c t r a l methods Alexander amp Michael Bronstein 20062009 Michael Bronstein 2010 toscacstechnionacilbook 048921 Advanced topics in vision Processing and Analysis of Geometric Shapes ID: 323317
Download Presentation The PPT/PDF document "S p e" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
S
pectral methods
© Alexander & Michael Bronstein, 2006-2009
© Michael Bronstein, 2010tosca.cs.technion.ac.il/book
048921 Advanced topics in vision
Processing and Analysis of Geometric Shapes
EE
Technion
, Spring 2010Slide2
A mathematical exercise
Assume points with the metric are isometrically embeddable into
Then, there exists a canonical form such that
for all
We can also writeSlide3
A mathematical exercise
Since the canonical form is defined up to
isometry, we can arbitrarily setSlide4
A mathematical exercise
Conclusion: if points are isometrically embeddable into then
Element of
a matrix
Element of an
matrix
Note:
can be defined in different ways!Slide5
Gram matrices
A matrix of inner products of the formis called a Gram matrix
Jørgen Pedersen Gram(1850-1916)
Properties:
(
positive semidefinite
)
Slide6
Back to our problem…
Isaac Schoenberg(1903-1990)
[Schoenberg, 1935]:
Points with the metric can be isometrically embedded into a Euclidean space if and only if
If points with the metric
can be isometrically embedded into , then
can be realized as a Gram matrix of rank ,
which is positive semidefinite
A positive semidefinite matrix of rank
can be written as
giving the canonical formSlide7
Classic MDS
Usually, a shape is not isometrically embeddable into a Eucludean space, implying that (has negative eignevalues)
We can
approximate by a Gram matrix of rank
Keep
m
largest eignevalues
Canonical form computed as
Method known as
classic MDS
(or
classical scaling
)Slide8
Properties of classic MDS
Nested dimensions: the first dimensions of an -dimensionalcanonical form are equal to an -dimensional canonical form
Global optimization problem
– no local convergenceRequires computing a few largest eigenvalues of a real symmetric matrix, which can be efficiently solved numerically (e.g. Arnoldi and Lanczos)
The error introduced by taking instead of can be quantified as
Classic MDS minimizes the
strainSlide9
MATLAB® intermezzo
Classic MDSCanonical formsSlide10
Classical scaling example
1
B
D
A
C
1
1
1
1
B
A
C
2
A
A
1
B
C
D
B
C
D
2
1
1
1
1
2
1
1
1
1
1
D
1Slide11
Deformation
Deformation
+Topology
Topological invarianceSlide12
Local methods
Make the embedding preserve
local
properties of the shape
If , then is small. We want the corresponding distance in the embedding space to be small
Map neighboring points to neighboring pointsSlide13
Local methods
Think globally, act locally
Local criterion how far apart the embedding takes neighboring points
“
”
David Brower
Global criterion
whereSlide14
Laplacian matrix
where is an matrix with elementsMatrix formulation
Recall stress derivation
in LS-MDS
is called the
Laplacian matrix
has zero eigenvalueSlide15
Local methods
Compute canonical form by solving the optimization problem
Trivial solution ( ): points can collapse to a single point
Introduce a constraint avoiding trivial solutionSlide16
Minimum eigenvalue problems
Lets look at a simplified case: one-dimensional embedding
Geometric intuition:
find a unit vector shortened the most by the action of the matrix
Express the problem using eigendecompositionSlide17
Solution of the problem
is given as the
smallest non-trivial eigenvectors of
The smallest eigenvalue is
zero
and the corresponding eigenvector is constant (collapsing to a point)
Minimum eigenvalue problemsSlide18
Laplacian eigenmaps
Compute the canonical form by finding the smallest non-trivial eigenvectors of
Method called
Laplacian eigenmap [Belkin&Niyogi]
is sparse (computational advantage for eigendecomposition)
We need the lower part of the spectrum of
Nested dimensions like in classic MDSSlide19
Laplacian eigenmaps example
Classic MDS
Laplacian eigenmapSlide20
Continuous case
Consider a one-dimensional embedding (due to nested dimension property, each dimension can be considered separately)
We were trying to find a map that maps neighboring points to neighboring points
In the continuous case, we have a smooth map on surface
Let be a point on and be a point obtained by an infinitesimal displacement from by a vector in the tangent plane
By Taylor expansion,
Inner product on tangent space (metric tensor)Slide21
Continuous case
By the Cauchy-Schwarz inequality
implying that is small if is small: i.e., points close to are mapped close to
Continuous local criterion:
Continuous global criterion:Slide22
Continuous analog of Laplacian eigenmaps
Canonical form computed as the minimization problem
where:
Stokes theorem
We can rewrite
is the space of
square-integrable
functions onSlide23
Laplace-Beltrami operator
The operator is called Laplace-Beltrami operator
Laplace-Beltrami operator is a generalization of Laplacian to manifolds
In the Euclidean plane,
Intrinsic
property of the shape (invariant to isometries)
Note:
we define Laplace-Beltrami operator with minus, unlike many books
In coordinate notation Slide24
Laplace-Beltrami
Pierre Simon de Laplace (1749-1827)
Eugenio Beltrami
(1835-1899)Slide25
Properties of Laplace-Beltrami operator
Let be smooth functions on the surface . Then the Laplace-Beltrami operator has the following properties
Constant eigenfunction: for any
Symmetry:Locality: is independent of for any pointsEuclidean case: if is Euclidean plane and
then
Positive semidefinite:Slide26
Continuous vs discrete problem
Continuous:
Discrete:
Laplace-Beltrami operator
LaplacianSlide27
To see the sound
Chladni’s experimental setup allowing to visualize acoustic wavesErnst Chladni ['kladnɪ]
(1715-1782)
E. Chladni, Entdeckungen über die Theorie des KlangesSlide28
Chladni plates
Patterns seen by Chladni are solutions to stationary Helmholtz equation
Solutions of this equation are
eigenfunction of Laplace-Beltrami operatorSlide29
The first eigenfunctions of the Laplace-Beltrami operator
Laplace-Beltrami operatorSlide30
Laplace-Beltrami eigenfunctions
An eigenfunction of the Laplace-Beltrami operator computed on different deformations of the shape, showing the invariance of theLaplace-Beltrami operator to isometriesSlide31
Laplace-Beltrami spectrum
Eigendecomposition of Laplace-Beltrami operator of a compact shape gives a discrete set of eigenvalues and eigenfunctions
The eigenvalues and eigenfunctions are isometry invariant
Since the Laplace-Beltrami operator is symmetric, eigenfunctions
form an
orthogonal basis
for Slide32
Shape DNA
[Reuter et al. 2006]: use the Laplace-Beltrami spectrum as an isometry-invariant shape descriptor (“shape DNA”)
Laplace-Beltrami spectrum
Images: Reuter
et al.Slide33
Shape DNA
Shape similarity using Laplace-Beltrami spectrum
Images: Reuter
et al.Slide34
Uniqueness of representation
ISOMETRIC SHAPES ARE ISOSPECTRAL
ARE ISOSPECTRAL SHAPES ISOMETRIC?Slide35
Mark Kac
(1914-1984)
Can one hear the shape of the drum?
“”
More prosaically: can one reconstruct the shape
(up to an isometry) from its Laplace-Beltrami spectrum?Slide36
To hear the shape
In Chladni’s experiments, the spectrum describes acoustic characteristics of the plates (“modes” of vibrations)
What can be “heard” from the spectrum:
Total Gaussian curvatureEuler characteristicArea
Can we “hear” the metric?Slide37
One cannot hear the shape of the drum!
[Gordon et al. 1991]:Counter-example of isospectral but not isometric shapesSlide38
Discrete Laplace-Beltrami operator
Let the surface be sampled at points and represented as a triangular mesh , and let
Discrete version of the Laplace-Beltrami operator
In matrix notation
whereSlide39
Find the discrete eigenfunctions of the Laplace-Beltrami operator by solving the
generalized eigenvalue problem
Levy 2006Reuter, Biasotti, Giorgi, Patane & Spagnuolo 2009
where is an matrix whose columns are the eigenfunctions
is a diagonal matrix of corresponding eigenvalues
Discrete Laplace-Beltrami eigenfunctionsSlide40
Discrete vs discretized
Continuous surface
Laplace-Beltrami operator
Discretize the surfaceas a graph
Discretize Laplace-Beltrami
operator, preserving some
of the continuous properties
Graph Laplacian
Eigendecomposition
Continuous eigenfunctions and eigenvalues
Eigendecomposition
Discretize
eigenfunctions
and eigenvalues
“Discrete Laplacian”
“Discretized Laplacian”
FEMSlide41
1. Tutte 1963; Zhang 2004
2. Pinkall 1993; Meyer 2003
Cotangent weight
2
sum of areas of triangles
sharing vertex
Discrete Laplacian
1
(umbrella operator); or
valence of vertex (Tutte)
Discrete Laplace-Beltrami operatorSlide42
Properties of discrete Laplace-Beltrami operator
The discrete analog of the properties of the continuous Laplace-Betrami operator is
Symmetry:Locality: if are not directly connected
Euclidean case: if is Euclidean plane,Positive semidefinite:
In order for the discretization to be consistent,
Convergence:
solution of discrete PDE with converges to the solution
of continuous PDE with for Slide43
No free lunch
Laplacian matrix we used in Laplacian eigenmaps does not converge to the continuous Laplace-Beltrami operatorThere exist many other approximations of the Laplace-Beltrami operator, satisfying different properties
[Wardetzky
et al. 2007]: there is no discretization of the Laplace-Beltrami operator satisfying simultaneously all the desired propertiesSlide44
Finite elements method
Reuter, Biasotti, Giorgi, Patane & Spagnuolo 2009
Eigendecomposition problem in the weak form
for any smooth
Given a finite basis spanning a subspace of
can be expanded as
Write a system of equation
posed as a generalized eigenvalue problem