﻿Surface Mesh Smoothing, Regularization and Feature Detection
H. Huang ∗ U. Ascher †
February 21, 2008
Abstract
We describe a hybrid algorithm that is designed to reconstruct a piecewise smooth
surface mesh from noisy input. While denoising, our method simultaneously regularizes
triangle meshes on flat regions for further mesh processing and preserves crease sharpness
for faithful reconstruction. A clustering technique, which combines K-means and geometric
a priori information, is first developed and refined. It is then used to implement vertex
classification so that we can not only apply different smoothing operators on different vertex
groups for different purposes, but also succeed in crease detection, where the tangent plane
of the surface is discontinuous, without any significant cost increase. Consequently we are
capable of efficiently obtaining different mesh segmentations, depending on user input and
thus suitable for various applications.
Key words: Multiscale Anisotropic Laplacian, Umbrella Operator, Mesh Regularization,
Crease Detection, Mesh Segmentation, K-means Clustering.
1 Introduction
(a) (b) (c) (d)
Figure 1: Mesh smoothing: (a) noisy Double-Torus model; (b) reconstructed model; (c) crease
detection; (d) data classification.
∗
Institute of Applied Mathematics, University of British Columbia, Vancouver, BC, V6T 1Z2, Canada
(hhzhiyan@math.ubc.ca). Supported in part under NSERC Research Grant 84306.
†
Department of Computer Science, University of British Columbia, Vancouver, BC, V6T 1Z4, Canada.
(ascher@cs.ubc.ca). Supported in part under NSERC Research Grant 84306.
12
The rapid development of 3D scanning and acquisition technology has necessitated efficient
denoising algorithms for triangular surface meshes. Furthermore, simulations in time involving
such surfaces often require subsequent smoothing and regularization. Fast and simple 3D
mesh smoothing operators based upon geometric isotropic diffusion were proposed in the 1990s
[38, 16]. Later, algorithms based on anisotropic diffusion were introduced to avoid smearing
out of real important features [11, 8, 3, 37, 4, 17, 9]. These methods are typically expensive,
both in terms of cost per iteration and in the number of iterations required to achieve
satisfactory results. Moreover, they often require the user to provide unintuitive parameter
values, including a threshold value and a time step for the anisotropic diffusion process. The
approach of bilateral filtering has given rise to methods that more rapidly yield results of a
quality similar to anisotropic diffusion, albeit with less theoretical justification [13, 21, 44].
The latter methods usually require very few, cheap iterations, and cost but a tiny fraction
of the computational effort required to carry out an elaborate anisotropic diffusion process,
especially on large meshes. However, the results generated by these bilateral filtering variants
may strongly depend on vertex neighborhood choices and the manner in which tangent planes
are approximated.
Further, sampling irregularities in the given mesh occasionally distort results and significantly
slow algorithms down [14, 25, 10, 18]. Several discrete operators were designed to
overcome this numerical difficulty and maintain sampling irregularity in the smoothed meshes.
To satisfy different application requirements, a smoothing method with simultaneous mesh
regularization is proposed in [31].
The methods mentioned above can be thought of as having just one stage or one pass.
Another approach for adaptive mesh smoothing is to recognize mesh features first. In [34],
a method is proposed that consists of three stages: feature-preserving pre-smoothing, feature
and non-feature region partitioning, and feature and non-feature region smoothing using two
separate approaches. In [36], the authors first use the eigen analysis of a normal voting tensor
to extract sharp edges, then apply bilateral filtering to smooth along the direction of the sharp
edge, and finally, employ modified bilateral filtering to the overall mesh to obtain a smoothed
mesh model with sharp features. The algorithm proposed here has a similar structure in that
edges and corners are recognized at an earlier stage. However, our methods for detecting features
and for denoising meshes are totally different. Our methods can be simply implemented,
are rather efficient computationally, and can be easily extended to handle a wide variety of
applications as described below.
Polygonal surface models with features that are challenging to preserve can be roughly
divided into two classes. In the first class objects usually contain many visually meaningful
fine scale components or details, and typically require very fine meshes to represent them well.
Examples are the Dragon and Igea models used in [18]; see also Figs. 13 and 12. The second
class consists of piecewise smooth CAD-like models, which usually have large flat regions, long
sharp edges and distinct corners. The required mesh for representation can often be much
coarser. Examples are the Star and Fandisk models, here depicted in Figs. 4 and 6. All of
these models may be sampled very irregularly, as in Fig. 11. In [18], we have methodically
developed a fast adaptive multiscale mesh denoising algorithm that is capable of retaining
fine scale texture as well as mesh irregularities. This algorithm performs particularly well for
models of the first class. In the present article we focus on denoising, as in Fig. 1, and on mesh
regularizing, as in Fig. 2, for the second class of models. Thus, the goal becomes to generate
enough smoothing on the flat regions while preserving edge sharpness and corner distinction.
In addition, we extend our algorithm to efficiently detect all sharp creases and consequently
complete various mesh segmentation options based on user desire, see Figs. 7 and 8.3
(a) (b) (c)
Figure 2: Mesh regularization: (a) original clean but nonuniform Torus mesh; (b) corrupted
by heavy noise, this is our input data; (c) smoothed and regularized by our hybrid algorithm.
Let us refer to vertices in flat regions as non-feature vertices. Both edge and corner vertices
are called feature vertices. Since different vertex groups should be dealt with in different ways,
it is advantageous to cluster vertices accurately and then choose the most suitable discrete
smoothing operator on each vertex cluster. Data classification on polygonal meshes is not
new as such. In [26], Lavou et al designed a constant curvature region decomposition of
3D meshes following vertex classification. Fuzzy clustering was applied on triangular mesh
faces for hierarchical mesh decomposition in [24]. Liu and Zhang [27] proposed a 3D mesh
segmentation algorithm through spectral clustering of mesh faces. In [7], Chen et al used
Bayesian discriminant analysis to determine the decision boundary for separating potential
feature and non-feature vertices in curvature space. However, we believe that the present
paper is the first to develop a specific clustering technique using the first order height intensity
and properly scaled Gaussian curvature, which is good enough to be applied directly on noisy
meshes. All related parameters are highly intuitive in our geometric context.
In the following Section 2 we first recall several mesh denoising algorithms. After analyzing
the strengths and weaknesses of the corresponding discrete operators we develop our clustering
technique in Section 3, and subsequently choose the most suitable operator for smoothing on
each vertex cluster with a specific feature. Further, based on the constructed vertex classification
we extend our algorithm in Section 4 to detect all sharp creases, and segment meshes
with respect to mesh sharp features and user requirements. Implementation details, results
and discussion are then presented in Section 5. Summary and conclusions follow.
2 Discrete Laplacian Operators
Before describing different discrete Laplacian operators, we introduce notation. A manifold
M is discretized by a triangular surface mesh S with its sets of vertices V (S) = x = {xi; i =
1, . . . , N} and directed edges E(S). If two distinct vertices xi and xk are linked by an edge
ei,k = xk − xi then we denote k ∈ N (i) and the edge length li,k = |ei,k|. F(i) represents the
one-ring-face set that is adjacent to the vertex xi. The given data is a noisy mesh of this sort,4
and we denote its vertices x(0) = v = {vi; i = 1, . . . , N}.
2.1 Discrete isotropic Laplacian
The simplest discrete isotropic Laplacian operator is the Umbrella operator [38]. It averages
the neighboring edges
∆xi = 1
∑
ei,k, (1)
mi
k∈N (i)
where mi = |N (i)|, the number of neighbors of vertex xi. This is a linear form implying the
assumption that all neighboring edge lengths are roughly equal to one. Hence it can serve
as an effective smoother if the targeted mesh is close enough to being regular. On the other
hand, when the model has different discretization rates significant local mesh regularization,
which may or may not be desirable, is introduced by this Umbrella operator (see [31]). To
retain mesh sampling rates, a better choice is the scale-dependent version [14]. Further, to
solve problems arising from unequal face angles, a better approximation to the mean curvature
normal was proposed in [10, 28] which doesn’t produce vertex tangential drifting when surfaces
are relatively flat and compensates both for unequal edge lengths and for unequal face angles.
This approach does not improve the mesh irregularity sampling rate [31]. All these schemes
are based on isotropic diffusion, though, which implies that they all easily smear sharp features
during the smoothing process.
2.2 Discrete anisotropic Laplacian
To better reconstruct sharp features, consider next designing an anisotropic Laplacian operator
[18]. For each vertex xi of the given data mesh we estimate the corresponding normal ni as
the mean of the adjacent face normals and define the local height intensity hi,k = e T i,k ni, the
projection of ei,k along the normal ni. A corresponding discrete anisotropic Laplacian (AL)
operator is then given by
∆xi = 1
Ci
⎛
⎝ ∑
k∈N (i)
g(hi,k)hi,k
⎞
⎠ ni, (2)
where Ci = ∑
k∈N (i) g(hi,k).
Further, we employ a Gaussian filter in this context, which is simple, robust and clearly
reduces the influence of neighbors that contain large discontinuities in the normal space
g(hi,k) = exp(− h2 i,k
2σ2 ), (3)
i
where hi = {hi,k} k∈N (i). The robust scale σi may be automatically estimated by either the
mean absolute deviation scaling
or the median absolute deviation scaling
σi = 2 ∗ mean(|hi − mean(hi)|), (4a)
σi = median(|hi − median(hi)|). (4b)5
2.3 Multiscale evolution
The normalization in (2) scales the operator such that, from the point of view of geometric
diffusion, a step size τ = 1 can be stably used in an explicit integration method such as forward
Euler
xi ← xi + ∆xi, i = 1, . . . , N. (5)
Though AL is very efficient for smoothing some models, it cannot avoid over-smoothing
in the presence of significant intrinsic texture, see [18]. A simple first step toward a better
solution is then to add a data fidelity term
xi ← xi + τ∆xi + λi(vi − xi), i = 1, . . . , N, λ > 0, (6)
where {vi; i = 1, . . . , N} is the given vertex set, thus increasing the influence of the given data
during the denoising process at all iterates [39]. This also helps to reduce the effect of volume
shrinkage over several iterations [38, 10, 18].
Since the surface scale is local and the mesh is generally nonuniform, we choose λi at each
such iteration depending on the spatial location i. Also, since noise is carried in the position
of vertices and changes all the neighboring height intensities hi,k, we expect that λi should
depend on this noise effect. Recall that the Gaussian parameter σi, adaptively determined by
(4), is a robust estimator on local height intensity, larger over feature regions and smaller over
flat regions. This naturally gives rise to the choice
λi = σi/¯σ, where ¯σ = max{σi; i = 1, . . . , N}. (7)
For details, see [18]. The importance of the entire function λ(x) obtained this way is magnified
in the next iteration step by damping out the AL operator (2), to recapture a higher frequency
band. Thus, in the jth iteration we calculate λ by (7) and set
xi ← xi + K j ∆xi + λi(vi − xi), i = 1, . . . , N, (8)
where K is an input parameter, 0 < K < 1, with the default setting K = 0.5. We refer to this
denoising scheme as the multiscale anisotropic Laplacian (MSAL) method.
In [18] we have shown that MSAL with the mean absolute deviation scaling (4a) performs
rather well for recapturing fine scale texture. However, MSAL with (4a) sometime tends to
over-smooth long sharp creases, for instance in the Fandisk model of Fig. 6. We must use
another robust scale σ in the Gaussian filter, suitable for the current purpose. The image
processing literature uses tools from robust statistics to sharpen image edges and corners by
automatically estimating σ as the median absolute deviation of the given image intensity gradient,
see [5, 33]. Extending this selection to polygonal meshes, our numerical experiments
show that for preserving edge sharpness (4b) works much better than (4a). Similar conclusions
are reached in [42], where mesh smoothing algorithms via mean and median filters applied to
face normals are compared. Unfortunately, median absolute deviation scaling cannot generate
enough smoothing on large flat regions and occasionally destroys corners due to its inherent
limitations [15, 6]. This suggests that different methods may best be applied for different purposes.
Vertex classification is thus a good way to go further. For example, if we classify vertices
on a CAD-like model into three non-feature, edge and corner groups, we may apply different
smoothing operators with different parameter selections to pursue a better reconstruction.6
3 Vertex Classification
3.1 Data set computation for clustering
One way to classify mesh vertices is according to their principal curvatures. The magnitude
of both principal curvatures is small for non-feature vertices and large for corner vertices. For
an edge vertex, the magnitude of one of its principal curvatures is small and the other is quite
large. Thus, if we have curvature information of the mesh, we can partition mesh vertices into
three clusters denoted corner, edge and non-feature.
In [28] the discrete mean curvature normal at vertex xi is defined by
K(xi) = 3
2Ai
∑
k∈N (i)
(cot αi,k + cot βi,k)(xi − xk)/2, (9)
where αi,k and βi,k are the two angles opposite to the edge ei,k in the two triangles sharing it.
Here we use one third of the simple one ring face area Ai at the vertex xi to approximate the
complicated surface area Amixed defined in [28]. See [41, 29] for approximation convergence
estimates. This yields the curvature expression κM(xi) = 1
2 ||K(xi)||. The discrete Gauss
curvature κG at vertex xi is approximated as
κG(xi) = 3(2π − ∑
k∈F(i)
θk)/Ai, (10)
where the sum is over the faces in the set F(i) and θk is the angle of the k-th face at the vertex
xi. Since κM = (κ1 + κ2)/2 and κG = κ1κ2, discrete principal curvatures at the vertex xi can
be consequently computed by the quadratic formula
κ1(xi) = κM(xi) + √ ∆(xi), κ2(xi) = κM(xi) − √ ∆(xi), (11)
with ∆(xi) = max(κ 2 M (xi) − κG(xi), 0). Note that in the continuous case and almost always
in the discrete case, κ 2 M (xi) > κG(xi). The maximum principal curvature κ1 is alway positive
whereas the minimum principal κ2 follows the sign of the Gaussian curvature κG which is
negative at hyperbolic vertices. Since it is not necessary to differentiate ellipticity and hyperbolicity
in our classification, we just consider the absolute value of both Gaussian curvature
κG and minimum principal curvature κ2. The quantitative 2D data matrix C ∈ IR N×2 for
classification is then generated as
C = [κ1(x1), κ1(x2), . . . , κ1(xN); |κ2(x1)|, |κ2(x2)|, . . . , |κ2(xN)|] T . (12)
However, since we are directly working on noisy vertices, the second-order curvature information
is more sensitive and more easily influenced by noise effects, so that we often cannot
obtain the desirable classification, see Figs. 4 and 9. Moreover, computation of principal curvatures
of each vertex is obviously not cheap. Fortunately, we already have first-order information
to roughly partition noisy meshes. Recall the local height intensity hi,k = e T i,k ni, which is the
projection of ei,k along the vertex normal ni. Now define
hmax(xi) = max
k∈N (i) |hi,k|,
hmin(xi) = min
k∈N (i) |hi,k|. (13)
Similarly to principal curvatures, for an edge vertex, hmin is small whereas hmax should be
relatively quite larger. For a corner vertex, both hmin and hmax should be large, and for7
a non-feature vertex they should both be relatively small. Thus, for a given vertex matrix
V = [x1, x2, . . . , xN] T , xi ∈ IR 3×1 , i = 1, . . . , N, we can generate a corresponding quantitative
data matrix H ∈ IR N×2 with respect to height intensity
H = [hmax(x1), hmax(x2), . . . , hmax(xN); hmin(x1), hmin(x2), . . . , hmin(xN)] T . (14)
In pattern recognition terminology [12], the rows of C or H are called the patterns or
objects and the columns are called the features or attributes. The objective of clustering is to
partition the data set into several clusters. Generally, a cluster is a group of objects that are
more similar to one another than to members of other clusters. The term similarity should
be understood as mathematical similarity, measured in some well-defined sense. In our case,
considering the first column of data matrices as X-coordinate and the second column as Y-
coordinate, the similarity can be measured simply by the Euclidean distance based on the
geometric information we already have. The whole data set sits in the first quadrant. Objects
that correspond to non-feature vertices should be near the origin; objects that correspond to
edge vertices should be close to the X-axis and relatively far from the Y -axis; and objects that
correspond to corner vertices should be close to neither X-axis nor Y -axis.
However, a height intensity value depends on the approximation of the vertex normal.
For some corner vertices, the value of their hmin could be very small where the average of
the neighboring face normals might point along one edge such that they would be mistakenly
grouped into the edge cluster, see Figs. 4, 5 and 9. To address this, we employ the Gaussian
curvature κG. There are two reasons. The first is that |κG| is larger at corner vertices and
much smaller at both non-feature and edge vertices. The second reason is that computing κG
by (10) is inexpensive. We can get θ and A very quickly by applying simple operations to the
cross product of two edge vectors in each triangle. These cross products must be performed
anyway when computing face normals. In order to generate the data matrix similar to C or
H with respect to the Gaussian curvature, we set
κG1(xi) = 6π
Ai
+ √ Θ(xi), κG2(xi) = 6π
Ai
− √ Θ(xi), (15)
with Θ(xi) = ( 6π
Ai )2 − κG(xi). The corresponding quantitative data matrix G ∈ IRN×2 for
corner identification is then generated as
G = [κG1(x1), κG1(x2), . . . , κG1(xN); |κG2(x1)|, |κG2(x2)|, . . . , |κG2(xN)|] T . (16)
Clearly, |κG2| is larger at the corner vertices than at the rest. Therefore, objects in G
corresponding to corner vertices should be relatively far above the X-axis, whereas objects
corresponding to non-feature and edge vertices should be close to the X-axis. For a better
clustering we linearly rescale the first column of G (X-coordinate in the Euclidean axis) such
that all objects are bounded by a square box.
3.2 K-means clustering
The classic unsupervised fast clustering algorithm is K-means [12], which allocates each data
point to one of c clusters to minimize the within-cluster sum of squares. Taking the partition
data matrix H as an example, we seek a partition H1, . . . , Hc to minimize the objective function
c∑
∑
|| ˆ hi − µj||2, (17)
j=1 ˆhi∈Hj8
(a) (b) (c) (d)
Figure 3: (a) Corrupted Ring model; (b) smoothed model based on partition in (d): nonfeature
vertices are marked by smaller pink dots and feature (edge) vertices are marked by
larger pink dots; (c) data set H computed for the noisy Ring model in (a); (d) two clusters by
K-means: red → non-feature and green → feature.
where ˆ hi = [hmax(xi), hmin(xi)], Hj is a set of objects (data points) in the j-th cluster and
µj = mean( ∑ ˆhi)
hi∈Hj
ˆ is the the center point over the j-th cluster. The number of clusters c
is usually two in the current context. For example, corner and non-corner clusters are always
expected in the data matrix G, see Figs. 4(i) and 5(c). After the corner cluster is identified
and temporarily removed from H, nonempty edge and non-feature clusters are expected in the
objects that remain, see Figs. 4(p) and 5(f). Or, as in the Ring model in Fig. 3, there is no
corner vertex at all, so we only need to classify vertices into the two clusters of feature and
non-feature.
One important issue here is the need to supply a good initial guess for the center points
µj for starting the K-means clustering. According to our analysis of the distribution of data
matrices G, H, or C above, we intuitively set
µcorner = [κG1(xj), |κG2(xj)|], j = arg max
1≤i≤N |κG2(xi)|,
µnoncorner = [κG1(xk), |κG2(xk)|], k = arg min
1≤i≤N |κG2(xi)|,
for corner identification in the data matrix G, and then
µedge = [ max
1≤i≤N hmax(xi),
min
1≤i≤N hmin(xi)],
µnonedge = [ min
1≤i≤N hmax(xi),
min
1≤i≤N hmin(xi)],
for edge detection in the data matrix H. The same initial setting as above has been applied
in C for comparison purposes.
K-means clustering is simple and efficient, but it is based on estimating explicit models of
the data. When the data distribution is arranged in a complex pattern or corrupted by heavy
noise, K-means may not give us a very satisfactory vertex partition. One more advanced
clustering approach, which has been shown to handle more complicated structured data, is
spectral clustering [1, 22, 32, 40, 35, 30]. It does not require estimating an explicit model
of data distribution, but rather employs a spectral analysis of the matrix of point-to-point
similarities. However, the real power of spectral clustering is not utilized in the present context.
Moreover, in our mesh smoothing application, especially for large meshes, spectral clustering
may significantly slow the algorithm down, because it involves the calculation of the c leading9
(a) (b) (c) (d) (e)
(f) (g) (h)
(i) (j) (k)
(l) (m) (n)
(o) (p) (q)
Figure 4: (a) Corrupted Star model; (b) smoothed model based on partition in (n); (c)
smoothed model based on partition in (q); (d) 3D plot of feature vertices classified by principal
curvatures (blue and green clusters in (n)); (e) 3D plot of feature vertices classified by height
intensities (blue and green clusters in (q)); (f) data set H computed for the noisy Star model
in (a); (g) data set C; (h) data set G; (i) two clusters by K-means on data set G for corner
identification : blue → corner; (j) marked corner vertices classified by Gaussian curvature in
data set H; (k) marked corner vertices in data set C; (l) temporarily removing the corner
cluster from data set C; (m) two clusters by K-means on data that has remained in (l) : red10
eigenpairs for the large and full affinity matrix. To maintain high efficiency in our algorithm,
we propose another method to improve it in a cheaper way, namely hierarchical K-means.
Thus, we first implement K-means clustering repeatedly on the unresolved cluster until all
sub-clusters consist of vertices with only one type of feature - corner, edge or non-feature.
Then we simply combine sub-clusters that contain vertices with the same feature. When this
process terminates we have at most three clusters, corner, edge and non-feature, ready for
smoothing and further processing. See the column marked Km-N in Table 1 for the number
of clustering applications required for different models.
Our numerical results show that, for denoising non-feature vertices, both the umbrella
operator and AL with (4a) perform well. The main difference between them is that the
umbrella operator regularizes irregular meshes during smoothing, while AL with (4a) keeps the
mesh irregularity sampling rate relatively unchanged. Since the umbrella operator is simpler,
faster, and produces simultaneous mesh regularization, we apply it by default on the nonfeature
cluster. Further, we apply MSAL with (4b) on the edge cluster, and keep the corner
cluster untouched, see Figs. 3(b) and 4(c). This approach is referred to as our hybrid denoising
algorithm.
The Star example in Fig. 4 demonstrates a case where using the height intensity data
matrix H instead of the principal curvature data matrix C is preferable. Even though in
Figs. 4(l) and 4(m) edge and non-feature groups with respect to principal curvatures look
well separated, Fig. 4(d) clearly demonstrates that vertices on 12 edges that form a cube are
mistakenly grouped into the non-feature cluster with the result that those 12 edges are oversmoothed
in the reconstructed model, see Fig. 4(b). Even if we continue to do reclustering
on the red non-feature cluster in Fig. 4(m), the result cannot be improved since principal
curvature data points corresponding to vertices on the cube are highly mixed with those of
vertices on flat regions. In comparison, height intensity data points in Fig. 4(o) provide a
better distribution pattern such that only one application of K-means clustering generates a
very good partition between edge and non-feature, see Figs. 4(p) and 4(e). Combining with
the corner cluster identified in Fig. 4(i), the vertex classification depicted in Fig. 4(q) yields
the high quality reconstruction in Fig. 4(c).
3.3 Classification Refinement
Since we are directly working with noisy data, we may not get exact vertex classification
even when the more advanced clustering techniques are employed. The approximation of
some vertex normals could be badly influenced by heavy noise and likewise for local height
intensities. 1
Some non-feature vertices with high frequency noise might be wrongly grouped
into the edge or corner clusters; some corner vertices, especially at saddle structures, might be
wrongly grouped into the edge or non-feature clusters; some edge vertices on curved creases
might be wrongly grouped into the non-feature cluster; and so on. These could result in
undesirable bumps in flat regions, defects at edges and corners, or the vanishing of visually
meaningful curved ridges, as e.g. in Fig. 6(b). Hence, a post-clustering procedure is proposed
to refine the vertex classification.
Take the Fandisk model as an example. Firstly, partition the scaled Gaussian curvature
data G into two groups to identify corner vertices; secondly, locate and temporarily remove
corners from the data H, see Figs. 5(c) and 5(d); thirdly, employ K-means clustering on the
data remaining in H to get the non-feature and edge clusters, see Figs. 5(e) and 5(f); finally,
1
Other methods for estimating the vertex normals [19, 20, 18] have not produced significant improvement
in this regard.11
(a) (b) (c) (d) (e) (f)
(g) (h) (i) (j) (k)
Figure 5: (a) data set H computed for the noisy Fandisk model depicted in Fig. 6(a); (b)
data set G; (c) corner identification in (b); (d) marking corner vertices classified by Gaussian
curvature in (a); (e) temporarily removing the corner cluster from (d); (f) two clusters by
K-means on the data that has remained in (e) : red → non-feature and green → edge; (g)
adding the corner cluster back into (f); (h) recapturing potential edge vertices back into the
edge cluster through the classification refinement procedure (P = 25); (i) 3D plot of edge and
corner vertices based on clusters in (g), see corresponding model in Fig. 6(b); (j) 3D plot of
edge and corner vertices based on refined clusters in (h) where red dots mark the recaptured
vertices, see corresponding model in Fig. 6(c); (k) crease detection: each crease is traced by
one specific color.
(a) (b) (c)
Figure 6: Recovering curved ridges on the Fandisk model: (a) noisy model; (b) smoothed
model based on the vertex partition in Fig. 5(g); (c) smoothed model based on the refined
vertex partition in Fig. 5(h).12
combine these three clusters and apply the refinement procedure that we design below to
improve clustering quality, see Figs. 5(g), 5(h), 5(i) and 5(j).
Refinement Algorithm:
• A corner vertex should have at least three edge vertices in its one-ring neighbor, referred
to as edge neighbors. Thus, for each unchecked vertex in the corner cluster:
1. send it into the non-feature cluster if it does not have any edge neighbor;
2. send it into the edge cluster if it has fewer than three edge neighbors;
• An edge vertex should have at least two neighboring feature vertices, referred to as feature
neighbors, each belonging to one of the edge cluster, corner cluster or potential edge list.
Thus, for each unchecked vertex in the edge cluster with the empty potential edge list:
1. send it into the non-feature cluster if it does not have any feature neighbor;
2. if it only has one feature neighbor, apply recapturing:
2.1. add the non-feature neighbor with minimal local height intensity to the potential
edge list;
2.2. if the total number of vertices on the list exceeds P , send all these vertices
back into the non-feature cluster and then move on to the next unchecked edge
vertex; otherwise continue;
2.3. begin to check the newest vertex added into the potential edge list: if it has
more than one feature neighbor, recapture all vertices on the list from the nonfeature
cluster into the edge cluster and label them as checked edge vertices
and then move on to the next unchecked edge vertex; otherwise do step 2.1 to
pick another potential feature neighbor of this edge vertex.
The confidence integer P controls the extent of edge searching. It is the only parameter
appearing in our classification refinement algorithm. It should be neither too small nor too
large because we want to find as many potential edge vertices as possible whereas we must
avoid recapturing by mistake non-feature vertices. Fortunately, it is very intuitive to adjust
P up or down according to the 3D plot of feature vertices after running refinement, such as
depicted in Fig. 5(j). In addition, the whole refinement procedure is very efficient (see Table
1), so there is no costly trial and error process. In our experience, for most cases the default
setting P = 25 works very well.
4 Crease Detection and Mesh Segmentation
Since we are able to mark all feature vertices, one quick extension is to detect all sharp creases
on CAD-like models. Taking the Donut model in Fig. 7(a) as one example, we consider its
feature vertices as a complete, directed and weighted graph. If there is an edge in the original
mesh from feature vertex xi to xk, then set the weight wi,k = hi,k; otherwise set wi,k = ∞.
Thus, by implementing the crease detection algorithm described below we are able to locate
and label all visually distinct creases on the Donut, see Fig. 7(b).
Subsequently, if some closed piecewise smooth curves have been determined by ordering
the labels of creases to form cutting paths, then we are able to carry out the corresponding13
(a)
(b)
(c) (d) (e) (f)
Figure 7: (a) Noisy Donut model and vertex classification in the data set H; (b) all creases
detected and labeled; (c) segmenting the smoothed mesh into 3 patches; (d) 5 patches; (e) 10
patches; (f) 50 patches.
(a) (b) (c) (d) (e)
Figure 8: Various mesh segmentations for the Star model in Fig. 4 according to input cutting
paths sketched in pink: (a) 2 patches; (b) 3 patches; (c) 6 patches; (d) 12 patches; (e) 24
patches.14
specific mesh segmentation. This is different from general mesh segmentation problems studied
in various articles [27, 43, 2, 23], where the main challenge is to automatically produce
segmentation results that are in close agreement with human shape perception. In the current
context, the segmentability of a CAD-like shape is quite clear, and all creases that can define
different cutting paths are already located. So what we really are concerned with here is how
to efficiently make use of available data while respecting user specifications as much as we
can. The algorithm is given below. Figs. 7(c) - 7(f) and 8(a) - 8(e) clearly demonstrate the
flexibility of our mesh segmentation. Note that we are making one major assumption, namely,
that input cutting paths must be closed.
Crease Detection Algorithm:
• Input data: corner vertex set Vc and edge vertex set Ve;
• Initialize: compute weight matrix using mesh connectivity and height intensity;
• Starting point: one random corner vertex in Vc;
1. from starting point mark edge vertices along path with the smallest weight until
arriving at any corner vertex or marked edge vertex;
2. label the found crease and remove marked edge vertices from Ve;
3. if the starting corner vertex is not isolated, repeat step 1; otherwise mark it, set
another unmarked corner vertex as the starting point and repeat step 1;
• Check convergence if the whole set Vc is marked: exit successfully if Ve is empty, otherwise
randomly pick a vertex in Ve as a new starting point and repeat step 1. Note that in this
case the found crease may be closed without containing a corner, as in Figs. 3 and 9.
Mesh Segmentation Algorithm:
• Available data: vertex set V and connectivity structure;
• Input data: the list of labels of the creases that form several closed cutting paths;
• Initialize: Vcut = vertices on input creases and Vinterior = V/Vcut;
• Starting point: one random vertex in Vinterior;
1. send multiple-ring vertex neighbor of the starting point into one patch-vertex set
Vp; stop neighbor chasing at vertices in Vcut such that all vertices in the outermost
ring are in Vcut;
⋂
2. paint one-ring face neighbor of all members in Vp Vinterior with the same color and
set Vinterior = Vinterior/Vp;
⋂
3. check if there is any face with all three vertices in Vp Vcut; if yes paint it too;
• Checking convergence: exit successfully if Vinterior is empty; otherwise start a new patch
painting with a different color.Figure 9: (a) Height intensity data matrix H computed for the noisy Bearing model in (m);
(b) principal curvature data matrix C; (c) scaled Gaussian curvature data matrix G; (d) two
clusters in G classified by K-means; (e) reclassifying the corner cluster (blue) in (d) by K-
means; (f) replacing the corner cluster in (d) with new corner cluster generated in (e) and
combining new non-corner cluster (green) with old non-corner cluster in (d); (g) locating
corner vertices identified by Gaussian curvature in (a); (h) locating corner vertices in (b); (i)
temporarily removing the corner cluster from (g); (j) two clusters by K-means for the data
that has remained in (i) : red → non-feature and green → edge; (k) adding the corner cluster
back into (j); (l) classification refinement; (m) noisy Bearing model; (n) smoothed model based
on the vertex partition in (k); (o) smoothed model based on the refined vertex partition (l).
15
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
(m) (n) (o)16
(a) (b) (c) (d)
(e) (f) (g)
Figure 10: Reconstruction with edge sharpness preservation: (a) noisy Sliced-Sphere model;
(b) smoothed model based on the refined vertex classification in (g); (c) another view of (a);
(d) corresponding view of (b); (e) detected creases represented in different colors; (f) data set
H computed for the noisy model in (a); (g) refined vertex classification.
5 Numerical Results and Discussion
The first set of results presented in Figs. 1 and 2 demonstrates how well edge sharpness is
preserved by our reconstructed 3D surface meshes and how efficiently mesh sampling is regularized.
The second set of results in Figs. 3, 4, 5 and 6 shows step by step how the data set
for the best vertex partition is cheaply built, how K-means clustering efficiently works, how
classification quality is optimized and how powerful our corresponding hybrid denoising algorithm
can be. Next, in Figs. 7 and 8 we demonstrate the successful extension of our algorithm
to sharp crease detection and adaptive mesh segmentation.
A more complicated data pattern is considered in Fig. 9. Just one application of K-means
clustering using the data set G is not sufficient to give satisfactory corner identification, see
Fig. 9(d). Instead of using the attractive but time-consuming spectral clustering technique,
we employ hierarchical K-means developed in Sec. 3.2 to solve this problem in a much cheaper
way. After reclustering, combining and refining, we obtain a perfect restoration of the Bearing
model in Fig. 9(o). Figs. 9(a) and 9(b) demonstrate that for the noisy Bearing model, the
height intensity data set H provides a better pattern distribution for vertex classification than
the principal curvature data set C. Figs. 10 and 11 depict more examples where our algorithm17
(a) (b) (c)
(d) (e) (f) (g)
Figure 11: (a) Vertex classification for the noisy Torus model in (d); (b) 3D plot of edge
and corner vertices based on clusters in (a); (c) crease detection; (d) noisy Torus model; (e)
smoothed model based on the vertex partition in (a); (f) mesh segmentation (16 patches) when
all detected creases in (c) are input to construct cutting paths; (g) another view of (f).
Model Vertices Faces Km-N Km-T Re-T Cr-T Sm-T Sm-N T-T
Ring 3(b) 2.3K 4.6K 1 0.04 0.06 0.06 0.38 4 0.54
Torus 11(e) 2.7K 5.4K 2 0.13 0.09 0.06 0.92 4 1.20
S-Sharp 10(b) 4.3K 8.6K 5 0.29 0.45 0.09 1.05 4 1.88
Fandisk 6(c) 6.5K 13K 2 0.18 0.37 0.11 1.28 4 1.94
Idol 12(b) 10K 20K 1 0.28 1.97 - 0.94 3 3.19
Bearing 9(o) 14K 28K 3 0.48 1.41 0.29 1.67 4 3.85
Star 4(c) 28K 56K 3 0.65 1.06 0.37 3.82 5 5.90
D-Torus 1(b) 35K 70K 4 0.83 1.67 0.32 5.60 5 8.42
M-Neko 13(b) 62K 125K 3 3.76 7.94 - 5.41 3 17.1
Table 1: Runtime data for different models. Km-N: number of applications of K-means; Km-T:
total time for K-means clustering; Re-T: time for classification refinement; Cr-T: time for crease
detection; Sm-N: number of smoothing iterations; Sm-T: total time for hybrid smoothing; T-T:
total execution time for the whole process. Reported times are in seconds; all examples are
performed on an Intel Pentium 4 CPU 3.2 GHz machine with 512 RAM.18
(a) (b) (c) (d)
Figure 12: Reconstructed results with large featureless regions and intrinsic texture: (a)
scanned Idol model containing unknown noise; (b) smoothed model by our hybrid algorithm;
(c) back view of (a); (d) back view of (b).
(a) (b) (c) (d)
Figure 13: (a) Scanned Maneki-Neko model with noise; (b) smoothed model with intrinsic
texture by our hybrid algorithm; (c) back view of (a) - featureless region with noise; (d) back
view of (b) - featureless region without noise.19
is employed to yield rather pleasing results in terms of mesh smoothing, segmentation and
crease detection.
Figs. 12 and 13 demonstrate that our method not only preserves sharp model features
but also retains visually meaningful fine scale components referred to as intrinsic texture, even
when the model contains large featureless regions as well. In this case we simply divide vertices
into feature and non-feature groups, and then use the umbrella operator on the non-feature
cluster to obtain enough smoothing while applying MSAL to the feature cluster to capture
intrinsic texture. Note that for both these models MSAL can in fact be directly applied
without any vertex classification, yielding pleasing results that are comparable to those shown
here at somewhat faster total execution times than those listed in Table 1. Our purpose here
is to demonstrate that the hybrid method can successfully and efficiently combine different
aspects such as mesh regularization on flat regions and careful denoising in the presence of fine
scale model features.
In [18] we have demonstrated the power for anti-shrinking of the MSAL scheme as well.
Since here we keep corner vertices unchanged and apply MSAL on edges, the hybrid algorithm
also helps toward restoring the original volumes, achieving even better results in this regard.
Table 1 emphasizes the efficiency of our hybrid algorithm. The number of smoothing iterations
Sm-N need not be large: four or five iterations are sufficient for most cases. Although the
computation time for vertex classification and crease detection depends on the particular model
structure, the total execution time recorded in the last column increases almost linearly with
the size of the mesh.
6 Conclusions
We have designed an efficient hybrid algorithm based on specific vertex classification that is
capable of denoising 3D surface meshes of models with long edges while preserving edge sharpness,
and of generating sufficient smoothing while simultaneously regularizing over featureless
regions. Combined with the multiscale method MSAL developed in [18] for models with intrinsic
texture we now have a set of algorithms that efficiently handle smoothing and regularization
of meshes large and small in a variety of situations. Subsequent crease detection and adaptive
mesh segmentation algorithms are developed. These operations can be carried out without
any significant cost increase.
Whereas our segmentation algorithm is new as far as we know, we emphasize its restricted
nature. The assumption that cutting paths have been detected and are closed is a major one,
and in this sense our algorithm is not comparable to the usual, more general mesh segmentation
approaches. It is merely an inexpensive, direct extension of our system.
Our method employs hierarchic K-means and a confidence integer P in the classification
refinement. This implies that some user intervention in the form of parameter specification
must be required. Since a fully automatic mechanism is always desirable, a future challenge
is to deduce the optimal parameter P from the mesh structure itself and to introduce some
computationally efficient fuzzy clustering techniques that may be more capable of classifying
vertices with respect to their features while requiring little or no user intervention.
Acknowledgments
We wish to thank Drs. Nando de Freitas and Alla Sheffer for fruitful discussions. Models
presented in this paper have been downloaded off the websites of Klaus Hildebrandt, Alexander20
G. Belyaev and AIM@SHAPE Shape Repository.
References
[1] Y. N. Andrew, I. J. Michael, and W. Yair. On spectral clustering: Analysis and an
algorithm. In Proceedings of Neural Information and Processing Systems, 2002.
[2] M. Attenea, B. Falcidieno, and M. Spagnuolo. Hierarchical segmentation based on fitting
primitives. The Visual Computer, 22:181–193, 2006.
[3] C. Bajaj and G. Xu. Adaptive surface fairing by geometric diffusion. In Proceedings
of Symposium on Computer Aided Geometric Design, pages 731–737. IEEE Computer
Society, 2001.
[4] C. Bajaj and G. Xu. Anisotropic diffusion on surfaces and functions on surfaces. ACM
Trans. Graphics (SIGGRAPH), 22(1):4–32, 2003.
[5] M. J. Black, G. Sapiro, D. H. Marimont, and D. Heeger. Robust anisotropic diffusion.
IEEE trans. image processing, 7(3):421–432, 1998.
[6] C. Y. Chen and K. Y. Cheng. A sharpness dependent filter for mesh smoothing. Computer
Aided Geometric Design, 22:376–391, 2005.
[7] C. Y. Chen, K. Y. Cheng, and H. Y. Liao. Fairing of polygon meshes via bayesian
discriminant analysis. J. WSCG, 12:1–3, 2004.
[8] U. Clarenz, U. Diewald, and M. Rumpf. Anisotropic geometric diffusion in surface processing.
In Proceedings of IEEE Visualization, pages 397–405, 2000.
[9] U. Clarenz, U. Diewald, and M. Rumpf. Processing textured surfaces via anisotropic
geometric diffusion. IEEE Transactions on Image Processing, 13(2):248–261, 2004.
[10] M. Desbrun, M. Meyer, P. Schröder, and A. H. Barr. Implicit fairing of irregular meshes
using diffusion and curvature flow. In Proceedings of SIGGRAPH, pages 317–324, 1999.
[11] M. Desbrun, M. Meyer, P. Schröder, and A. H. Barr. Anisotropic feature-preserving
denoising of height fields and bivariate data. Graphics Interface, pages 145–152, 2000.
[12] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. second ed. John Wiley
& Sons, INC, 2001.
[13] S. Fleishman, I. Drori, and D. Cohen-Or. Bilateral mesh denoising. ACM Trans. Graphics
SIGGRAPH, 22(3):950–953, 2003.
[14] K. Fujiwara. Eigenvalues of laplacians on a closed riemannian manifold and its nets. In
Proceedings of AMS, volume 123, pages 2585–2594, 1995.
[15] R. C. Gonzalez and R. E. Woods. Digital image processing. second ed. Prentice-Hall,
Engewood Cliffs, NJ, 2002.
[16] I. Guskov, W. Sweldens, and P. Schröder. Multiresolution signal processing for meshes.
In Proceedings SIGGRAPH, pages 325–334, 1999.22
[34] J. Shen, B. Maxim, and K. Akingbehin. Accurate correction of surface noises of polygonal
meshes. Int’l J. Numerical Methods in Eng., 64(12):1678–1698, 2005.
[35] J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 22(8):888–905, 2000.
[36] T. Shimizu, H. Date, S. Kanai, and T. Kishinami. A new bilateral mesh smoothing
method by recognizing features. In Proceedings of the Ninth International Conference on
Computer Aided Design and Computer Graphics (CAD-CG’05), pages 281–286, 2005.
[37] T. Tasdizen, R. Whitaker, P. Burchard, and S. Osher. Geometric surface smoothing via
anisotropic diffusion of normals. In Proceedings of IEEE Visualization, pages 125–132,
2002.
[38] G. Taubin. A signal processing approach to fair surface design. ACM Transactions on
Graphics (SIGGRAPH), pages 351–358, 1995.
[39] J. Vollmer and R. Mencl. Improved laplacian smoothing of noisy surface meshes. EURO-
GRAPHICS, 18(3):131–139, 1999.
[40] Y. Weiss. Segmentation using eigenvectors: a unifying view. In International Conference
on Computer Vision, pages 975–982, 1999.
[41] G. L. Xu. Discrete laplace-beltrami operators and their convergence. Computer Aided
Geometric Design, 21(8):767–784, 2004.
[42] H. Yagou, Y. Ohtake, and A. Belyaev. Mesh smoothing via mean and median filtering
applied to face normals. In Proceedings of Geometric Modeling and Processing, pages
124–131, 2002.
[43] D. M. Yan, Y. Liu, and W. P. Wang. Quadric surface extraction by variational shape
approximation. In Proceedings of Geometric Modeling and Processing, pages 73–86, 2006.
[44] Y. Yu, K. Zhou, D. Xu, X. Shi, H. Bao, B. Guo, and H.-Y. Shum. Mesh editing
with poisson-based gradient field manipulation. ACM Trans. Graphics (SIGGRAPH),
23(3):641–648, 2004.