Mumford
Mumford - - Shah Shah
Energy Functional Energy Functional
Zoltan
Zoltan Kato Kato
http://www.cab.u
http://www.cab.u - - szeged.hu/~kato/variational/ szeged.hu/~kato/variational/
Introduction Introduction
! ! Proposed in their influential paper by Proposed in their influential paper by
! ! David David Mumford Mumford
! ! http:// http:// www.dam.brown.edu/people/mumford www.dam.brown.edu/people/mumford / /
! ! Jayant Jayant Shah Shah
! ! http://www.math.neu.edu http:// www.math.neu.edu/~shah/ /~shah/
Optimal Approximations by Piecewise Smooth Optimal Approximations by Piecewise Smooth
Functions and Associated
Functions and Associated Variational Variational Problems
Problems . . Communications on Pure and Applied Communications on Pure and Applied Mathematics, Vol. XLII, pp 577
Mathematics, Vol. XLII, pp 577 - - 685, 1989 685, 1989
Images as functions Images as functions
! ! A gray A gray - - level image represents the light level image represents the light intensity recorded in a plan domain
intensity recorded in a plan domain R R
! ! We may introduce coordinates We may introduce coordinates x x , , y y
! ! Let Let g(x,y g(x,y ) ) denote the intensity recorded at the denote the intensity recorded at the point
point ( ( x,y x,y ) ) of of R R
! ! The function The function g(x,y g(x,y ) ) defined on the domain defined on the domain R R is is called an image.
called an image.
What kind of function is g?
What kind of function is g?
! ! The light reflected by the The light reflected by the surfaces
surfaces S S
iiof various of various objects
objects O O
iiwill reach the will reach the domain
domain R R in various open in various open subsets
subsets R R
ii! ! When When O O
11appears as the appears as the background to the sides of background to the sides of O O
22then the open sets then the open sets R R
11and and R R
22will have a will have a
common boundary (
common boundary ( edge edge ) )
! ! One usually expects One usually expects g(x,y g(x,y ) ) to be discontinuous along to be discontinuous along
this boundary
this boundary
FigureFigurefromfromD. MumfordD. Mumford& J. & J. ShahShah: : Optimal Approximations by Piecewise Smooth Functions Optimal Approximations by Piecewise Smooth Functions and Associatedand Associated VariationalVariationalProblemsProblems. . Communications on Pure and Applied Mathematics, Communications on Pure and Applied Mathematics, Vol. XLII, pp 577
Vol. XLII, pp 577--685, 1989685, 1989
Other discontinuities Other discontinuities
! ! Surface orientation of visible objects (cube) Surface orientation of visible objects (cube)
! ! Surface markings Surface markings
! ! Illumination (shadows, uneven light) Illumination (shadows, uneven light)
Piece
Piece - - wise smooth g wise smooth g
! ! In all cases, we expect In all cases, we expect g(x,y g(x,y ) ) to be piece to be piece - - wise wise smooth to the first approximation.
smooth to the first approximation.
! ! It is well It is well modelled modelled by a set of smooth functions by a set of smooth functions f f
iidefined on a set of disjoint regions
defined on a set of disjoint regions R R
iicovering covering R R . .
! ! Problems: Problems:
! ! Textured objects (regions perceived homogeneous but Textured objects (regions perceived homogeneous but lots of discontinuities in intensity)
lots of discontinuities in intensity)
! ! Sahdows Sahdows are not true discontinuities are not true discontinuities
! ! Partially transparent objects Partially transparent objects
! ! Noise Noise
! ! Still widely and Still widely and succesfully succesfully applied model! applied model!
Segmentation problem Segmentation problem
! ! Consists in computing a decomposition of Consists in computing a decomposition of the domain of the image
the domain of the image g(x,y g(x,y ) )
1. 1. g g varies varies smootly smootly and/or slowly within and/or slowly within R R i i 2. 2. g g varies discontinuously and/or rapidly varies discontinuously and/or rapidly
across most of the boundary
across most of the boundary Γ Γ between between regions
regions R R i i
U n
i
R i
R
= 1
=
Optimal approximation Optimal approximation
! ! Segmentation problem may be restated as Segmentation problem may be restated as
! ! finding optimal approximations of a general function finding optimal approximations of a general function g g
! ! by piece by piece - - wise smooth functions wise smooth functions f f , whose restrictions , whose restrictions f f
iito the regions
to the regions R R
iiare are differentiable differentiable
! ! Many other applications: Many other applications:
! ! Speech recognition Speech recognition
! ! Sonar, radar or laser range data Sonar, radar or laser range data
! ! CAT scans CAT scans
! ! etc etc … …
Optimal segmentation Optimal segmentation
! ! Mumford Mumford and Shah studied 3 and Shah studied 3 functionals functionals which which measure the degree of match between an image measure the degree of match between an image
g(x,y
g(x,y ) ) and a segmentation. and a segmentation.
! ! First, they defined a general functional First, they defined a general functional E E (the (the famous
famous Mumford Mumford - - Shah functional): Shah functional):
! ! R R
iiwill be disjoint connected open subsets of the planar will be disjoint connected open subsets of the planar domain
domain R R , each one with a piece , each one with a piece - - wise smooth boundary wise smooth boundary
! ! Γ Γ will be the union of the boundaries. will be the union of the boundaries.
C C Γ
=
= n
i
R i
R
1
Mumford
Mumford - - Shah functional Shah functional
! ! Let Let f f differentiable on differentiable on ∪ ∪ R R i i and allowed to and allowed to be discontinuous across
be discontinuous across Γ Γ . .
! ! The smaller The smaller E E , the better , the better (f, (f, Γ Γ ) ) segments segments g g 1. 1. f f approximates approximates g g
2. 2. f f (hence (hence g g ) does not vary much on ) does not vary much on R R i i s s
3. 3. The boundary The boundary Γ Γ be as short as possible. be as short as possible.
! ! Dropping any term would cause Dropping any term would cause inf inf E=0 E=0 . .
Γ +
∇ +
−
=
Γ µ ∫∫ f g dxdy ∫∫ − Γ f dxdy ν
f
E R R
2 2
2 ( )
) ,
(
Cartoon image Cartoon image
! ! (f, (f, Γ Γ ) ) is simply a cartoon of the original image is simply a cartoon of the original image g g . .
! ! Basically Basically f f is a new image with edges drawn sharply. is a new image with edges drawn sharply.
! ! The objects are drawn The objects are drawn smootly smootly without texture without texture
! ! (f, (f, Γ Γ ) ) is essentially an idealization of is essentially an idealization of g g by the sort of by the sort of image created by an artist.
image created by an artist.
! ! Such cartoons are perceived correctly as representing Such cartoons are perceived correctly as representing the same
the same scane scane as g as g " " f f is a simplification of the scene is a simplification of the scene containing most of its essential features.
containing most of its essential features.
Cartoon image example
Cartoon image example
Related problems Related problems
! ! D. D. Geman Geman & S. & S. Geman Geman : Stochastic relaxation, : Stochastic relaxation,
Gibbs distribution and the Bayesian restoration of Gibbs distribution and the Bayesian restoration of
images.
images. IEEE Trans. on PAMI 6, pp 721 IEEE Trans. on PAMI 6, pp 721 - - 741, 741, 1984.
1984.
! ! MRF model MRF model
! ! A. Blake & A. A. Blake & A. Zisserman Zisserman : Visual Reconstruction. : Visual Reconstruction.
MIT Press, 198 MIT Press, 198 7 7
! ! Weak membrane model Weak membrane model
! ! M. M. Kass Kass , A. , A. Witkin Witkin & D. & D. Terzopoulos Terzopoulos : Snakes: : Snakes:
Active contour Models.
Active contour Models. International Journal of International Journal of Computer Vision, vol. 1, pp 321
Computer Vision, vol. 1, pp 321 - - 332, 1988 332, 1988 . .
! ! Active contour model Active contour model
approximation approximation
! ! A special case of A special case of E E where where f= f= a a i i is constant is constant on each open set
on each open set R R i i . .
! ! Obviously, it is minimized in Obviously, it is minimized in a a i i by setting by setting a a i i to the mean of
to the mean of g g in in R R i i : :
Γ +
−
=
Γ ∑ ∫∫
−
2 2
2 ( , ) ( )
µ µ ν
R
ii
i
dxdy a
g f
E
) ) (
(
i R
R
i area R
gdxdy g
mean
a
ii
= ∫∫
=
approximation approximation
! ! It can be proven that minimizing It can be proven that minimizing E E 0 0 is well is well posed:
posed:
! ! For any continuous For any continuous g g , there exists a , there exists a Γ Γ made up made up of of finit finit number of singular points joined by a number of singular points joined by a finit finit
number of arcs on which
number of arcs on which E E
00atteins atteins a minimum. a minimum.
! ! It can also be shown that It can also be shown that E E 0 0 is the natural is the natural limit functional of
limit functional of E E as as µ µ # # 0 0
Γ +
−
=
Γ ∑ ∫∫
2 20
( ) ( ( ))
µ ν
i i
R R
i
dxdy g
mean g
E
Relation to the
Relation to the Ising Ising model model
! ! If we further restrict If we further restrict f f to to
! ! take only values of +/ take only values of +/ - - 1, 1,
! ! assume that assume that g g and and f f are defined on a lattice are defined on a lattice
! ! then then E E
00becomes the energy of the becomes the energy of the Ising Ising model
model . .
Relation to the
Relation to the Ising Ising model model
! ! Γ Γ is the path between all pairs of lattice points on is the path between all pairs of lattice points on which
which f f changes sign: changes sign:
∑
∑ − + −
=
) , ( ), , (
2 ,
2
*
0
( ) ( ( , ) ( , ))
2( ( , ) ( , ))
l k j i j
i
l k f
j i f j
i g j
i f f
E
µνFigure
FigurefromfromD. MumfordD. Mumford& J. & J.
ShahShah: : Optimal Optimal Approximations by Approximations by Piecewise Smooth Piecewise Smooth Functions and Associated Functions and Associated Variational
VariationalProblems. Problems. Communications on Pure and Communications on Pure and Applied Mathematics, Vol.
Applied Mathematics, Vol.
XLII, pp 577
XLII, pp 577--685, 1989685, 1989
Weak Weak string string
! ! Fitting Fitting an an elastic elastic spline spline with with possible possible breaks breaks ( ( line line process
process or or local local edges edges ) )
! ! Remove Remove noise noise
! ! Approximate Approximate with with smooth smooth curves curves
! ! Breaks Breaks where where smoothness smoothness is is not not satisfied satisfied
Image
Image fromfromCMU 15CMU 15--385 Computer 385 Computer VisionVisioncourse, course, SpringSpring2002 2002 bybyTaiTaiSing Sing LeeLee
Energy
Energy of of a a weak weak string string
! ! α α is is the the cost cost of of inserting inserting a a break break ( ( local local edge edge element
element ) ) l l i i
! ! l l i i may may take take binary binary values values [0,1] [0,1]
! ! l l i i is is turned turned on on when when ( ( f f i+1 i+1 - - f f i i ) ) 2 2 > > α α / / λ λ
∑
∑
∑ − + − − +
=
+i i i
i i
i i
i
i
d f f l l
f f
E ( ) ( )
2λ (
1)
2( 1 ) α
Weak Weak membrane membrane model model
Image
Image fromfromCMU 15-CMU 15-385 Computer 385 Computer VisionVisioncourse, course, SpringSpring2002 2002 bybyTaiTaiSing LeeSing Lee
∑
∑
∑
∑
∑
+ +
+
−
− +
+
−
− +
−
=
+
+
j i
c j
i
j i j
i j
i j
i
j i j
i
j i j
i
j i j
i j
i
j i j
i j
i j i j
i
j i V h
v v
f f
h f
f d
f v
h f
E
, ,
, ,
, ,
2 ,
1 ,
, ,
2 ,
, 1 ,
2 ,
, ,
, ,
) , ( )
( )
1 ( ) (
) 1
( ) (
) (
) ,
, (
κ α
λ
constraint constraint
! ! V V
cc(i,j (i,j ) ) energy energy term term : :
! ! Low Low for for 0, 2 0, 2 lines lines
! ! Medium Medium for for 3 lines 3 lines
! ! High High for for 1, 4 1, 4 lines lines
Image
Image fromfromCMU 15-CMU 15-385 Computer 385 Computer VisionVisioncourse, course, SpringSpring2002 2002 bybyTaiTaiSing LeeSing Lee
the the Euler Euler equation equation
! ! The The extrema extrema of of a a function function f(x f(x ) ) are are attained attained where where f’ = 0
f’ = 0
! ! Similarly Similarly , , the the extrema extrema of of the the functional functional E(u E(u ) ) are are obtained
obtained where where E’ = 0 E’ = 0 . .
! ! E = ( E = ( ∂ ∂ E / E / ∂ ∂ u) u) is the is the first first variation variation . .
! ! Assuming Assuming a common a common formulation formulation where where u(x):[0,1] u(x):[0,1] # # R R , , u(0)=a
u(0)=a and and u(1)=b u(1)=b , ,
the the basic basic problem problem is is to to minimize minimize : :
! ! The The necessary necessary condition condition for for u u to to be be an an extremum extremum of of E(u E(u ) ) is is the the
Euler
Euler equation equation
of of a a one one dimensional dimensional problem problem . .
∫
=
10
( , ' ) )
( u F u u dx
E
' = 0
∂
− ∂
∂
∂
u F dx
d u
F
Gradient
Gradient descent descent
! ! The The Euler Euler equation equation can can be solved be solved by by numerically numerically solving solving (1). (1).
! ! Can Can be be formulated formulated as as an an evolution evolution equation
equation (t ( t – – time time ): ):
!! ForFor exampleexample, , updatingupdating ff whilewhile ll is is fixed (
fixed (stringstring):):
Image
Image fromfromCMU 15CMU 15--385 Computer 385 Computer VisionVisioncourse, course, SpringSpring2002 2002 bybyTaiTaiSing Sing LeeLee
) 3 (
) 2 ( ) 1
)(
( 2 ) (
2
) 1 (
1
1
i t
i t
i i
i i
i i
i i
i i
f f E
f f
l f
f d
f f E
f E dt
df
∂
− ∂
=
−
=
∆
−
−
−
−
∂ =
∂ ∂
− ∂
=
+
+
µ λ
Works
Works only only for for convex convex functions
functions ! !
Simulated
Simulated Annealing Annealing example example
! ! Works Works for for non non - - convex convex energy
energy functions functions
! ! λ λ =6 =6 , , α α =0.04 =0.04
Images
ImagesfromfromCMU 15-CMU 15-385 Computer 385 Computer VisionVisioncourse, course, SpringSpring 2002
2002 bybyTaiTaiSing LeeSing Lee
Line process (l) Surface signal (f)
Graduated
Graduated non non - - convexity convexity
! ! Proposed Proposed by by Blake Blake & & Zisserman Zisserman for for the the weak weak membrane’s membrane’s energy energy
! ! Basic idea: Basic idea:
1. 1. Approximate Approximate the the original original energy energy functional functional by by a a convex convex one one
2. 2. Do Do a a gradient gradient descent descent on on the the approximation approximation 3. 3. Gradually Gradually morph morph back back the the approximation approximation into into
the the original original energy energy while while repeating repeating step step 2. 2.
! ! In In case case of of a a weak weak membrane membrane energy energy , , the the morphing
morphing can can be be parametrized parametrized ! !
Graduated
Graduated non non - - convexity convexity
! ! GNC GNC runs runs downhill downhill on on each each of of a a sequence sequence of of
functions functions
! ! It It reaches reaches a a global global optimum
optimum assuming assuming a a seuence
seuence of of
approximating
approximating and and locally
locally convex convex funtions funtions exist
exist . .
Image
Image fromfromCMU 15-CMU 15-385 Computer 385 Computer VisionVisioncoursecourse, , SpringSpring2002 by2002 byTaiTaiSing LeeSing Lee
the the weak weak membrane membrane energy energy
! ! F F (0) (0) =E =E the the original original functional
functional , ,
! ! F F (1) (1) =F* =F* the the convex convex approximation
approximation
∑
∑ − + −
−=
i
i i
p i
i i
p
f d g f f
F
( )( )
2 ( )(
1)
≥
<
≤
−
−
<
=
r t
if
r t
q if
r t
c
q t
if t
t g
p|
|
|
| 2
/ )
| (|
|
| )
(
2 22 2 )
(
α α λ
q r r c
membrane string
c p
c c
2
2
2 1
4 / 1
2 /
* 1 ,
/
*
λ α λ =
+
=
=
=
Contours
Contours are are defined
defined as as
the the set set of of i i for for which
which
|f |f
ii- - f f
ii-1-1|>0 |>0
the the weak weak membrane membrane energy energy
! ! It It is is shown shown that that F F
(p)(p)is is convex
convex for for p p ≥ ≥ 1 1
! ! F F
(1)(1)can can be be minimized minimized using
using gradient gradient descent descent
! ! As As p p # # 0 0
! ! Increased Increased localization localization of of boundaries boundaries ( ( l l ) )
! ! Gradual Gradual anisotropic anisotropic smoothing
smoothing of of surface surface ( ( f f ) )
! ! Parameters Parameters used used for for the the test: test: λ λ =6 =6 , , α α =0.03 =0.03
Image
Image fromfromCMU 15CMU 15--385 Computer 385 Computer VisionVisioncourse, course, SpringSpring2002 by2002 byTaiTaiSing Sing LeeLee
ff ll
equivalence equivalence
! ! Formal Formal equivalence equivalence between between the the two two approaches
approaches . . For For example example : :
! ! Taking Taking exponential exponential (~ (~ Hammersley Hammersley - - Clifford Clifford ) )
! ! T T ~ ~ uncertainty uncertainty („ („ temperature temperature ”) ”)
∑
∑
∑ − + − − +
=
+i
i i
i i
i i
i
i
d f f l l
f f
E ( ) ( )
2λ (
1)
2( 1 ) α
4 4 4
4 3 4
4 4
4 2 1
4 43 4
42
1
smoothnessprior (MRF)/ ) )
1 ( ) (
(
(Gaussian) term
data
/ ) / (
)
(
∏
− − 2∏
− 1− 2 − +−
=
+i
T l
l f
f i
T d
T f f
E i i i i i i