TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5546 ~ 55
5
1
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.520
3
5546
Re
cei
v
ed
No
vem
ber 2
5
, 2013; Re
vi
sed
F
ebruary 25,
2014; Accept
ed March 1
1
, 2014
Fuzzy c-Means and Mean Shift Algorithm for 3D-Point
Clouds Denoising
Tonggua
ng Ni
1
, Xiaoqing Gu
2
, Hong
y
u
an Wa
ng*
3
Schoo
l of Information Sci
enc
e and En
gi
neer
ing, Ch
an
gzho
u Univ
ersit
y
, C
han
gzh
ou 21
3
164, Ch
in
a
T
e
lp 86-05
19-8
633
05
58, F
a
x
86-0
519-
86
33
0
284
*Corres
p
o
n
id
n
g
author, e-ma
i
l
: hb
xtntg-12
@
163.com
1
, guxqcczu@gmail.com
2
, tidd
yddd
@16
3
.com
3
A
b
st
r
a
ct
In many a
p
p
l
i
c
ations, de
no
i
s
ing is n
e
ces
s
ary
since p
o
i
n
t-sampl
ed
mode
ls obtai
ne
d
by laser
scann
ers w
i
th insufficie
n
t prec
ision. An
alg
o
ri
thm for
po
int-s
a
mpl
ed surfac
e is prese
n
ted,
w
h
ich combi
n
es
fu
zz
y
c-
me
ans
clusterin
g
w
i
th me
an sh
ift filterin
g al
gorith
m
. By using fu
zzy c-me
ans cl
us
tering, the l
a
rg
e
-
scale no
ise
is
del
eted an
d
a part
of
s
m
a
ll-s
c
ale no
ise
also
is s
m
oot
h. T
he cluster
cent
e
r
s are re
gar
de
d a
s
the new
po
ints. After acquiri
ng
new
point sets
bein
g
le
ss
noisy, the remains noise
i
s
smoo
th
b
y
m
e
an
sh
i
f
t
meth
od. Ex
peri
m
e
n
tal
resu
lts
de
mo
nstr
ate th
at the
alg
o
rith
m c
an
pro
duce
a
mor
e
accur
a
te p
o
int-sa
mp
l
e
mo
de
l efficientl
y
w
h
ile havi
ng
better feature p
r
eservati
on.
Ke
y
w
ords
:
po
i
n
t-sampl
ed
mo
del, mea
n
-shift
proced
ure, fu
zz
y
c-
mea
n
s
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Point-sample
d mod
e
ls
are no
rmally g
ener
ated
by sampli
ng th
e
boun
da
ry surface of
physi
cal 3
D
-scanni
ng de
vices. Despite the impr
ov
ement of scannin
g
accu
racy, the data is
invariably
noi
sy. Moreover, the incre
a
si
ng u
s
e
of
3
D
scanne
rs
has implie
d
a growth
in t
he
compl
e
xity of the scan
n
e
d
model
s. Therefo
r
e,
it is cru
c
ial tha
t
noisy mod
e
ls ne
ed to be
denoi
se
d or smooth
ed
before
perfo
rming any s
ubsequ
ent g
eometry p
r
o
c
e
ssi
ng such as
simplification,
re
con
s
tructi
on a
nd
parameteri
z
at
io
n
.
There i
s
a challe
nge
to remove
the
inevitable noi
se while pre
s
ervin
g
the unde
rlying
surface features in compu
t
er gra
phi
cs.
In
particula
r, fine feature
s
are often lost if
no sp
eci
a
l tre
a
tment is pro
v
ided [1, 2].
In rece
nt year, a variety of denoisin
g
me
thod
s h
a
ve been int
r
udu
ce
d, su
ch as the
Lapla
c
ian o
p
e
rato
r [3], anisotro
p
ic diffu
sion [4, 5], diffusion of the normal field [6], and local
l
y
adaptive
Wie
ner filtering
[7
]. A mean
-shi
ft-base
d
ani
sotropi
c
denoi
sing
alg
o
rith
m [8] is p
r
op
o
s
ed
for poi
nt-sam
pled
su
rfaces. Takin
g
into
accou
n
t
the
vertex no
rma
l
and
cu
rvatu
r
e a
s
th
e ran
ge
comp
one
nt, the algo
rithm
extend mean shift filter
ing to 3D surf
ace
smoothi
n
g
. By clusteri
ng
adja
c
ent sa
mple point
s of similar l
o
cal m
ode
s,
the method
also p
r
ovid
es a me
ani
ngful
segm
entation
of the point model. The
neighbo
rs
of each sam
p
l
e
point are
colle
cted un
d
e
r
spatial an
d ra
nge co
nst
r
ain
t
s. Finally, th
e prop
osed
trilateral poi
nt filtering alg
o
rit
h
m can remo
ve
noise
while
p
r
eserve
ge
om
etric f
eatures.
Although
the
method
is efficient, ove
r
smoothing
will
be
prod
uced wh
en the mesh is suffe
red fro
m
large
-
scal
e
noise.
In this pape
r, a two-stage
point clo
u
d
s
denoi
sing m
e
thod is propo
sed, which combine
s
fuzzy c-mea
n
s
with
the me
an-shift-b
a
se
d
ani
sotr
opi
c
denoi
sing
of
point-sam
ple
d
surfa
c
e
s
. T
h
is
algorith
m
can
handle the large
-
scale an
d small-
scale
noise an
d be formulated
for mesh-b
ase
d
geomet
ry and
even for gen
eral 3
D
geo
m
e
try.
2. Fuzzy
c-means
Fuzzy c-m
e
a
n
s (F
CM) i
s
a famous meth
od of cluste
ri
ng whi
c
h allo
ws on
e pie
c
e
of data
to belong to two or m
o
re
cl
usters. It is base
d
on
mini
mization of th
e followin
g
ob
jective functio
n
:
2
11
,,
1
NC
m
mj
k
j
k
jk
Jc
p
c
m
(1)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fuzzy c-Mea
n
s an
d Mean
Shift Algorithm
for 3D-Poi
nt Cloud
s De
noisi
ng (T
ong
guan
g Ni)
5547
Whe
r
e
m
is
an
y r
e
a
l
n
u
mbe
r
gr
ea
te
r
than
1
,
μ
jk
is the
deg
ree
of m
e
mbe
r
ship
p
j
in
the clu
s
ter
k
,
p
j
is the
j
th d
-
dime
nsi
onal
points,
c
k
is the cl
uste
r d
-
d
i
mensi
onal
center of
the cluster,
is
any
norm exp
r
e
s
sing the simila
rity between
any measure
d
data and th
e cen
c
ter.
This p
ape
r a
dopts th
e imp
r
oved m
e
tho
d
of the fuzzy clu
s
terin
g
[9, 10], which d
e
fined a
fuzzy
weig
hting coefficie
n
t, it makes
sh
ort di
sta
n
ces
become m
u
ch sh
orte
r an
d
long di
stan
ces
become mu
ch longe
r. So
the perfo
rm
ance of the
clu
s
terin
g
be
come
s b
e
tter clu
s
terin
g
. The
detail pro
c
e
d
u
re
s of the fuzzy cl
uste
ri
ng
algorithm a
r
e
given in [11].
3. Fuzzy
c-means Clu
s
te
r
i
ng
w
i
th Mea
n
Shift Filteri
ng Algorith
m
The main idea of our algorithm is that, fi
rstly the noi
sy data points
will be pre-processed
by an improved fuzzy c-m
ean
s cl
uste
ri
ng metho
d
. F
o
r ea
ch
data
point, we d
e
tect the n
u
mb
er of
the neigh
bori
ng point in th
e given sp
he
re in ord
e
r
to
determi
ne if it is noi
se or
n
o
t. If
the num
ber
of the neighb
oring
point is less th
an th
e given thre
shold, the poin
t
is a noise; otherwise we
will
clu
s
ter the p
o
i
nts in the sp
here
and reg
a
rd the
clu
s
te
r ce
nter a
s
th
e new p
o
int, whi
c
h can filter
the noise near the point sets. Thi
s
process w
ill preserve som
e
small-scale
noise. Using the
vertex no
rmal
and
the
curv
ature
a
s
the
range
co
mpo
n
ent an
d the
v
e
rtex p
o
sition
as the
spatia
l
comp
one
nt, the lo
cal
mo
d
e
of e
a
ch ve
rtex on p
o
int-based su
rfa
c
es
i
s
co
mput
ed
by a
3
D
mean
shift procedu
re dep
end
ent
on lo
cal
neig
hborhoo
ds th
at are
ada
ptively obtaine
d
by a kdtre
e
d
a
ta
stru
cture. Clusteri
ng pie
c
es
of point-based su
rfa
c
e
s
of si
milar local mo
de can p
r
ov
ide
meanin
g
ful m
odel
seg
m
ent
ation. The
n
, a trilateral
poi
nt filtering
scheme i
s
a
ppli
ed ba
se
d on
the
adaptively clu
s
tere
d nei
ghb
ors. T
he sch
e
m
e ca
n adju
s
t the position
of sampl
e
poi
nts alon
g thei
r
norm
a
l dire
cti
ons. Fin
a
lly the noi
se is
redu
ced fr
o
m
point-sam
ple
d
su
rfaces
succe
ssfully while
pre
s
e
r
ving ge
ometri
c features.
3.1. The Fuzz
y
c-means Algorithm fo
r Large Scal
e Noise
We d
e
fine th
at
s
is the
su
rro
undi
ng
sp
here,
r
is the
radiu
s
of
s
,
and
size is the given
threshold
of numbe
r of cl
ose
poi
nt
s in
the su
rroundi
ng sphe
re
s
and
m
i
is the
numbe
r of cl
ose
points in the
surro
undi
ng sphere
s
of poi
nt
p
i
.
Figure 1. Small-scal
e noi
se partly filtere
d
and La
rge
-
scale noi
se d
e
leted by FCM. The
para
m
eter
size is define
d
a
s
3. The re
d points a
r
e noi
se. (a
) The
seco
nd poi
nt is noi
se,
m
2
=2,
m
2
<size, we
deleted it. The se
con
d
poi
nt is noise,
m
3
=
3
=s
iz
e, s
o
it is
moved to c
l
us
tering c
e
nter
of points in th
e sph
e
re by clusteri
ng met
hod.
(b
) The f
i
rst and thi
r
d
points a
r
e noi
se. Because
the numbe
r o
f
close p
o
ints
in the surro
u
n
d
ing sphe
re i
s
larg
er tha
n
size, we re
ga
rd them a
s
sampl
e
point
s
Noi
s
e
p
oint
Surround
ing sp
here
Sample point b
e
fore clustering
Sample and no
ise point pos
itio
n after clusterin
g
Noise point dele
t
ed
2
3
1
1
2
3
(
a
)
(
b
)
Sam
p
le
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5546 – 55
51
5548
The followi
ng
is the pse
u
d
o
-code fo
r ap
plying fuzzy c-mea
n
s
clu
s
tering to a si
n
g
le point:
Large-scale DenoisePoint
(point
p
i
)
{
k
ij
} = neighbo
rhood (
p
i
)
For
i
: =
1
to
N
If (
m
i
<size)
Delete
p
i
Els
e
Call FCM ( )
End
c
i
=fuzzy c-me
ans clustering
center of
p
i
Return new point
q
i
=
c
i
In Figure 1,
we
can
se
e that larg
e-scal
e noi
se i
s
de
leted an
d sm
all-scal
e noi
se partly
filtered by FCM algorithm.
But FCM ca
n’
t delete sm
all
-
scale n
o
ise, it only partly smooth them. In
the next secti
on we
smoot
h them by bilateral filter.
3.2. Estimation of Norm
a
l
Vector and
Curv
ature
The n
o
rm
als and
cu
rvatu
r
es of the
p
o
int-
sampl
e
d
geom
etry can be
e
s
tim
a
ted by
various
methods
to es
timate [12,13]. As
s
u
me
p
j
,
j
=1, 2,…,
m
is a su
bset of the origin
al
measuri
ng p
o
int set
P
={
p
1
,
p
2
,…,
p
m
}, base
d
on th
e theory of prin
cipal
com
pone
nt analy
s
is
(PCA), the 3
×
3 covari
an
ce
matrix of
p
j
could be d
e
fin
ed as follo
w:
12
1
2
[
,
,
.
..,
]
[
,
,
...,
]
T
in
n
jj
j
j
j
j
C
p
pp
p
p
p
p
pp
p
p
p
(2)
Whe
r
e
C
i
s
a
symmetri
c
p
o
s
itive se
mi-d
e
f
inite matrix, and the
centroid of
p
j
is
1
1
n
i
j
i
p
p
n
. The
norm
a
l of
p
i
i
s
cho
s
en to
b
e
the unit ve
ctor
e
i,
1
, whi
c
h
co
rre
sp
ond
s
to the minima
l eigenval
ue
of
C
i
.
n
i
= e
i,
1
(3)
Pauly et al. [14] sho
w
e
d
th
at su
rface variat
ion is
clo
s
e
l
y related to
mean
cu
rvatu
r
e, an
d
here the
cu
rvature on p
i
i
s
taken a
s
the
surfa
c
e va
riat
ion,
H
i
=
λ
i,
1
/
(
λ
i,
1
+
λ
i,
2
+
λ
i,
3
)
(4)
Whe
r
e
λ
i,j
,
j
= 1, 2, 3 are eigenvalu
e
s of
C
i
and satisfy
λ
i,
1
≤
λ
i,
2
≤
λ
i,
3
.
3.3. Mean Shift De
noising
Since the
sp
a
t
ial and ra
ng
e domai
ns of
3D ge
om
etry
are
slightly di
fferent from t
hose of
image
s, 3D
positio
n of a
vertex is u
s
u
a
lly rega
rd
a
s
spatial info
rmation, b
u
t in this pa
pe
r we
rega
rd the
n
o
rmal a
nd
curvature of the local
surf
ace
as
ran
g
e
inform
ation
or feature space
informatio
n. We extend th
e mean shift algorith
m
to 3D domai
n directly.
We a
ssu
me that the data points
p
i
are the genera
lized poi
nts
of the input raw point
model
3
PR
, and the sp
atial po
sition informa
t
ion
v
i
=
(
x
i
,
y
i
,
z
i
) and
rang
e
information i
n
clu
d
ing
the norm
a
l vector
n
i
an
d th
e mean
curva
t
ure
H
i
of vert
ice
s
are in
clu
ed in the vect
or compo
nent
s
of
p
i
, which
can be written as:
p
i
=(
v
i
,
n
i
,
H
i
)
(5)
With
i
=
1, 2,
. . . ,
n
, and
n
being the
nu
mber
of point
s in
P
. He
re t
he dime
nsi
o
n
of vector
p
i
is 7
.
The
k
n
earest
neighb
orin
g points of ge
n
e
rali
zed p
o
int
s
p
i
are d
enot
ed by
N
(
p
i
)=
{
q
i,
1
,
q
i
,2
,…
, q
i,k
}.
Thus, the me
an shift vecto
r
of
p
i
can b
e
expre
s
sed a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fuzzy c-Mea
n
s an
d Mean
Shift Algorithm
for 3D-Poi
nt Cloud
s De
noisi
ng (T
ong
guan
g Ni)
5549
k
j
r
ij
r
i
k
j
i
ij
r
ij
r
i
i
v
q
p
g
p
M
q
q
p
g
p
M
1
1
(6)
Whe
r
e g
(
·) could be
eithe
r
a Ga
ussian
kernel o
r
an
Epane
chni
ko
v kern
el;
p
i
r
=(
n
i
,
H
i
)i
s the ra
ng
e
part of
p
i
, a
nd
M
(
p
i
) i
s
called the m
e
an shift poin
t
associ
ated
with
p
i
, and
M(
p
i
) co
uld be
initialized to
coincide with
p
i
,
M
v
(
p
i
) is
the mean
shi
ft vector associate
d
with
M(
p
i
). Th
en
we
define the me
an shift proce
dure
as
the
repeate
d
mov
e
ment of dat
a
points to the
sampl
e
mea
n
s
,
written a
s
:
M
(
p
i
)
:=
M
(
p
i
)
+
M
v
(
p
i
)
(7)
By running th
e pro
c
ed
ure for all
i
= 1, 2,…,
∞
, each d
a
ta point itera
t
es to a local
mode in
the joint spati
a
l rang
e dom
ain, and the mean shi
ft proce
dure ha
s provide
d
a st
able local mo
de
detectio
n
for the point-sa
m
pled mod
e
l.
3.4. Anisotr
o
pic Denoisin
g
Al
gorithm
for Small Scale Noise
The detaile
d mean shift denoisi
ng algo
ri
thm con
s
ist
s
of four pro
c
e
s
sing
stage
s.
Step 1. Ini
t
ializ
ation
.
Co
nstru
c
t a
kdtree
structu
r
e
for the
poi
n
t
model
and
se
arch
neigh
bors {
s
ij
q
} for ea
ch
v
i
in the spatial d
o
m
ain, then initialize ra
nge
comp
one
nt
r
i
p
by princip
a
l
comp
one
nt analysi
s
of the spatial n
e
i
ghbo
rs {
s
ij
q
}. The ran
ge ba
ndwi
d
ths
12
,
rr
r
hh
h
=
are
usu
a
lly defined as po
sitive values relat
ed to nor
mal
and curvatu
r
e, resp
ectivel
y
; for formulation
conve
n
ien
c
e,
we write
h
r
di
rectly.
Step 2. Mea
n
shift proc
e
dure
. Rep
eat
the mean shi
ft proced
ure discu
s
sed ab
ove until
conve
r
ge
nce.
Step 3.
Clu
s
tering
.
Buil
d clu
s
ters of
p
o
ints wh
ose mode
s are
similar.
G
ene
rally,
the
neigh
borhoo
d
size
k
is an
essential
pa
rameter fo
r g
ood
sha
pe
smoothing
re
sults. Pauly et
al.
[15] sugg
est
we sele
ct a
k
in the rang
e from 6 to 20 in
the spatial d
o
main.
Step 4. Verte
x
estimation
.
After updating the rang
e compon
ent of
pi
and its nei
ghbo
rs,
we ap
ply trilateral filtering i
n
the influen
ce r
egio
n
asso
ciated
with a fixed local m
ode. Com
p
a
r
ed
with bilate
ral
filtering, we n
o
t only sepa
rate s
patial a
n
d
ra
nge
sig
n
a
l
s to d
e
termi
n
e the lo
cal
area
with ge
ometri
c
cohe
re
nce, but al
so intro
duce a
cu
rvature
r
elate
d
ke
rnel to
sm
oot
h high
gradie
n
t
regio
n
s
effici
ently.Our ap
p
r
oa
ch i
s
slig
htly differ
ent fro
m
the trilatera
l normal filteri
ng propo
se
d in
[16, 17].
The cu
rvatures are
con
s
id
ered
a
s
se
co
nd-o
r
d
e
r
pro
p
e
rties of the
3D
geom
etry, and
the
perfo
rman
ce
of the cu
rvat
ure
r
elate
d
ke
rnel fo
r the
region
s n
ear
the sali
ent ri
dge a
nd ravine
st
ru
ct
ur
es i
s
sat
i
sf
a
c
t
o
ry
.
12
3
,
1
S
ij
ii
i
i
S
ii
i
i
i
j
i
q
ii
i
i
i
vv
n
wG
n
v
q
g
GG
d
G
h
G
e
r
(8)
Where
G
is a
Gaussian kernel,
d
i
is the
distance of
ǁ
v
i
-
s
ij
q
ǁ
,
h
i
is the pro
j
ection of
the vector (
v
i
-
s
ij
q
)
onto
the no
rmal
n
i
, and
e
i
is the in
verse of
the curvature difference between
v
i
and
s
ij
q
.
Step 5.
Ada
p
tiv
e
neighbors
. For l
a
rg
e point d
a
ta
sets, the
sel
e
ction of n
e
igh
bors i
s
a
trade
-off betwee
n
co
mpu
t
ational time
and sm
ooth
i
ng quality. Whe
n
ch
oo
si
ng a sm
all
and
uniform
spati
a
l size of a
neigh
borhoo
d
,
it cost
s le
ss in comp
uta
t
ional time, b
u
t smooth
s
t
h
e
model
poo
rly; on
the
other
hand, if
we
choo
se
a la
rg
e an
d u
n
iformly sp
aced
n
e
ighb
orh
ood,
the
model i
s
ove
r
smooth
ed.
We use
the
kdt
r
ee to
sea
r
ch
for the
k
ne
a
r
est
neigh
bo
rs, for in
stan
ce
k
= 1
2
, inste
a
d
of a n
e
igh
b
o
ring
spatial
ball,
whi
c
h
is unfortu
natel
y depen
dent
on the
sa
mpl
i
ng
den
sity of the point mod
e
l. Furthe
rmo
r
e
,
our
cluste
ri
ng sch
e
me p
r
ovide
s
an
a
daptive nei
gh
bor
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5546 – 55
51
5550
sea
r
ching me
thod, whe
r
e
different influ
ence regi
on
s are used ad
aptively to remove noise from
vertice
s
. Ap
plying trilate
ral filtering
t
o
t
he ada
ptive
neig
hbo
rhood
s gre
a
tly
improve
s
the
smoothi
ng
ca
pabilitie
s of th
e mean
shift f
ilter in
hig
h
g
r
adient regio
n
s
. Although th
e algo
rithm
we
pre
s
ent i
s
ordinarily n
on-i
t
erative whil
st denoi
si
ng,
we
can al
so
use the
re
sulting ada
ptive
neigh
bors
as
inputs to e
s
ti
mate rang
e i
n
formatio
n,
then ite
r
ativel
y perfo
rm th
e
four ste
p
s of
the
algorith
m
to make the p
o
i
n
t model sm
o
o
ther.
4. Experimental Re
sults
and An
aly
s
is
Usi
ng a PC o
f
Intel core2 Q955
0 and 8
G
B memory, this pape
r poi
nts of differen
t
models
and prede
ce
ssors of the d
enoi
si
ng alg
o
r
ithms exp
e
ri
ment and
co
mpared, We i
m
pleme
n
ted the
point clo
u
d
s
denoi
sing
alg
o
rithm a
s
de
scrib
ed in the
previou
s
sect
ion and g
a
ve our results, a
nd
our alg
o
rithm
and previou
s
algorith
m
s
were comp
are
d
.
A compa
r
iso
n
to the mean shift smoot
hing app
ro
ach is sho
w
n in
Figure 3. Fi
gure 3
(
a
)
rand
om
noi
se
is ad
ded
to t
he m
oai
mod
e
l (th
e
n
u
mb
er
of the
noi
se is 6
543,
an
d the
num
ber of
the point i
s
2000
0). Fig
u
r
e 3
(
b
)
is th
e re
sult
s u
s
i
ng the m
ean
shift poi
nt cloud
s de
noi
si
ng
method. In Fi
gure
3(c) the
noise
delete
d
by FC
M
with paramete
r
size=6. In Figure
3(d
)
sm
all-
scale noi
se
smoothi
ng b
y
mean shift point cl
oud
s denoi
sing m
e
thod. We
can se
e that our
results of the
two mod
e
ls
are b
e
tter th
an the me
an
shift point cl
oud
s den
oisi
ng metho
d
while
handli
ng the
large
-
scale n
o
ise a
nd the
mean
shift p
o
int clou
ds
d
enoi
sing m
e
thod will
pro
d
u
ce
oversmoothi
n
g
in sha
r
p fea
t
ures.
5. Conclusio
n
This p
ape
r h
a
s p
r
e
s
ente
d
a two-stage
point clo
ud d
enoi
sing met
hod which co
mbining
fuzzy
c-mea
n
s
with me
a
n
shift filterin
g app
ro
ac
h.
Our algo
rith
ms
h
a
ve
a g
ood re
sult
while
workin
g with unorgani
ze
d and
la
rg
e-scale
noi
sy
p
o
int set
s
. But it has a di
sa
d
v
antage that
the
improved FCM will partly smooth sharp
feature
while
clustering. In Figure 3(d),
we may see
hat
the se
con
d
p
o
int moves to
wards its
clo
s
e points.
In the future, we h
ope to improve our ap
proa
ch
in orde
r to prese
r
ve the sh
arp featu
r
e
s
of models b
e
tter.
(a)
(b)
(c
)
(d)
Figure 2. Noi
s
y point delet
ed by our alg
o
rithm. (a
) Th
e noisy poi
nt clou
ds, (b
) La
rge
-
scal
e noi
se
deleted by F
C
M, (c) Small
s
cale noi
se
smoothi
ng by
mean shift point cloud
s de
noisi
ng meth
od,
(d) T
he mod
e
l
feature is p
r
ese
r
ved
(a)
(b)
(c
)
(d)
Figure 3. Co
mpari
s
o
n
wit
h
mean shift point clo
u
d
s
denoi
sing m
e
thod for the
moai. (a) i
s
th
e
noisy poi
nt cl
oud
s, (b) i
s
the result usin
g
mean
shift a
ppro
a
ch, (c) is Large
-scal
e
noise d
e
lete
d
by FCM, (d)
shows noi
se
smoothing by
mean shift point clou
ds d
enoi
sing
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fuzzy c-Mea
n
s an
d Mean
Shift Algorithm
for 3D-Poi
nt Cloud
s De
noisi
ng (T
ong
guan
g Ni)
5551
Ackn
o
w
l
e
dg
ements
The mod
e
ls
are of co
urte
sy Y. Ohtake
, and
thanks to Y. Ohtake forproviding the ba
sic
platform, so
we
can i
m
ple
m
ent ou
r alg
o
rithm effici
e
n
tly. This work was fu
nde
d
by the Natio
nal
Natural Scie
n
c
e Fou
ndatio
n of China (No. 61070
121
).
Referen
ces
[1]
Z
hang
X, Li H,
Chen
g Z
,
Z
hang Y. Rob
u
st Curvat
ur
e Estimation a
nd Ge
ometr
y
A
nal
ys
i
s
of 3d Poin
t
C
l
o
ud
Su
rfa
c
es.
Journal of C
o
mputati
o
n
a
l Informatio
n
Systems
. 20
09; 6(
5): 1983-
19
90.
[2]
Ni T
,
Ma Z.
F
a
st algor
ith
m
for
poi
nt-sa
m
pl
ed
mode
l d
e
n
o
isi
n
g
. Intern
ation
a
l C
onfer
ence
on C
o
mp
ute
r
Engi
neer
in
g an
d T
e
chnolo
g
y
.
2010; 7: 71
9-7
22.
[3] T
aubin
G.
A si
gna
l process
i
n
g
appr
oac
h to fair surface d
e
si
gn
. Proc. SIGGRAPH. 95: 35
1
–35
8.
[4]
Baja
j C,
Xu G. Anisotro
pic d
i
ffusion
on surfa
c
e an
d functi
o
n
on s
u
rfaces.
ACM T
r
ans. o
n
Graph
ics.
200
3; 22: 4–3
2
.
[5]
Clare
ntz U,
Di
e
w
al
d
U, R
u
m
p
f M.
Anis
otro
pic
geo
metric
diffusio
n
i
n
s
u
r
f
ace pr
ocessi
n
g
. Proc. IEEE
Visua
lizati
on. 2
000: 39
7–
40
5.
[6]
Ohtake Y, Baly
aev A, S
e
id
el
H,
Mesh s
m
oot
hin
g
by
ad
aptiv
e a
nd
anis
o
tro
p
ic Ga
ussia
n
fil
t
er ap
pli
ed t
o
the mesh nor
mal
. Proc. Vision
, Modeli
ng, an
d Visua
lizati
on.
2002; 2
03–
21
0.
[7] Alex
a
M.
W
i
en
er filterin
g of meshes
. Proc. Shap
e Mod
e
li
ng
Internation
a
l.
200
2; 51–
57.
[8]
Hu G, Peng Q,
F
o
rrest A. Mean sh
ift deno
isi
ng of p
o
int-sa
mple
d surfaces
.
T
he Visua
l
Co
mp
uter
. 20
06
;
22(3): 14
7-1
5
7
.
[9]
Liu W, Chen Z
,
Bai P, Fang
S, Shi Y.
A
ki
n
d
of
i
m
prov
ed
meth
od
of fu
zzy clusteri
ng
. Pr
ocee
din
g
s
of
the F
ourth Inte
rnatio
nal C
onfe
r
ence o
n
Mach
ine
L
earn
i
n
g
a
nd C
y
b
e
rnetics
.
2005; 5: 264
6
- 2649.
[10]
Ni T
,
W
ang H,
Gu
X. A fast
surface rec
ons
tr
uction
alg
o
rit
h
m b
y
r
obust
fuzz
y
C me
an
s an
d d
o
w
n
sampli
ng.
Jour
nal of Co
mputa
t
iona
l Informati
on Syste
m
s
. 2013; 9(3): 9
07-
914.
[11]
W
ang L, Y
uan B,
T
ang
X.
Imp
licit Surface Re
constructio
n
from N
o
isy 3D S
c
attered Data
. Procee
din
g
s
of the Eighth In
ternatio
nal C
o
n
f
erence o
n
Sig
nal Proc
essin
g
.
2006; 2: 134-
142.
[12]
Hu G, Peng Q
.
Bilatera
l esti
mati
on
of nor
ma
l o
n
p
o
int-s
a
mpl
ed
mo
de
l
s
. Internatio
nal
Confer
enc
e o
n
Comp
utation
a
l
Scienc
e an
d Its Applic
ations.
200
5; 758
–7
68
.
[13]
Hop
pe H, DeR
o
se T
,
Duch
a
m
p T
,
McDonal
d J, Stuetzle W
.
Surface reconstructio
n
from un
org
ani
z
e
d
poi
nts
. Proceedings of ACM SIGGRAPH. 1992; 71–78.
[14]
Paul
y
M, Gros
s M, Kob
b
l
e
t L
.
Efficient Si
mp
lificatio
n
of Po
i
n
t-Sampl
ed G
e
ometry
. IEEE
Visualization.
200
2; 163
–1
70
.
[15]
Paul
y M, Keis
er
R, Kobbl
et L, Gross M. Shap
e mod
e
li
ng
w
i
th p
o
i
n
t-sa
mple
d geom
etry.
ACM T
r
ans.
Graph
. 200
3; 2
2
(3): 641
–6
50.
[16]
Cho
udh
ur
y P, T
u
mblin J.
T
he trilatera
l filte
r
for high co
ntrast ima
ges a
nd
mesh
es
. Procee
din
g
s o
f
Eurogr
aph
ics. 200
3; 186
–1
96
.
[17]
Lou
ie
H, Bur
n
s M, Lim
a
C.
An introduction and user's
guide
to the IEEE S
m
art Grid Web P
o
rtal
.
Innovativ
e Smart Grid
T
e
chn
o
lo
gies C
onf. Europ
e
. 201
0; 1–5.
Evaluation Warning : The document was created with Spire.PDF for Python.