You are on page 1of 22

Quality Measure of Quadrilateral Meshes

Phillip Lu
Advised by Dianna Xu
Haverford College, Department of Computer Science
Abstract
A method of analyzing the quality of quadrilateral meshing algorithm is presented. By running each quadrilateral of the mesh through
John Robinsons Continuum-Region-Element (CRE) method, the resultant numeric values correlate with the shape parameters of each
quadrilateral. We find that Robinsons Jacobian determinant, even
after accounting for scale, is only as good as a mesh quality measure
as aspect ratio as defined by the CRE method.
Keywords. quadrilateral mesh, mesh quality, mesh analysis

Introduction

Polygonal meshes are discrete, approximate representations of 2D or 3D


shapes constructed from simple polygons: a collection of points, edges, and
polygonal faces that represents a 2D or 3D object. Though mainly used in
computer graphics, these meshes have a variety of applications, from highorder surface modelling to data compression [3].
Since most 3D geometrical objects have two dominant directions, quadrilateral meshes are particularly well suited for representing such objects, given
that individual quadrilaterals have two pairs of edges that run (ideally) orthogonal to each other [3]. Applications for which preserving the shape and
form of the 3D object is of importance would therefore prefer quadrilateral
meshes over triangular meshes, since transformations of a triangular mesh
would not guarantee that the overall shape is preserved. However, due to the
difficulties inherent in creating quadrilateral meshes, the majority of research
1

effort has gone into triangle mesh construction. As such, though there are
several methods of measuring the quality of a triangle mesh [2, 5], there is
no widely accepted similar metric to measure the quality of a quadrilateral
mesh.
Beyond simple metrics like the amount of polygons in a mesh, it is difficult
to find a measurement of quality that represents the entire mesh without
resorting to measurement of the individual polygonal elements. As such,
almost all mesh quality measures examine the shape quality of each polygon.
As Eppstein states in his presentation to the Meshing Roundtable in 2001,
there are numerous measures of a triangle, such as ratio of circumcricle to
incircle radii, ratio of diameter to height, or perimeter squared to area [5].
However, he mentions that the quality guarantees for quad/hex meshes [are]
much less developed. It is easy to see that finding a metric for triangles is
easier than finding one for quadrilaterals: while triangles can always be inscribed within a circle (the smaller the circle is, the better a triangle is),
not all quadrilaterals can be inscribed within a circle (i.e. non-cyclic quadrilaterals). While diameter to height is a consistent measure of a triangles
quality, it is not for that of a quadrilaterals, as using that measure would
return the same value for both squares and a rhombus (same diameter, same
height). However, we would not want the measure for a square and rhombus to be the same, since the preference for quadrilateral meshes over other
types of meshes is partially due to the quadrilateral having two dominant
local directions, typically associated with principal curvature directions. We
would thus prefer to have a mesh composed of mostly squares, where the two
dominant local directions are, for the most part, orthogonal to each other,
and thus accurately representing curvature directions, as opposed to a mesh
composed of mostly rhombuses, especially extremely tapered ones.
Measuring the quality of a quadrilateral is significantly more difficult than
that of a triangle, since there are more geometric features in a quadrilateral
that need to be properly reflected in the metric itself (i.e. skew, aspect ratio)
than in a triangle. While John Robinson was primarily concerned about
finite element stress testing in his Continuum-Region-Element (CRE) method
paper, he shows that the determinant of the quadrilaterals Jacobian matrix
is a numerical variable that reflects all significant features of a quadrilateral
[8]. This paper will attempt to evaluate the Jacobian determinants viability
as a measurement of quadrilateral mesh.
Throughout this paper, I will use the shorter term quad to refer to
quadrilateral.
2

Related Works

As mentioned before, this paper will be an evaluation of the validity of Robinsons CRE method [8] in mesh quality measurement. Thus, this paper will
draw heavily from the terminology and methods of Robinsons paper and
book (i.e. aspect ratio, skew, and taper are all defined within his works,
and will be integrated within my analysis). Robinson was concerned about
having a general method of finite element stress testing: a method of analyzing the forces on and within physical structures (planes, bridges, dams, etc.)
by breaking them down into many small discrete structural pieces (finite
element) [9]. Since these individual finite elements are made up of simple polygons (polygons with non-intersecting perimeters), there is significant
work done on analyzing the effects of pressure, torque, and bending on materials modeled by simple polygons. As such, the CRE method, while originally
developed as an element testing procedure, can be used to measure a quads
aspect ratio, skew, and taper, amongst other measurements of shape.
The first quad mesh algorithm I will be testing my quality measurement
on will be Atalay, Ramaswami, and Xus quad-tree algorithm [1] for generating a quad mesh. I will be using the mesh quality measures section of
their paper and comparing my results with theirs. The algorithm itself will
be discussed in section 2.1
While the Bommes et al. survey of quad-mesh algorithms [3] contains
multiple quad-mesh algorithms, few would mesh over a 2D point cloud, fewer
still have publicly available executables. Thus, I will be using Jonathan
Richard Shewchuks publicly available triangle mesh generator [10] to create
an initial triangle mesh, from which I will convert to a quadrilateral mesh.
A discussion of this process will be in section 2.2.

2.1

Atalay, Ramaswami, and Xus Quadtree Mesh Algorithm

Atalay, Ramaswami, and Xus algorithm for generating quad meshes relies on
using quadtrees effectively [1]. A quadtree, effectively a tree data structure
such that each node has either no children (leaf), or exactly 4 children. Atalay
et al.s algorithm first takes the point set, and creates a quadtree, each node
splitting whenever it has more than two elements in it. This effectively
means that each node will contain at most one point. Further, Atalay et al.s
algorithm ensures that no two neighbouring structure are two or more levels
3

Figure 1: A visualization of a quadtree with associated point set. [6]


apart, and uses this property to apply quad templates for the deepest level
of the subdivision. After that, the algorithm can apply general templates
for stitching together nodes of arbitrary levels. The existence of general
templates allows geometric analysis of the worst-case configuration, which
yields the minimum angle of 18.43 and a maximum angle of 171.86.
Atalay et al.s algorithm takes 2D point set information and outputs a
Geomview Object File Format (.OFF) file, that consists of the meshs vertices
and faces.

2.2

Shewchuks Triangle

As previously mentioned, I will be running the analysis program over the


outputs of Shewchuks Triangle program. Triangle can take in point sets,
4

and output a Delaunay triangulation (a triangulation such that no point is


inside the circumcircle of any triangle in the triangulation). A Delaunay triangulation will maximize the minimum angle of the resultant triangle mesh,
given that no additional points can be added [4]. Every point set is also
guaranteed to have a Delaunay triangulation.
An advantage to using Shewchuks Triangle program is that I am able to
specify certain quality guarantees. I will thus be using Triangle to output two
sets of data: one with no quality guarantee, and one with a quality guarantee
that the minimum angle will be at least 20 degrees.
This by itself should not be of any concern. After all, my goal here is not
to analyze the quad meshing algorithms in-depth, but instead to use Robinsons metrics on the meshes and comparing the result. However, Shewchuks
Triangle only outputs triangle meshes, and so we need a way to convert the
triangle mesh into a quad mesh. While there exists several methods of tri-toquad mesh algorithms, few of them offer an implementation of them online.
Instead, we will do a crude tri-to-quad algorithm that is simple to implement,
though offers nothing in terms of quality guarantees.
For each triangle in the triangle mesh, we add four additional points at
the centroid, and the three midpoints at their respective edges. We then
add three edges, from centroid to the three midpoints. We now have three
quadrilaterals in place of the triangle. We do this conversion for all the
triangles in the mesh, and we will end up with a quad mesh. See Figure 2
for example.
While this method is crude and offers no specific guarantee, this algorithm
will suffice for providing a mesh generator to compare to. Generally, we
expect a tri-to-quad mesh algorithm to have worse metrics than direct quad
mesh generation algorithms, especially tri-to-quad mesh algorithms that do
not apply mesh smoothing or mesh simplification afterwards. This gives us a
very general hypothesis to test: our metric for mesh quality should be better
on Atalay et. als algorithm than on Shewchuks algorithm after tri-to-quad
mesh conversion.

Method

The CRE method was originally developed as a procedure for understanding


how forces work on and within 3-dimensional, complex objects, represented
by and broken down into numerous discrete, elementary pieces. While the
5

Figure 2: Making a quad mesh from a triangle mesh.

(a) Shewchuks Triangles Delaunay triangulation of 50 random points.

(b) Quad mesh by running the tri-to-quad mesh algorithm over the triangulation.

method was designed to work on any shape of element with any number of
nodes, the elementary shape he derives all the equations in the paper from
is the quad [8]. Since bad shapes can affect the results of finite element
analysis, Robinson developed the CRE method to find the shape parameters that differentiate one shape to another, so as to be able to identify
bad shapes prior to conducting finite element analysis via an automated
analysis system.
In his paper, Robinson shows that any quad shape can be constructed
given four parameters: aspect ratio (AR), skew (), and the respective taper
in two principle directions (Tx , Ty ) [8]. He further shows that given the coordinates of the four points on the quad, one can plug in equations to find all
four shape parameters, the proof of which can be found in Robinsons paper.
I will explain these equations in detail in the sections 3.1 and 3.2 below.
The method and structure for comparing the different algorithms will be
discussed in section 3.3.

3.1

Local Cartesian Coordinates

Each of our quads exist as four points in some global Cartesian coordinate
system. To ease calculations and to make sure the quad is in the right
orientation (see Figure 3), we need to convert to a local Cartesian coordinate
system [9]. To do this, the four vectors from the centroid going through

the four bisectors of the quad edges is used (i.e. V 05 being the vector from
the centroid to the bisector of the segment from vertex 1 to vertex 2. See

Figure 4). The corresponding V 06 is used as the vector to define the local x

vector (
x ). To find the corresponding y vector, we cross V 06 with V 07 to
get the normal vector to the plane of the quad (corresponding to the local

z vector,
z ). With
x and
z defined, we can then find the
y by crossing

z with x . x and y will be referred to as x and y in the rest of the paper,

x-axis being the axis that is in the direction of


x.
With the local coordinate system defined, we can now convert the four
points into the new coordinate system by projecting the old Cartesian coordinates onto the new vectors, like so:

xi = V 0i

yi = V 0i
7

where i from 1 to 4 indicates the four respective vertices of the quad, and

V 0i is the vector from the centroid to that particular vertex. Thus, we now
have all four vertices in the local coordinate system.
The interpolation function from the old system to the new system can
be also expressed in the following equation, where and are curvilinear
coordinates with limits 1;
x = e1 + e2 + e3 + e4
y = f1 + f2 + f3 + f4
The curvilinear coordinates (, ) of the four points are (1, 1), (1, 1),
(1, 1), (1, 1) for points 1, 2, 3, and 4 respectively (Figure 3). The e and f
coefficients are given in the following equations:

Figure 3: Points oriented counter-clockwise starting from bottom left.


Point 0 indicates centroid. Bisectors shown.

Figure 4: Midpoints defined in counter-clockwise fashion

1
1
f1 = (y1 + y2 + y3 + y4 )
e1 = (x1 + x2 + x3 + x4 )
4
4
1
1
e2 = (x1 + x2 + x3 x4 )
f2 = (y1 + y2 + y3 y4 )
4
4
1
1
f3 = (y1 y2 + y3 + y4 )
e3 = (x1 x2 + x3 + x4 )
4
4
1
1
e4 = (x1 x2 + x3 x4 )
f4 = (y1 y2 + y3 y4 )
4
4
We can now produce the shape parameters by using the coefficients produced above.

3.2

Shape Parameters

The four coefficients of e and f each are closely related to the features of the
quad. While I will be explaining how each coefficients affect the features,
Figure 5 provides a quick glance at what the coefficients do.
e1 and f1 may be the most straight-forward coefficients to explain. Since
e1 is the sum of the four x coordinates divided by four, this is the average
value of x. Likewise, f1 represents the average value of y. The coordinates
(e1 , f1 ) thus points to the centroid of the quad.
9

e2 and f3 both represent a similar concept: the half-length of the edges of


the original rectangle. We can see this by construing e2 as actually being
1
1
4 ((x2 x1 ) + (x3 x4 )) and f3 as being 4 ((y3 y2 ) + (y4 y1 )). Keeping in
mind the orientation of the points (Figure 3), (x2 x1 ) and (x3 x4 ) represent
the edges of the original rectangle parallel to the x-axis, while (y3 y2 ) and

Figure 5: Physical meaning of the e and f coefficients [8]


10

(y4 y1 ) represent the edges of the original rectangle parallel to the y-axis. By
halving the respective sums, we find the average length of the line segments
parallel to their respective axis. By halving the result again, we find the
half-length of the line segments.
e3 and f2 both represent a similar concept: the length between the bisector of a segment on the original rectangle to its respective point on a
skewed parallelogram (see Figure 5c). Again, we can see that e3 is actually
1
1 1
2 ( 2 (x3 x2 ) + 2 (x4 x1 )). Since (x3 x2 ) and (x4 x1 ) represent twice the
distance between the vertex to the respective point on the original rectangle in the x-axis, 12 (x3 x2 ) and 12 (x4 x1 ) represent just the distance.
1
1 1
2 ( 2 (x3 x2 ) + 2 (x4 x1 )) thus represents the average distance of the two.
The logic follows for f2 on the y-axis.
e4 and f4 both represent the distance from a point on the tapered quadrilateral (resembling a trapezoid) to its respective point on the original rectangle. The logic is similar to that of the other coefficients. Taper causes
points 1 and 3 to move in one direction, and points 2 and 4 to move in the
other (as seen when comparing the movement of the vertices on a rectangle
as it gradually becomes a trapezoid). Thus, to measure the average distance
of movement on the x-axis, we add x1 and x3 while subtracting x2 and x4 to
find the total distance of taper. Dividing by 4 yields the average distance
of taper.
The shape parameters are given in Robinsons paper as the following [8]:
e2
f3
e3
skew =
f3
f4
taper in the x-direction(Tx ) =
f3
e4
taper in the y-direction(Ty ) =
e2
aspect ratio =

or

f3
e2

(largest)

We can easily see how e2 and f3 are related to the aspect ratio. e2
represents the half-width, f3 represents the half-length; by dividing one over
the other and taking the bigger of the two, we get the aspect ratio (width
over length, or length over width, whichever is bigger).
Regarding skew, since e3 represents the absolute distance of skew in the
x direction, we need a way of scaling so that if the shape is identical, our
11

metric for skew does not increase as the shape becomes larger. f3 is chosen
for this, as it is a scale of the associated parallelogram in the y direction.
e3
Thus, skew =
f3
Regarding the two tapers, e4 and f4 represent the emphabsolute distance
of taper in the x and y direction, respectively. Similar to what we did with
skew, we need to ground the distance to the size of the quadrilateral. Thus,
taper in the x direction would use the half-width e2 , and taper in the y
f4
e4
direction would use the half-length f3 . Thus, taperx = , and tapery =
f3
f2
From Robinsons paper [8]: the Jacobian matrix for a flat (projected)
quadrilateral is given by:
x

[J] =
x

y
(e2 + e4 ) (f2 + f4 )
=[
]
(e3 + e4 ) (f3 + f4 )
y

The determinant of [J] can written as det[J] = e2 f3 +e2 f4 +(e4 f3 e3 f4 ).


Which, in turn, can be refactored using the shape metrics in the following:
det[J] = f32 (AR)(1 + Tx + (Ty

Skew
Tx ))
AR

As we can see, all four shape parameters are present within the determinant, in addition to the half-length of the basic rectangle. For analysis,
we will be taking the determinant at the third vertex of the quadrilateral
(top-right), as that vertex has the curvilinear coordinate where = = 1.
Due to the presence of the half-length, the size of the quadrilateral will
affect its Jacobian determinant. This, plus the possibility of having negative
tapers which can actually reduce the determinant to below the determinant
of a square, have led me to suggest a new metric, Modified determinant, for
quadrilateral quality, by removing the mentioned metrics from the equation:
M odDet[J] = (AR)(1 + Tx + (Ty

Skew
Tx ))
AR

Again, we will use only the third vertex of the quadrilateral for analysis.
12

Since the aspect ratio, skew, both tapers, and the Jacobian determinant
are all calculable from only the four points of a quad, these calculations can
be implemented in a relatively straight-forward way. Since quad meshes are
made up of quad elements by definition, I can run the CRE method over all
the quads in the mesh and find out the average of all four respective shape
elements and the Jacobian determinant.
Running the mesh generation algorithm over a set of different inputs
would yield many quad meshes derived from an algorithm. By using the
average value of the shape elements from all the quads from all the meshes
generated by a single algorithm, we can compare that average to that of other
algorithms.

3.3

Comparison Method

I have written a program that takes in a quad mesh in Geomview Object File
Format (.OFF), and outputs the average and worst metric of quads in the
mesh using the method discussed previously. Using this program, for each
mesh generation algorithm I examine, I will run and record the metrics over
several selected point sets. This includes 6 sets of points placed randomly in
various quantities (10, 50, 100, 200, 500, 1000) and a set of 303 points that
is a polygonal representation of Lake Superior. The 6 sets of random points,
once generated, is tested on all the algorithms: all the meshing algorithms
are tested on the same data set.
The metrics that I will specifically compare with are the following:
Total quads
Average/worst
Average/worst
Average/worst
Average/worst
Average/worst
Average/worst

AR
Skew value
Taper in the x-axis
Taper in the y-axis
Jacobian determinant
Modified Jacobian determinant

Comparisons and Results

The tables below show the average (Table 1) and worst (Table 2) metrics of
the resultant meshes using the specified algorithm and the specified input.

13

Table 1: Average metrics

Total quads
Avg. AR
Avg. skew value
Avg. taperx
Avg. tapery
Avg. Jacobian det.
Avg. mod. det.

10 pts

50 pts

100 pts

200 pts

500 pts

1000 pts

801
1.182521
0.187232
0.15145
0.173889
119.4661
1.609149

3768
1.207838
0.176465
0.156579
0.16305
31.66326
1.631948

7613
1.191341
0.176154
0.151035
0.163772
16.43845
1.601269

15074
1.200362
0.176752
0.150312
0.159986
8.39404
1.608572

37219
1.202959
0.175277
0.149911
0.158285
3.417593
1.610947

75057
1.197838
0.176673
0.149348
0.161002
1.698783
1.605722

Lake Superior
(303 pts)
24444
1.198281
0.185968
0.155988
0.172206
0.013589
1.629233

Table 1.a Atalay et al. quadtree meshing average metrics

Total quads
Avg. AR
Avg. skew value
Avg. taperx
Avg. tapery
Avg. Jacobian det.
Avg. mod. det.

10 pts

50 pts

100 pts

200 pts

500 pts

1000 pts

39
5.587552
2.81124
0.25
0.361675
338.2453
8.636394

273
4.619727
2.498871
0.25
0.320704
70.20836
7.061956

555
4.916503
2.540989
0.25
0.329417
51.46118
7.643117

1152
3.703014
1.768101
0.25
0.321119
20.99207
5.718958

2949
4.028027
1.965916
0.25
0.312787
50.31810
6.274875

5934
3.658026
1.790858
0.25
0.309766
4.078406
5.601143

Lake Superior
(303 pts)
1758
5.456112
2.884117
0.25
0.358115
0.031532
8.545263

Table 1.b Shewchuk Triangle tri-to-quad meshing average metrics

Total quads
Avg. AR
Avg. skew value
Avg. taperx
Avg. tapery
Avg. Jacobian det.
Avg. mod. det.

10 pts

50 pts

100 pts

200 pts

500 pts

1000 pts

126
1.649084
0.510837
0.25
0.289354
35.49717
2.493724

489
1.647021
0.513869
0.25
0.284134
28.91249
2.475079

1032
1.639505
0.505065
0.25
0.281548
14.93490
2.456549

1935
1.647352
0.508918
0.25
0.282497
8.27568
2.470872

4830
1.643218
0.500012
0.25
0.282193
3.483061
2.464108

9921
1.660241
0.516299
0.25
0.281665
1.69971
2.488185

Lake Superior
(303 pts)
3339
1.678473
0.526546
0.25
0.281482
0.005471
2.514221

Table 1.c Shewchuk Triangle with minimum 20 angles, tri-to-quad meshing average metrics

14

Table 2: Worst metrics

Worst
Worst
Worst
Worst
Worst
Worst

AR
skew value
taperx
tapery
Jacobian det.
mod. det.

10 pts

50 pts

100 pts

200 pts

500 pts

1000 pts

3.000000
1.145833
0.6
0.730798
4783.264
4.89

3.000004
1.333333
0.6
0.730798
2990.188
4.890005

3.000003
1.333333
0.6
0.730798
3140.657
4.890003

3.000007
1.333334
0.6
0.730799
3153.046
4.890004

3.000027
1.875
0.600002
0.7308
3157.515
4.89001

3.000055
1.875002
0.600002
0.730805
3160.102
4.890047

Lake Superior
(303 pts)
3.00073
1.875004
0.600029
0.730854
17.503433
4.890585

Table 2.a Atalay et al. quadtree meshing worst metrics

Worst.
Worst.
Worst.
Worst.
Worst.
Worst.

AR
skew value
taperx
tapery
Jacobian det.
mod. det.

10 pts

50 pts

100 pts

200 pts

500 pts

1000 pts

42.79535
28.77125
0.25
1.823832
2448.629
64.19302

281.9245
127.9088
0.25
1.044526
3334.426
422.8869

466.2196
294.6275
0.25
7.098919
4191.789
699.3325

559.2251
329.0671
0.25
4.005084
1702.153
838.8374

1777.939
1254.819
0.25
9.556236
107129.7
2666.950

1174.521
624.1278
0.25
3.096078
964.6050
1761.781

Lake Superior
(303 pts)
530.1884
351.1376
0.25
3.855682
8.472251
795.2556

Table 2.b Shewchuk Triangle tri-to-quad meshing worst metrics

Worst.
Worst.
Worst.
Worst.
Worst.
Worst.

AR
skew value
taperx
tapery
Jacobian det.
mod. det.

10 pts

50 pts

100 pts

200 pts

500 pts

1000 pts

3.34753
1.656701
0.25
0.018427
276.9116
5.021296

4.425987
1.820182
0.25
0.066429
661.7325
6.63898

4.043448
1.757625
0.25
0.072425
200.6794
6.065173

4.180376
1.744029
0.25
0.079237
136.0718
6.270564

4.376373
1.812709
0.25
0.077191
47.29270
6.56456

4.558736
1.899074
0.25
0.086745
32.69729
6.838106

Lake Superior
(303 pts)
4.615184
1.925705
0.25
0.087063
0.273995
6.922818

Table 2.c Shewchuk Triangle with minimum 20 angles, tri-to-quad meshing worst
metrics

15

Figure 6: Jacobian determinant against size of input set


An initial glance yields some interesting observations. Despite the glaring
disadvantage of the tri-to-quad algorithm, the quad meshes created by it
have much fewer faces than those created by Atalay et al.s direct quad mesh
generation algorithm. While running Triangle with a bounded minimum
angle gives us quad meshes with more faces, they still have significantly
fewer faces than those created by Atalay et al.s algorithm.
The change in average metric over quantity of points does not seem to be
significant; that is, more random points in a set does not have a significant
impact on the average metrics, except for the Jacobian determinant. However, knowing the way the Jacobian determinant is calculated (the Jacobian
determinant has the coefficient of aspect ratio times the square of the midlength), this can be attributed to the smaller elements in quad meshes with
more faces. This is supported by the fact that the Lake Superior input data
yields a much smaller Jacobian determinant, no matter the algorithm. See
Figure 6.
16

Figure 7: Modified Jacobian determinant against size of input set


I initially thought there would be a significant difference between randomly generated point set input and a polygonal point set, but plotting
modified Jacobian determinant against input size (Figure ??) yields almost
no difference between randomly generated and polygonal point set input, at
least for Shewchuks Triangle algorithm with angle guarantee and Atalay et
al.s quadtree algorithm.
Comparing the average aspect ratio of non-bounded and bounded minimum angles on Shewchuks Triangle shows a dramatic improvement from
the 3.5 to 5.5 range to a more manageable and consistent 1.6 to 1.7 range
(Figure 8). Meanwhile, Atalay et al.s quadtree algorithm generates aspect
ratio in the range between 1.18 to 1.2.
Curiously, the average taper in the x direction is consistently 0.25, bounded
minimum angle or not. Perhaps this is an artifact from the tri-to-quad algorithm. After examining the individual analysis, all the quads seem to have a

17

Figure 8: Aspect ratio against size of input set


taper of -0.25 under the algorithm.
Looking at the worst metrics, taper in the x-direction exhibits a very
strange characteristic. While we know that the Shewchuk tri-to-quad mesh
contain only quads of taper -0.25, and so the identical worst taperx is no
surprise, taper for Atalay et al.s algorithm generate meshes with extremely
similar tapers. I can not offer any thoughts as to why this would be the case,
given that six of the input point sets are essentially randomly generated, and
so there would be no guarantee that a quad with a given taper would exist.
This may be worth further investigation.
It is interesting to see that the worst modified Jacobian determinant for
Atalay et al.s quad meshes are also very similar to each other. This can perhaps to attributed to the angle bounds that their algorithm guarantees. The
same effect is not seen in Shewchuks Triangle with bounded angles, possibly
because while the original triangle mesh has angle guarantees, the tri-to-quad
conversion may introduce extreme angles. This can be seen in Shewchuks
18

Triangle without bounded angles: the worst modified determinant increases


dramatically for bigger input sets.

Figure 9: Modified Jacobian determinant against aspect ratio

Evaluation

The metrics do provide the expected results: Shewchuks Triangle without


bounded angles gives the worst meshes, followed by Shewchuks Triangle
with bounded angles, and followed by the quadtree direct meshing algorithm.
While the modified determinant is indicative of four shape parameters (aspect
ratio, skew, taperx and tapery ), Robinsons method of calculating aspect ratio
seems to be sufficient for general mesh quality assessment. As we can see
in Figure 9, modified Jacobian determinant correlates extremely highly with
aspect ratio. This would indicate that for the input set we have analyzed,
modified Jacobian determinant is as good as a mesh quality measure as
Robinsons aspect ratio.
19

For individual quad analysis, the modified Jacobian determinant will be


more effective in determining its quality. The determinant is affected by all
four shape parameters, and so is the ideal candidate should algorithms need
to maximize for a certain quality metric.
However, there are limitations to the Jacobian determinant. As Robinson
states, a quadrilateral has two sets of shape parameters depending on which
of the oblique axes is taken as the local x-axis, although the skew parameter is
the same in each case [8]. This means that the same quad, given a different
ordering of points, could give one of two metrics. See Figure 10.

Conclusion

This paper examines Robinsons CRE method in-depth, and provided an


implementation of the analysis to be run on quad meshes in .OFF format.
The modified Jacobian determinant, given in this paper, is a valid measure
of the quality of a quad mesh, although it should be noted that for mesh
analysis, Robinsons method of measuring aspect ratio works similarly as
an indicator of quality. Further research could be conducted by analyzing
other quad mesh generation programs. Since the source code for the analysis
program is given below, the only impediment to analyzing other quad mesh
algorithms would be the acquisition of executable implementation of the
algorithms, and the necessary file type conversion, as .OFF is one of many
mesh formats available.
The code for the analysis program and input data sets is available at
https://github.com/plu97/analyzeOffFile.

References
[1] Atalay F. B., Ramaswami S. and Xu D. (2012) Quadrilateral Meshes
with Provable Angle Bounds, Engineering with Computers 28, Issue 1,
pp 31-56
[2] Bern M., Eppstein D., Gilbert J. (1994) Provably Good Mesh Generation
Journal of Computer and System Sciences 48, Issue 1, pp 384-409

20

[3] Bommes D., Levy B., Pietroni N., Puppo E., Silva C., Tarini M., Zorin
D. (2013) Quad-mesh Generation and Processing: A Survey, Computer
Graphics Forum 32, Issue 6, pp 51-76
[4] Devadoss S. L. and ORourke J. (2011) Discrete and Computational Geometry
[5] Eppstein D. (2001) Global Optimization of Mesh Quality, from http:
//www.ics.uci.edu/~eppstein/pubs/Epp-IMR-01.pdf [Accessed Dec.
19 2014]
[6] Eppstein D., Goodrich M. T., Sun J. Z. (2005) The Skip Quadtree:
A Simple Dynamic Data Structure for Multidimensional Data, from
http://www.ics.uci.edu/~eppstein/pubs/EppGooSun-SoCG-05.pdf
[Accessed Apr. 19 2015]
[7] Remacle, J.-F., Lambrechts, J., Seny, B., Marchandise, E., Johnen, A.
and Geuzainet, C. (2012) Blossom-Quad: A non-uniform quadrilateral
mesh generator using a minimum-cost perfect-matching algorithm. International Journal for Numerical Methods in Engineering 89, Issue 9,
pp 1102-1119
[8] Robinson J. (1987) CRE Method of Element Testing and the Jacobian
Shape Parameters, Engineering Computations 4, Issue 2, pp 113-118
[9] Robinson J. (1988) Understanding Finite Element Stress Analysis
[10] Shewchuk J. R. (1996) Triangle: Engineering a 2D Quality Mesh Generator and Delaunay Triangulator, Applied Computational Geometry Towards Geometric Engineering, pp 203-222
[11] Tarini M., Pietroni N., Cignoni P., Panozzo D. and Puppo E. (2010)
Practical Quad Mesh Simplification, Computer Graphics Forum 29, Number 2, pp 407-418

21

Figure 10: Jacobian Determinant measurement

(a) A quad, with aspect ratio = 1.2, mod. Jacobian determinant = 1.6733

(b) The same quad, rotated 90 to the right. Aspect ratio = 1.5, mod. Jacobian
determinant = 2.0

22

You might also like