using shape transformations of form features.
Proceedings of All India Manufacturing Technology,
Design and Research, 5:981–985.
Kulkarni, Y. H., Sahasrabudhe, A., and Kale, M. (2017a).
Dimension-reduction technique for polygons. Inter-
national Journal of Computer-Aided Engineering and
Technology, 9:1–17.
Kulkarni, Y. H., Sahasrabudhe, A., and Kale, M. (2017b).
Leveraging feature generalization and decomposition
to compute a wellconnected midsurface. Engineering
with Computers, 33:159–170.
Kwok, P. C. K. and Ranjan, V. (1991). A survey of 3-d
thinning algorithms. Vision Interface, 6:13–20.
Lee, T.-C. and Kashyap, R. L. (1994). Building skele-
ton models via 3-d medial surface/axis thinning al-
gorithms. Graphical Models and Image Processing
(CVGIP), 56:462–478.
Lockett, H. and Guenov, M. (2008). Similarity measures
for mid-surface quality evaluation. Computer-Aided
Design, 40:368–380.
Lorensen, W. E. and Cline, H. E. (1987). Marching cubes:
A high resolution 3d surface construction algorithm.
ACM SIGGRAPH Computer Graphics, 21:163–169.
Ma, C.-M. and Wan, S.-Y. (2001). A medial-surface ori-
ented 3-d two-subfield thinning algorithm. Pattern
Recognition Letters, 22:1439–1446.
Michikawa, T. and Suzuki, H. (2010). Non-manifold medial
surface reconstruction from volumetric data. Geomet-
ric Modeling and Processing, 6130 (LNCS):124–136.
N
´
emeth, G., Kardos, P., and Pal
´
agyi, K. (2010). Topol-
ogy preserving 3d thinning algorithms using four and
eight subfields. Image Analysis and Recognition,
6111:316–325.
Nguyen, Q. T. and Gomes, A. J. P. (2016). Healed marching
cubes algorithm for non-manifold implicit surfaces.
Encontro Portugues de Computacao Grafica e Inter-
acao (EPCGI), 23:2:1–8.
Nolan, D. C., Tierney, C. M., Armstrong, C. G., Robinson,
T. T., and Makem, J. E. (2014). Automatic dimen-
sional reduction and meshing of stiffened thin-wall
structures. Engineering with Computers, 30:689–701.
Ogayar-Anguita, C.-J., Rueda-Ruiz, A.-J., Segura-Sanchez,
R.-J., Diaz-Medina, M., and Garcia-Fernandez, A. L.
(2020). A gpu-based framework for generating im-
plicit datasets of voxelized polygonal models for the
training of 3d convolutional neural networks. IEEE
Access, 8:12675–12687.
Pal
´
agyi, K. and N
´
emeth, G. (2009). Fully parallel 3d
thinning algorithms based on sufficient conditions for
topology preservation. Discrete Geometry for Com-
puter Imagery (DGCI), 5810 (LNCS):481–492.
Prohaska, S. and Hege, H.-C. (2002). Fast visualization of
plane-like structures in voxel data. IEEE Visualiza-
tion, 13:29–36.
Rezayat, M. (1996). Midsurface abstraction from 3d solid
models: general theory and applications. Computer-
Aided Design, 28:905–915.
Schinko, C., Riffnaller-Schiefer, A., Krispel, U., Eggeling,
E., and Ullrich, T. (2017). State-of-the-art overview
on 3d model representations and transformations in
the context of computer-aided design. International
Journal on Advances in Software, 3 & 4:446–458.
Schinko, C. and Ullrich, T. (2020). A new grid-based mid-
surface generation algorithm. Hyperseeing – the Pub-
lication of the International Society of the Arts, Math-
ematics, and Architecture, 19:81–84.
Schinko, C., Ullrich, T., Schiffer, T., and Fellner, D. W.
(2011). Variance Analysis and Comparison in
Computer-Aided Design. Proceedings of the Inter-
national Workshop on 3D Virtual Reconstruction and
Visualization of Complex Architectures, XXXVIII-
5/W16:3B21–25.
Shawn, M. and Watson, J.-P. (2011). Non-manifold
surface reconstruction from high-dimensional point
cloud data. Computational Geometry, 44:427–441.
Sun, L., Tierney, C. M., Armstrong, C. G., and Robin-
son, T. T. (2016). Automatic decomposition of com-
plex thin walled cad models for hexahedral dominant
meshing. Procedia Engineering, 163:225–237.
Toriwaki, J. and Yonekura, T. (2002). Euler number and
connectivity indexes of a three dimensional digital
picture. Forma, 17:183–209.
Ullrich, T., Settgast, V., and Fellner, D. W. (2008). Ab-
stand: Distance Visualization for Geometric Analy-
sis. Project Paper Proceedings of the Conference on
Virtual Systems and MultiMedia Dedicated to Digital
Heritage (VSMM), 14:334–340.
Woo, Y.-H. and Choo, C.-U. (2009). Automatic genera-
tion of mid-surfaces of solid models by maximal vol-
ume decomposition. Transactions of the Society of
CAD/CAM Engineers, 14:297–305.
Yamzaki, S., Kae, K., and Ikeuchi, K. (2002). Non-
manifold implicit surfaces based on discontinuous im-
plicitization and polygonization. Geometric Modeling
and Processing, 12:138–146.
GRAPP 2021 - 16th International Conference on Computer Graphics Theory and Applications
192