Metrics of curves in shape optimization and

analysis
A. Mennucci1
Lectures at Scuola Normale Superiore,
Feb 2011

Contents

I The matter at hand 4

1 Shapes & Curves 4
1.1 Shapes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Curve–related quantities . . . . . . . . . . . . . . . . . . . 5
1.2.2 Curvature(s) . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Gâteaux differentials in the space of immersed curves . . 7

2 Shapes in applications 9
2.1 Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Shape Optimization & Active Contours . . . . . . . . . . . . . . 10
2.3 Geometric Active Contour Method . . . . . . . . . . . . . . . . . 12
2.4 Shape Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Problems & Goals . . . . . . . . . . . . . . . . . . . . . . . . . . 16

II Basic notions, Deeper review 21

3 Basic mathematical notions 21
3.1 Distance and metric space . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Banach, Hilbert and Fréchet spaces . . . . . . . . . . . . . . . . . 22
3.3 Manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Fréchet Manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Riemann & Finsler Geometry . . . . . . . . . . . . . . . . . . . . 26
3.6 Drawbacks in infinite dimensions . . . . . . . . . . . . . . . . . . 28
3.7 Gradient in abstract differentiable manifold . . . . . . . . . . . . 29
3.8 Group actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4 Spaces and metrics of curves 31

4.1 Classes of curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.1 Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2 Group actions on curves. . . . . . . . . . . . . . . . . . . . . . . . 32

1
 4.3 Geometric curves, and functionals . . . . . . . . . . . . . . . . . 33
4.4 The two categories of “Shape Spaces” . . . . . . . . . . . . . . . 34
4.5 Geometric curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.5.1 Bi,f is a Fréchet manifold . . . . . . . . . . . . . . . . . . 35
4.6 Goals (revisited) . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

III Literature Overview 37

5 Representation/embedding/quotient 38

6 Metrics of sets 39
6.1 Some more maths on distance and geodesics . . . . . . . . . . . . 39
6.2 Hausdorff metric . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.3 A Hausdorff like distance of compact sets . . . . . . . . . . . . . 42

7 Finsler geometries of curves 46

7.1 L∞ metric and Fréchet distance . . . . . . . . . . . . . . . . . . . 47
7.2 L1 metric and Plateau problem . . . . . . . . . . . . . . . . . . . 47

8 Riemannian metrics of curves 48

8.1 Curves up to pose, as a submanifold of a Hilbert space . . . . . 48
8.1.1 Shape Representation using Direction Functions [Klassen
et al., 2004] . . . . . . . . . . . . . . . . . . . . . . . . . 48
8.1.2 A metric with explicit geodesics [Younes et al., 2008] . . . 52
8.2 Riemannian spaces of immersed curves . . . . . . . . . . . . . . . 54
8.2.1 H 0 (traditional) . . . . . . . . . . . . . . . . . . . . . . . 54
8.2.2 H A (Michor-Mumford) . . . . . . . . . . . . . . . . . . . 55
8.2.3 Conformal metrics . . . . . . . . . . . . . . . . . . . . . . 55
8.3 Sobolev type Riemannian metrics . . . . . . . . . . . . . . . . . . 56
8.3.1 Properties of H j metrics . . . . . . . . . . . . . . . . . . . 56
8.3.2 Proof of local ∃! of flows, (prop. 8.12) . . . . . . . . . . . 59
8.3.3 Mathematical properties . . . . . . . . . . . . . . . . . . . 61
8.4 A new Sobolev-type Riemannian metric . . . . . . . . . . . . . . 62
8.5 Some final remark (trade-off; quotient order) . . . . . . . . . . . 65

IV Sobolev Active Contours 65

9 Sobolev Active Contours 66
9.1 Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
9.2 Why H̃ j is better than H j . . . . . . . . . . . . . . . . . . . . . . 67
9.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

10 New Applications With Sobolev Active Contours 70

10.1 A Few Useful Energies Precluded By H 0 . . . . . . . . . . . . . . 70
10.2 Segmentation Applications for New Energies . . . . . . . . . . . . 71

11 A new Sobolev-type Riemannian metric 73

11.1 Dynamical models . . . . . . . . . . . . . . . . . . . . . . . . . . 73

2
V Riemannian manifolds of curves 75
12 Horizontality 77
12.1 Horizontal and vertical space . . . . . . . . . . . . . . . . . . . . 77
12.2 From curve-wise p.i to homotopy-wise p.i . . . . . . . . . . . . . 79

13 Momenta 81

14 Existence of Geodesics 83
14.1 The H 0 distance is degenerate . . . . . . . . . . . . . . . . . . . 83
14.2 (Critical) geodesics for H j . . . . . . . . . . . . . . . . . . . . . . 84
14.3 The direct method . . . . . . . . . . . . . . . . . . . . . . . . . . 84
14.4 Winding and pulley . . . . . . . . . . . . . . . . . . . . . . . . . 85
14.5 Curvature(s) as measures . . . . . . . . . . . . . . . . . . . . . . 86
14.6 Existence of geodesics for H 0 . . . . . . . . . . . . . . . . . . . . 86

3
 Start of lecture: I. The matter at hand

Part I
The matter at hand
1 Shapes & Curves
What is this course about? We begin by summarizing in a simpler form the
definitions, review the goals, and infer what are the needed mathematical tools.

1.1 Shapes
1. Examples
Some examples of “Shape Spaces”.
• The family of all collections of k points in lRn .
• The family of all non empty compact subsets of lRn .
• The family of all closed curves in lRn .
This is the shape space that we will study in the main part of these
lectures.
• Other: surfaces, measures, metric spaces, etc.

2. Shape as a variable
In all of the above models, shapes are collected in a Shape Space M , where
the shape is a point, i.e. a variable x ∈ M .
A tentative and vague list of operations.
calculation/calculus → math. analysis , differential geometry
nearness → topology, distance
optimization → analysis, topology, differential geometry...
categorization →...distance...
compression (reduce
complexity/dimensionality) → ...quantization...PCA...
estimation of unobservables → probability/statistics
3. Computer vision
A more precise list of applications of Shape Theory in Computer Vision.
There are two different (but interconnected) fields.
Shape Optimization where we want to find the shape that best satisfies a
design goal; a topic of interest in Engineering at large;
e.g. segmentation, tracking
Shape Analysis where we study a family of Shapes for purposes of statistics,
(automatic) cataloging, probabilistic modeling, among others, and possi-
bly create an a-priori model for a better Shape Optimization.

4
1.2 Curves
4. Curves
(Most of the theory following will be developed for curves in lRn , when this does
not complicate the math).
S 1 = {x ∈ lR2 | |x| = 1} is the circle in the plane. It is the template for all
possible (closed) curves. (Open curves will be called paths).
A C 1 curve is a continuously differentiable map c : S 1 → lRn such that
0
c (θ) := ∂θ c(θ) exists at all points θ ∈ S 1 .
Definition 1.1. An immersed curve is a C 1 curve c such that c0 (θ) 6= 0 at
all points θ ∈ S 1 .
: S1 →
c c(S 1 )
7→

c(S 1 ) is the image, or trace, of the curve.

5. Manifold of curves
For the purposes of this introduction, we define the manifold of (paramet-
ric) curves M to be the set of all closed immersed curves. It is an open subset
of all C 1 curves (so it is manifold). Let’s see what this means, practically.
Suppose that c : S 1 → lRn is a closed immersed curve.
A deformation of c is a function h : S 1 → lRn .
The set of all such h is the tangent space Tc M of M at c.
An infinitesimal deformation of the curve c0 in “direction” h will yield (on
first order) the curve c0 (u) + εh(u).
A homotopy C connecting c0 to c1 is a continuous function
h C : [0, 1] × S 1 → S 1 such that c0 (θ) = C(0, θ) and c1 (θ) =
c1 C(1, θ).
M c0 By defining γ(t) = C(t, ·) we can associate C to a path
γ : [0, 1] → M in the space of curves M , connecting c0
to c1 .
When dealing with homotopies C = C(t, θ), we will use the “prime notation”
C 0 for the derivative ∂θ C in the “curve parameter” θ, and the “dot notation” Ċ
for the derivative ∂t C in the “time parameter” t.

1.2.1 Curve–related quantities

8. Towards our goal

A good way to specify the design goal for Shape Optimization is to define an
objective function (a.k.a. energy) F : M → lR that is minimum in the curve
that is most fit for the task.
When designing our F , we will want it to be geometrical, that is, indepen-
dent of reparameterizations; this is easily accomplished if we use geometrical
quantities to start from. We now list the most important such quantities.
9. Derivations, Tangent
Definition 1.2. If the curve c is immersed, we can define the derivation with
respect to the arc parameter
1
∂s = ∂θ .
|c0 |

5
(We will sometimes also write Ds instead of ∂s .)
Definition 1.3. At all points where c0 (θ) 6= 0, we define the tangent vector

c0 (θ)
T (θ) = = ∂s c(θ) .
|c0 (θ)|

(At the points where c0 = 0 we may define T = 0).

10. Length
Definition 1.4. The length of the curve c is
Z
len(c) := |c0 (θ)| dθ . (1)
S1

11. Integration by arc parameter

We can define formally the arc parameter s by

ds := |c0 (θ)| dθ ;

we use it only in integration, as follows.

Definition 1.5. We define the integration by arc-parameter of a function
g : S 1 → lRk along the curve c by
Z Z
g(s) ds := g(θ)|c0 (θ)| dθ .
c S1

and the average integral

Z Z
1
g(s) ds := g(s) ds
c len(c) c

and we will sometimes shorten this as avgc (g).

1.2.2 Curvature(s)
There are two different definitions of curvature of an immersed curve: mean
curvature H and signed curvature κ, which is defined when c is valued in lR2 .
H and k are extrinsic curvatures, they are properties of the embedding of c
into lRn .
13. Curvature H
Definition 1.6 (H). If c is C 2 regular and immersed, we can define the (mean)
curvature H of c as
H = ∂s ∂s c = ∂s T
1
where ∂s = |c0 | ∂θ is the derivation with respect to the arc parameter.

Properties. • It is easy to prove that H ⊥ T .

• H is a geometrical quantity. If c̃ = c ◦ φ and H̃ is its curvature, then

H̃ = H ◦ φ.

6
 • 1/|H| is the radious of a tangent circle that best approximates the curve
to second order.

14. Curvature κ
Let c be an immersed curve.
Definition 1.7 (N). When the curve c is in lR2 we can define a normal vector
N to the curve, by requiring that |N | = 1, N ⊥ T and N is rotated π/2 degree
anticlockwise with respect to T .
Definition 1.8 (κ). If c is in lR2 and C 2 , then we can define a signed scalar
curvature κ = hH, N i, so that

∂s T = κN = H and ∂s N = −κT .

κ>0

H N

N
H

H
N
κ<0

Figure 1: curvature

1.2.3 Gâteaux differentials in the space of immersed curves

Let M be the family of all immersed curves.
We will be interested in the Gâteaux derivation
E(c + th) − E(c)
DE(c; h) := lim = [∂t E(c + th)]|t=0 . (2)
t→0 t

of operators O and functions E : M → lRk .

By operator we mean any one of these possible operations.
• O : M → Tc M
O(c) is a smooth a vector field; examples are the operations of computing
the tangent field T , and the curvature H.
• O:M →M
O(c) is an immersed curve; examples are all the usual actions such as
translations, rotations, etc. . .
• O : M → (Tc M → Tc M )
O(c) is itself an operator on vector fields along c; an example is the deriva-
tive Ds .

7
 In all above cases, the evaluated operator O(c) is itself a function; for this
reason, to avoid confusion in the notation, in this proposition we write the
Gâteaux derivation as Dc,h O instead of DO(c; h). These rules following are
the building blocks that are used (by means of standard calculus rules for deriva-
tives) to compute the derivation of all other geometric operators usualy found
in computer vision applications.
Proposition 1.9.

Dc,h (Ds O) = −(Ds h · Ds c)(Ds O) + Ds (Dc,h O) , (3)

Z Z
Dc,h O ds = Dc,h O + O. (Ds h · Ds c) ds , (4)
Zc Zc
Dc,h O ds = Dc,h O + O.(Ds h · Ds c) ds −
c
Zc Z
− O ds (Ds h · Ds c) ds (5)
c c

The proof is by direct computation.

The above also holds for homotopies when substituting formally h = ∂t .
A consequence is that
Z Z
Dc,h len(c) = (Ds h · Ds c) ds = − (h · Ds2 c) ds , (6)
c c

18. Gâteaux differential of the elastica

We propose an example of application of the above proposition We can
Gâteaux-derive the “second arc parameter derivative operator Dss ” by repeated
applications of (3), as follows

Dc,h (Dss g) = −(Ds h · Ds c)Dss g + Ds [Dc,h (Ds g)] =

= −(Ds h · Ds c)Dss g + Ds [−(Ds h · Ds c)(Ds g) + Ds (Dc,h g)] =
= Dss (Dc,h g) − 2(Ds h · Ds c)Dss g − [Ds (Ds h · Ds c)](Ds g)

and hence obtain the Gâteaux differential of the curvature

Dc,h Dss c = Dss h − 2(Ds h · Ds c)Dss c − [Ds (Ds h · Ds c)]Ds c

and eventually obtain the Gâteaux differential of the elastica energy by

substituting this last equality in (4)
Z Z
2
Dc,h |Dss c| ds = 2Dss c · (Dc,h Dss c) + |Dss c|2 . (Ds h · Ds c) ds =
c c
Z
= 2Dss c · Dss h − 3(Ds h · Ds c)|Dss c|2 ds
c

since (Dss c · Ds c) = 0.

8
2 Shapes in applications
A number of methods have been proposed in Shape Analysis to define distances
between shapes, averages of shapes and optimal morphings between shapes. At
the same time, there has been much previous work in Shape Optimization, for
example Image Segmentation via Active Contours, 3D Stereo Reconstruction via
Deformable Surfaces; in these later methods, many authors have defined Energy
Functionals E(c) on curves (or on surfaces), whose minima represent the de-
sired Segmentation/Reconstruction; and then utilized the Calculus of Variations
to derive curve evolutions to search minima of E(c), often referring to these
evolutions as Gradient Flows. The reference to these flows as gradient flows
implies a certain Riemannian metric on the space of curves; but this fact has
been largely overlooked. We call this metric H 0 , and define it by
Z
hh, kiH 0 := hh(s), k(s)i ds
S1

where h, k ∈ Tc M , ds := |c0 (θ)| dθ is integration by arc-parameter, and hh(s), k(s)i

is the usual Euclidean scalar product in lRn (that sometimes we will also write
as h(s) · k(s)).

2.1 Images
20. Cosa è un’immagine

Prendiamo una
immagine:

Canaletto, Campo di Rialto

9
Ingrandiamo molto un dettaglio: notiamo
dei “quadratini”, che sono detti “pixel”.
Ogni pixel P è identificato da un’ascissa i
e un’ordinata j nel piano dello schermo.

Ogni pixel ha un colore, che è dato da una combinazione addittiva

di rosso, verde e blu: questo modello del colore è indicato dalla sigla
RGB (dall’inglese Red Green Blue).
La intensità di ognuno di questi 3 colori è convenzionalmente indicata da un
numero intero, che assume un valore compreso fra 0 e 255.
Un’immagine del computer di dimensioni l × a è un vettore in lR3la .
azzurro bianco

verde
Tutti i colori che si possono visualizzare G giallo
sono contenuti in un cubo. Il vertice
(0, 0, 0) rappresenta il nero, il (255, 255, 255) grigio
blu
il bianco. B magenta

Esistono anche altre scelte di coordinate per Cb

i colori.
nero rosso
R
Cr
Una molto diffusa si ottiene definendo Y, Cr , Cb come

 Y = R+G+B

3  R = Y + 23 Cr
C =R− 2 G+B
G = Y − 13 Cr − 12 Cb
 r
Cb = B − G B = Y − 13 Cr + 12 Cb


I tre nuovi assi Y, Cr , Cb sono ortogonali (ma non ortonormali).

Questa è una versione semplificata della definizione ufficiale; si veda in http://en.wikipedia.org/wiki/YCbCr
Nella trattazione matematica astratta che segue, supporremo che un’immagine
sia una funzione I : Ω → lR3 (se immagine a colori) oppure I : Ω → lR (se a
scala di grigi). Usualmente si sceglie Ω = [0, 1]2

2.2 Shape Optimization & Active Contours

25. A short history of Active Contours

10
Active contours, introduced by [Kass et al., 1987], have been widely used for the
segmentation problem. The idea is to minimize an energy, defined on contours
or curves, that contains an image based edge attraction term and a smoothness
term, which becomes large when the curve is irregular. An evolution is derived to
minimize the energy based on principles from the calculus of variations. There
have been many variations to original model of [Kass et al., 1987]; for example
[Cohen, 1991], and a survey in [Blake and Isard, 1998]. An unjustified feature
of the model of [Kass et al., 1987] is that the evolution is dependent on the way
the contour is parameterized. Hence there have been geometric evolutions sim-
ilar to the idea of [Kass et al., 1987] in [Caselles et al., 1993; Malladi et al.,
1995], which can be implemented by level set methods [Osher and Sethian, 1988].
Thereafter, [Kichenassamy et al., 1995] [Caselles et al., 1995] considered min-
imizing a geometric energy, which is a generalization of Euclidean arclength,
defined on curves for the edge-detection problem. The authors derived the gra-
dient descent flow in order to minimize the geometric energy. In the works of
[Siddiqi et al., 1998] and [Xu and Prince, 1998], the authors build on the active
contour edge-detection approach to segmentation.
In contrast to the edge-based approaches for active contours (mentioned above),
region-based energies [Ronfard, 1994] [Zhu et al., 1995] [Yezzi et al., 1999]
[Chan and Vese, 2001] for active contours have provided many desirable fea-
tures; for example, they provide less sensitivity to noise, better ability to cap-
ture concavities of objects, more dependence on global features of the image,
and less sensitivity to the initial contour placement. In these approaches, an
energy is designed to be minimized when the curve partitions the image into
statistically distinct regions. In [Mumford and Shah, 1985, 1989], the authors
introduce and rigorously study a region-based energy that is designed to both
extract the boundary of distinct regions while also smoothing the image within
these regions. Subsequently, [Tsai et al., 2001b] [Vese and Chan, 2002] gave a
curve evolution implementation of minimizing the energy functional considered
by Mumford&Shah in a level set framework [Osher and Sethian, 1988]. Another
work that makes use of a region-based energy is [Paragios and Deriche, 2002a,b],
which also combines edge-based information. Like the edge-based methods, the
gradient descent flows are calculated to minimize these energies.
More recent works that use active contours for segmentation are not only
based on information from the image to be segmented (edge-based or region-
based), but also a-priori information known about the shape of the desired object
to be segmented. The work of [Leventon et al., 2000] showed how to incorporate
prior information into the active contour paradigm. Subsequently, there have
been a number of works, for example [Tsai et al., 2001a; Rousson and Paragios,
2002; Chen et al., 2002; Cremers and Soatto, 2003; Raviv et al., 2004], which
design energy functionals that incorporate prior shape information of the desired
object. In these works, the main idea is designing a prior term of the energy
that is small when the curve is close, in some sense, to a pre-specified shape.
27. Energies in Computer Vision
A variational approach to solving Computer Vision problems: minimizing
geometric energy functionals F (c) where c is a curve enclosing the region R.

11
 R2
• Boundary Based Energies:
Z
F (c) = g(c(s)) ds
c

ds=arclength element
c
Rc R • Region Based Energies:
Z Z
F (c) = fin (x) dx + fout (x) dx
R Rc

dx=area element
R and Rc are the interior and exterior areas.
28. Examples of geometric energy functionals for segmentation
I : Ω → [0, 1] is the image (in black & white).
• Chan-Vese
R2

Z 2
F (c) = I(x) − avgR (I) dx +
ZR 2
c
+ I(x) − avgRc (I) dx
Rc R
Rc
Z
1
where avgR I = I(x)dx is the average
|R| R
on the region.
• Geodesic “Edge detector”
Z
1
F (c) = g(c(s)) ds where g(x) =
c 1 + |∇I(x)|2

g may be interpreted as a conformal metric on lR2 .

2.3 Geometric Active Contour Method

29. Geometric Active Contour Paradigm
General Procedure for Geometric Active Contours:
1. Choose an appropriate geometric energy functional, E.
2. Compute the Gâteaux differential (i.e. directional derivative)
d
DE(c; h) := E(c + th)|t=0
dt
where c is a curve and h is an arbitrary perturbation of c.
3. Manipulate DE(c; h) into the form
Z
h · v(s) ds
c

12
 4. Consider v to be the gradient, the direction which increases E fastest.
5. Evolve ∂t c = −v; this is called the gradient descent flow.

30. Example: Geodesic Active Contour Edge Model

1. Start from an energy that is minimal along image contours:

Z
1
E(c) = g(c(s)) ds where g(x) := 2
c 1 + k∇I(x)k

2. Calculate the Directional Derivative:

Z
d


E c(t) = − ∂t c · gκN − (∇g · N )N ds
dt c

3. Deduce the Gradient:

∇E = − gκN − (∇g · N )N

(The part of the flow tangent to the curve is always zero – this will be
properly explained in the fourth lecture, in slide 247).
4. Write the gradient Flow:
∂c
= gκN − (∇g · N )N
∂t

(In the above, ∇I and ∇g are the gradients wrt x).

31. Implicit Assumption of H 0 Inner Product
We have made a critical assumption in going from the directional derivative
computation Z
DE(c; h) = h · v(s) ds
c

to deducing that the gradient of E is ∇E(c) = v. Namely, the definition of the

gradient is based on the following equality
Z
h∂t c, ∇EiH 0 = ∂t c · v(s) ds,
c |{z} |{z}
h1 =∂t c h2 =∇E

that needs an inner-product structure. (The precise definition of what the

gradient is is in frame 81 in section 3.7). This implies that we have been
presuming all along that curves are equipped with a H 0 -type inner-product
defined as follows D E Z
h1 , h2 0 := h1 · h2 ds
H c

(note the integration according to the arclength measure)

13
2.4 Shape Analysis
32. Shape Analysis.
Many method and tools comprise the shape analysis.

• Distances between shapes.

• Averages for shapes.
• Principle Component Analysis for shapes.
• Probabilistic models of shapes.

33. (Signed) Distance function

We begin by defining the (signed) distance function, a tool that we will use
often in this theory.
uA (x) := inf y∈A |x − y| is the distance function, A, B ⊂ lR2 compact.

uA

uB

x2

B
A
x1

bA (x) := uA (x) − ulR2 \A (x) is the signed distance function.

bA

bB

x2

B
A
x1

34. Shape Distances

A variety of distances d(c1 , c2 ) have been proposed for measuring the differ-
ence between two given shapes c1 and c2 .

Examples:

• Hausdorff Distance
   
dH (c1 , c2 ) := sup uc2 (x) ∨ sup uc1 (x)
x∈c1 x∈c2

where uA (x) := inf y∈A |x − y| is the distance function, and where c1 , c2

is either the boundary curve or the internal region;
• Set Symmetric Distance

d(c1 , c2 ) = c̊1 ∆c̊1 = |(c̊1 \ c̊2 ) ∪ (c̊2 \ c̊1 )|

(c̊1 ,c̊2 denote the internal region of c1 , c2 , | · | is the area.)

14
 35. Shape Averages
Many definitions have also been proposed for the average c̄ of a finite set of
shapes c1 , . . . , cN .

Arithmetic Mean (representation dependent)

• Parametric or Landmark Averaging

N
1 X
c̄(p) = cn (p), where p = common parameter/landmark
N n=1

• Signed Distance Level Set Averaging

N
 1 X
c̄ = x | b̄(x) = 0 , where b̄(x) = bc (p)
N n=1 n

(bcn = signed distance function of cn )

36. Shape Averages (cont.)

Distance Based Averaging (Fréchet, 1948; [Karcher, 1977]):
Given a distance dM between shapes, an average shape c̄ may be found by
minimizing the sum of its squared distances.
N
X
c̄ = arg min dM (c, cn )2
c
n=1

(coincides with the arithmetic mean in Euclidean spaces)

Definition 2.1. Suppose that X is a random vector taking values in lRn ; let
X = E(X) be the mean of X. The principal component analysis is the
representation of X as
Xn
X=X+ Yi Si
i=1

where Si are constant vectors, and Yi are uncorrelated real random variables
with zero mean, and with decreasing variance. Si is known as the i-th mode
of principal variation.
The PCA is possible in general in any finite dimensional linear space equipped
with an inner product. In infinite dimensional spaces, or equivalently in case of
random processes, the PCA is obtained by the Karhunen-Loève theorem.
38. Principal Component Analysis (PCA)
With “linear” shape representations, PCA has become popular for imposing
shape priors in various shape optimization problems.
N
X
c̄ + αn Sn
n=1

15
where

c̄ = mean shape
Sn = n’th mode of variation
αn ∈ lR

For more general manifold representations we may replace Sn by tangent

vectors and follow geodesics to perform the summation.
39. PCA by Signed Distance representation [Tsai et al., 2001a]
In this slide we see an example of synergy between analysis and optimization,
from [Tsai et al., 2001a]. The first row of pictures represent a PCA of a family of
plane shapes; the second row of images represent an image of a plane shape (a)
occluded (b) and noise degraded (c), an initialization for the shape optimizer
(d) and the final optimal segmentation (e).
Principal component analysis of signed distance function of plane shapes.

(a) (b) (c) (d) (e) (f) (g)

| {z } | {z } | {z } | {z }
mean first mode second mode third mode
Segmentation:

(a) (b) (c) (d) (e)

2.5 Problems & Goals

The need for prior term arose from several factors including the fact that some images
contain limited information, the energies functions considered previously could not
incorporate complex information, the energies had too many extraneous local minima,
and the gradient flows to minimize these energies allowed for arbitrary deformations
that gave rise to unlikely shapes.
Works on incorporating prior shape knowledge into active contours have led to a
fundamental question on how to define distance between two curves or shapes. Many
works, for example [Younes, 1998a; Soatto and Yezzi, 2002; Mio and Srivastava,
2004a; Charpiat et al., 2005; Michor and Mumford, 2003; Yezzi and Mennucci, 2005a],
in the shape analysis literature have proposed different ways of defining this distance.
However, [Michor and Mumford, 2003; Yezzi and Mennucci, 2005b] observed that all
previous works on geometric active contours that derive gradient flows to minimize
energies, which were described earlier, imply a natural notion of distance induced by
a Riemannian structure. Indeed, the notion of gradient of such energies relies on an
inner product defined the set of perturbations of a curve. All of these previous works

16
on geometric active contours use the same geometric L2 -type inner product, which we
refer to as H 0 , to define gradient. However, [Michor and Mumford, 2003; Yezzi and
Mennucci, 2005a] have shown a surprising property: the Riemannian metric on the
space of curves induced by the H 0 inner product is pathological, i.e., the “distance”
between any two curves is zero.
In addition to the pathologies of the Riemannian structure induced by H 0 , there
are also undesirable features of H 0 gradient flows as we will demonstrate in this paper.
Some of these features are listed below.
1. There are no regularity terms in the definition of the H 0 inner product. That
is, there is nothing in the definition of H 0 that discourages flows that are not
smooth in the space of curves. By smooth in the spaces of curves, we mean that
the surface formed by plotting the evolving curve as a function of time is smooth.
Thus, when energies are designed to depend on the image that is to be segmented,
the H 0 gradient is very sensitive to noise in the image. As a result, the curve
becomes non-smooth instantly. Therefore, in geometric active contours models,
a penalty on the curve’s length is added to keep the curve smooth in addition to
producing a variational problem that is well-posed. However, this changes the
energy that is being optimized. Moreover, as we shall demonstrate, the length
penalty is sometimes insufficient, and leads to other problems.
2. H 0 gradients, evaluated at a particular point on the curve, depend locally on
derivatives of the curve. Therefore, as the curve becomes non-smooth, as men-
tioned above, the derivative estimates become inaccurate, and thus, the curve
evolution becomes inaccurate. Moreover, for region-based and edge-based active
contours, the H 0 gradient at a particular point on the curve depends locally on
image data at the particular point. Although region-based energies may depend
on global statistics, such as mean values of regions, the H 0 gradient still depends
on local image data. These facts imply that the H 0 is sensitive to noise and local
features.
3. The H 0 norm gives non-preferential treatment to arbitrary deformations regard-
less of whether the deformations are global motions (not changing the shape of
the curve) such as translations, rotations and scales or whether they are more
local deformations. This implies, as we shall show, that the H 0 -gradient of
image dependent energies “encourages” points on the evolving curve to move
“independently” to decrease energy rather than encouraging the points to move
more collectively. This restricts the gradient at a particular point from “seeing”
information located at other points of the curve. Therefore, the curve evolving
according to the H 0 gradient flow is susceptible to local minima of the energy.
4. Many geometric active contours (such as edge and region-based active contours)
require that the unit normal to the evolving curve be defined. As such, the evolu-
tion does not make sense for polygons. Moreover, since in general, an H 0 active
contour does not remain smooth, one needs special numerical schemes based on
viscosity theory in a level set framework to define the flow.
5. If the energy depends on n derivatives of the curve, then the H 0 gradient has 2n
derivatives of the curve. Since the corresponding level set flows with higher than
two derivatives are not known to have a maximum principle, level set methods
[Osher and Sethian, 1988] are difficult to use for the numerical implementation
[Chopp and Sethian, 1999] Therefore, one may try to use particle methods to im-
plement the flow. However, flows with higher than two derivatives are generally
difficult to implement because of numerical artifacts.
6. Lastly, as a specific example, the H 0 gradient ascent for arclength, i.e., backward
heat flow, is ill-posed. This is quite odd in an intuitive manner because there
is nothing in the definition of length itself that indicates that a flow to increase

17
 length should be ill-posed. We shall see that this ill-posedness is a property of
H 0 not just the energy itself.

41. Troubles
As of year ∼2000, a number of troubles were affecting Shape Theory.
• The active contours method was not robust against noise; as in this ex-
ample of segmentation using Chan-Vese

where the flowing contour gets trapped into noise.

• Many (apparently) simple energies could not be implemented practically,
for three reasons
– some energies generate ill posed H 0 flows;
– if an energy integrand uses derivatives of the curve of degree up to d,
then the PDE driving the flow has degree 2d; but derivatives of the
curves of high degree are noise sensitive and are difficult to implement
numerically, and
– the active contours method works best when implemented using level
set methods, but this is difficult for flows PDEs of degree higher
than 2.
• The methods used in Shape Optimization and in Shape Analysis were too
unrelated; so it was not easy to couple them.

42. Example: geometric heat flow

Let T = ∂s c is the tangent to the curve, H = ∂s ∂s c its curvature. By direct
computation, the Gâteaux differential of the length of a closed curve is
Z Z
D len(c)(c; h) = h∂s h · T i ds = − hh · Hi ds (7)
S1 S1

Let C = C(θ, t) be an evolving family of curves. The geometric heat flow

(also known as motion by mean curvature) is
∂C
= H = ∂s ∂s C
∂t
In the H 0 inner-product, this is the gradient descent for curve length.
This flow has been studied deeply by the mathematical community.
Properties (See [Gage and Hamilton, 1986; Grayson, 1987]).

• Embedded curves remain embedded.

• Embedded planar curves converge to a circular point in finite time.
• Comparison Principle (important for level set methods) : if two embedded
curves are used as initialization, and the first contains the second, then it
will contain the second for all time of evolution.

18
 • The flow is well posed only for positive times , that is, for increasing t
(similarly to the usual heat equation).

43. Short length bias

The usual edge–based active contour image is of the form
Z
E(c) = g(c(s)) ds
c

supposing that g is reasonably constant in the region where c is running, E(c) '
g len(c), due to the integral being performed wrt the arc parameter ds. So one
way to reduce E is to shrink the curve. Consequently, when the image is smooth
and featureless (as in medical imaging), the usual edge based energies would
often drive the curve to a point. To avoid it, we may normalize the energy to
obtain a more length–independent energy
Z
1
avgc (g(c)) := g(c(s)) ds
len(c) c

but this generates ill–posed H 0 –flow.

44. Flow computations
Here we write L := len(c). Let g : lRn → lRk ; let
Z Z
1 1
avgc (g(c)) := g(c(s)) ds = g(c(θ))|c0 (θ)| dθ ;
L c L c
then the Gâteaux differential of E(c) := avgc (g(c)) is
Z
1
DE(c; h) = ∇g(c) · h + g(c)h∂s h · T i ds −
L c
Z Z
1
− 2 g(c) ds h∂s h · T i ds =
L c c
Z


= ∇g(c) · h + g(c) − avgc (g(c)) h∂s h · T i ds
c

In the above, we omit the argument (s) for brevity; ∇g is the gradient of g wrt
x. If the curve is in the plane, that is n = 2, then we define the normal vector
N ⊥ T by rotating T counterclockwise, and define scalar curvature κ so that
∂s ∂s c = κN ; then, integrating by parts, the above becomes
Z
1

DE(c; h) = ∇g(c) · h − ∇g(c) · T hh · T i −
L c
− g(c) − avgc (g(c)) hh · Ds2 ci ds =


Z 


= ∇g(c) · N − κ g(c) − avgc (g(c)) hh · N i ds
c

Suppose now that we have a Shape Optimization functional E including a

term of the form avgc (g(c)); let C = C(θ, t) be an evolving family of curves
trying to minimize E; this flow would contain a term of the form
∂C

= . . . g(c(s)) − avgc (g) κN . . .
∂t

19
unfortunately the above flow is ill defined: it is a backward heat flow on roughly
half of the curve. We will now propose two simple such examples.
45. Example: A weird H 0 Active Contours
Let c = avgc (c) be the center of mass of the curve.
Let us fix a target point v ∈ lR2 .
E(c) := 12 |c − v|2 penalizes the distance from the center of mass to v.
The H 0 gradient descent flow is
∂c
= −∇H 0 E(c) = h(v − c) · N iN − κN (c − c) · (v − c)
∂t
P
v
The first term in this gradient descent flow tries to move
the whole curve towards v.
Let P := {w : h(w−c)·(v−c) ≥ 0} be the halfplane “on
c the v side” . The second term in this gradient descent
c flow tries to decrease the curve length out of P and
increase the curve length in P : and this is ill posed.

46. Example: standard deviation of the curve

Similarly if g(x) = |x − c|2 , then E(c) := avgc (|c − c|2 ) is the standard
deviation of the curve. The derivative is
Z 
2

h(c − c) · N i − κ g(c) − avgc (g(c)) hh · N i ds
L c

The flow to minimize this should be

∂C

= −h(c − c) · N iN + κN g(c) − avgc (g(c))
∂t
and this
pis ill posed where the curve is inside of the circle of center avgc (c) and
radius avgc (g(c)).
47. Conclusions

• The above phenomena are practically visible in many applications, where

the Active Contour curve would “fractalize” in an attempt to minimize
the task energy. For this reason, a regularization term is often added to
the energy: this remedy, though, does change the energy, and ends up
solving a different problem.
• More recent works that use active contours for segmentation are not only
based on information from the image to be segmented (edge-based or
region-based), but also prior information, that is information known
about the shape of the desired object to be segmented.

The need for prior information terms arose from several factors such as

• the fact that some images contain limited information,

• the energies functions considered previously could not incorporate complex
information,

20
 • the energies had too many extraneous local minima, and the gradient
flows to minimize these energies allowed for arbitrary deformations that
gave rise to unlikely shapes; as in the example in the previous figure of
segmentation using the Chan-Vese energy, where the flowing contour gets
trapped into noise.
If one wishes to have a consistent view of the geometry of the space of curves
in both Shape Optimization and Shape Analysis, then one should use the H 0
metric when computing distances, averages and morphs between shapes.
Surprisingly, H 0 does not yield a well define metric structure, since the
associated distance is identically zero. (This striking fact was first described
in [Mumford]; it is generalized to spaces of submanifolds in [Michor and Mum-
ford, 2005]).
So to achieve our goal, we will need to devise a new metric.
Start of lecture: II. Basic notions, Deeper review

Part II
Basic notions, Deeper review
In this part we review all the program that was expressed in the first lecture,
providing the needed mathematical definitions in full detail.

3 Basic mathematical notions

3.1 Distance and metric space
51. Distance
Definition 3.1. Given a set M , a distance d = dM is a function d : M × M →
[0, ∞] such that
• d(x, x) = 0,
• if d(x, y) = 0 then x = y,
• d(x, y) = d(y, x) (d is symmetric)
• d(x, z) ≤ d(x, y) + d(y, z) (the triangular inequality).
The second request may be waived, in this case d is a semidistance.
The third request may be waived: then d would be an asymmetric metric.
Most theorems we will see can be adapted to this case.
52. Metric space
The pair (M, d) is a metric space. It has a topology, generated by balls
of the form B(x, r) := {y | d(x, y) < r}. According to this topology, xn → x iff
d(xn , x) → 0.
Definition 3.2. A metric space (M, d) is a complete iff, for all sequences
(cn ) ⊂ M ,
the fact that limm,n→∞ d(cm , cn ) = 0 implies that cn converges.

21
Example 3.3. • lRn has distance d(x, y) = |x − y|, and is complete.
• Any closed subset of a complete space is complete.
• If we cut away an accumulation point out of those spaces, the result is not
complete.

3.2 Banach, Hilbert and Fréchet spaces

53. Norm
Definition 3.4. Given a vector space E, a norm k · k is a function k · k : E →
[0, ∞]
• k · k is convex
• if kxk = 0 then x = 0.
• kλxk = |λ| kxk for λ ∈ lR
If the second request is waived, then it is a seminorm. If the third
request holds only for λ ≥ 0, then the norm is asymmetric; in this case, it
may happen that kxk = 6 k − xk. Each (semi/asymmetric)norm k · k defines a
(semi/asymmetric)distance

d(x, y) := kx − yk.

So a norm induces a topology.

54. Banach and Hilbert spaces

• A Banach space is a vector space E with a norm k · k defining a distance

d(x, y) := kx − yk such that E is metrically complete.
• A Hilbert space:
p a space with an inner product hf, gi, that defines a
norm kf k := hf, gi such that E is metrically complete. (It is also a
Banach space).

Let I ⊂ lRk be open; let p ∈ [1, ∞]. A standard example of Banach space is the
Lp space of functions f : I → lRn with norm
sZ
kf kLp := p
|f (x)|p dx for p ∈ [1, ∞) , kf kL∞ := sup |f (x)| ;
I I

if p = 2, L2 is a Hilbert space by inner product

Z
hf, gi := f (x) · g(x) dx .
I

Note that, in this spaces, f = g iff (by definition) the set {f 6= g} has measure
zero.

The following citations [Hamilton, 1982] are referred to the first part of
Hamilton’s 1982 survey on the Nash&Moser theorem.
55. Fréchet space [Hamilton, 1982]

22
 A locally-convex topological vector space (l.c.t.v.s.) E has a collection
of seminorms k · kk . The seminorms induce a topology, such that cn → c iff
kcn − ckk →n 0 for all k. All vector space operations are continuous w.r.t. this
topology.
A Fréchet space E is a complete Hausdorff metrizable locally-convex topo-
logical vector space; where
complete when, for any sequence (cn ), the fact that kcm − cn kk →m,n 0 for all
k implies that cn converges;
Hausdorff if kckk = 0 for all k then c = 0;
metrizable there are countably many seminorms associated to E.
The reason for the last definition is that, if E is metrizable, then we can define
a distance
∞
X 2−k kx − ykk
d(x, y) :=
1 + kx − ykk
k=0
that generates the same topology as the family of seminorms k · kk ; and the vice
versa is true as well, see [Rudin, 1973].
56. More examples of spaces of functions
Let I ⊂ lRk be open, or I be a finite dimensional manifold.
• The Banach spaces of C j functions; where
kf k = sup sup |f (k) (t)| ;
k≤j t

(k)
so cn → c when supt |cn (t) − c(k) (t)| →n 0 for all k ≤ j;
• A standard choice of Hilbert spaces of functions is provided by the Sobolev
spaces H j (I → lRn ), with scalar product
Z
hf, giH n := f (t) · g(t) + · · · + f (j) · g (j) dt .
c

• The Fréchet space of smooth functions C ∞ (I → lRn ).

The seminorms are kf kk = supx∈I |f (k) (x)| where f (k) is the k-th deriva-
(k)
tive. In this space, fn → f iff all the derivatives of fn converge uniformly
to derivatives f (k) .
This last is one of the strongest topology between the topologies usually asso-
ciated to spaces of functions.
57. Dual spaces.
Problem: the dual of a Fréchet space is not in general a Fréchet space (it
fails to be metrizable).
So given F, G Fréchet spaces, we cannot easily deal with “the space L(F, G)
of linear functions between F and G”.
As a workaround, given an auxiliary space H, we will consider “indexed
families of linear maps” L : H × F → G, where L(h; ·) is linear, and L is jointly
continuous; but we will not consider this as map
h 7→ (f 7→ L(h; f ))
(8)
H→ L(F, G)

23
 58. Example
One example is F = C 0 (S 1 ) and the map L : lR × F → F where L(t; f ) = g
with g defined as the translation g(x) = f (t + x) of f . This map is linear in the
second argument; it is easy to prove that it is jointly continuous; but the linear
operator map
t 7→ (f 7→ L(t; f ))
(9)
lR → L(F, F )
is not continuous, since the operator norm is

kL(t; ·) − L(s; ·)kL(F,F ) := sup kL(t; f ) − L(s; f )k =

kf k∞ ≤1

= sup sup |f (t + x) − f (s + x)| ≡ 2

kf k∞ ≤1 x

59. Derivatives
An example is the Gâteaux differential of map P : U ⊂ F → G, where F, G
are Fréchet spaces, U ⊂ F is open, and h ∈ F

P (f + th) − P (f )
DP (f ; h) := lim = [∂t P (f + th)]|t=0 . (10)
t→0 t
We say that P is C 1 if DP : U × F → G exists and is jointly continuous. (This
is weaker than what is usually required in Banach spaces).
The basics of the usual calculus hold.
Theorem 3.5 ([Hamilton, 1982, Thm. 3.2.5]). If P is C 1 then DP (f ; h) is
linear in h.
Theorem 3.6 (Chain rule [Hamilton, 1982, Thm. 3.3.4]). If P, Q are C 1 then
P ◦ Q is C 1 and


D(P ◦ Q)(f ; h) = DP Q(f ); DQ(f ; h) .

60. Troubles in calculus

But some important parts of calculus are instead lost.
Example 3.7 (See [Hamilton, 1982, Sec. 5.6]). Suppose P : U ⊂ F → F is
smooth, and consider the O.D.E.
d
f = Pf
dt
to be solved for a solution f : (−ε, ε) → F .

• if F is a Banach space, then, given initial condition f (0) = x, the solution

will exist (locally) and be unique;
• but if F is a Fréchet space, then f may fail to exist or be unique.

A consequence (that we will discuss more later on) is that the exponential
map in Riemannian manifolds may fail to be locally surjective. See [Hamilton,
1982, Sec. 5.5.2].

24
3.3 Manifold
To build a manifold of curves, we have ahead two main definitions of manifold
to choose from.
62. Differentiable Manifold.
An (abstract) differentiable manifold is a topological space M associ-
ated to a model l.c.t.v.s. E and equipped with an atlas of charts φk : Uk → Vk .

V2
φk are homeomorphisms , with c
V1
Uk ⊂ E open sets, and Vk ⊂ M
open sets that cover M . M
The composition φ−1 1 ◦φ2 restricted
to V1 ∩V2 will satisfy some require- φ1 φ2
ments, as e.g. being a diffeomor-
phism of class C r .

U U1 U2

The dimension dim(M ) of M is the dimension of E. When M is finite-

dimensional, the model space is always E = lRn .
63. (Sub)Manifold.
Let U be a fixed closed linear subspace of a l.c.t.v.s. X.
A (sub)manifold is a subset M of
X, such that, at all points c we may V1
c

find a chart φk : Uk → Vk , with M

Vk , Uk ⊂ X open sets, c ∈ Vk . φ1

X
φk are diffeomorphisms, and φk
maps U ∩ Uk onto M ∩ Vk . U1 U

M is itself an (abstract) manifold, and the model space is U .

In this case, dim(M ) = dim(U ) ≤ dim(X).
Most often, M = {Φ(c) = 0} where Φ : X → Y ; so, to prove that M is a
submanifold, we will use the implicit function theorem.
Viceversa any abstract manifold is a submanifold of some large X (by well
known embedding theorems).
64. Tangent space and tangent bundle
The tangent space Tc M is more eas-
ily described for submanifolds; in this Tc M
case, we choose a chart φk and a point
c
x ∈ Uk s.t. φk (x) = c; Tc M is the V1

image of the linear space U under the M

φ1
derivative Dx φk . Tc M is itself a linear
subspace in X. (We graphically rep- x

resent Tc M though as an affine sub- U1 U

space, by translating it to the point c.)

The tangent bundle T M is the collection of all tangent spaces.

25
3.4 Fréchet Manifold
When studying the space of all curces, we will deal with Fréchet manifolds,
where the model space E will be a Fréchet space, and the composition of local
charts φ−1
1 ◦ φ2 is smooth.
66. Examples of Fréchet manifolds
Given two finite-dimensional manifolds S, R, with S compact and n dimen-
sional,
• [Hamilton, 1982, Example 4.1.3]. the space C ∞ (S; R) of smooth maps
f : S → R is a Fréchet manifold. It is modeled on E = C ∞ (lRn ; R).
• [Hamilton, 1982, Example 4.4.6]. The space of diffeomorphisms Diff(S) of
S onto itself is a Fréchet manifold. The group operation φ, ψ → φ ◦ ψ is
smooth.
• But if we try to model Diff(S) on C k (S; S), then the group operation is
not even differentiable.

• [Hamilton, 1982, Example 4.6.6]. The quotient of the two above

C ∞ (S; R)/Diff(S)

is a Fréchet manifold. It contains “all smooth maps from S to R, up to

a diffeomorphism of S”. ( Does this sound familiar?)

We have found a math theory that seems apt to define and operate on the
manifold of geometric curves.

3.5 Riemann & Finsler Geometry

68. Riemann Metric, Length
Definition 3.8. • A Riemannian metric G on a differentiable manifold
M defines a scalar product hh1 , h2 iG|c on h1 , h2 ∈ Tc M
p
• The scalar product defines the norm |h| = |h|G|c = hh, hiG|c .
(All the above is generalized by Finsler Geometries.)
Suppose γ : [0, 1] → M is a path connecting c0 to c1 .
Z 1
• The length is Len(γ) := |γ̇(v)|γ(v) dv
γ c 0
1
M c0 Z 1
• The energy (or action) is E(γ) := |γ̇(v)|2γ(v) dv
0
69. Finsler Metric, Length
Definition 3.9. We define a Finsler metric to be a function F : T M → lR+ ,
such that
• F is continuous and,
• for all c ∈ M , v 7→ F (c, v) is a norm on Tc M .
We will sometimes write |v|c := F (c, v).

26
 (Sometimes F is called a “Minkowsky norm”).
As for the case of norms, a Finsler metric may be asymmetric; but, for sake
of simplicity, we will only consider the symmetric case.
Using the norm |v|c we can then again define the length of paths by
Z 1 Z 1
F
Len (γ) = |γ̇(t)|γ(t) dt = F (γ(t), γ̇(t)) dt
0 0

and similarly the action.

70. Distance & minimal geodesics
Definition 3.10. The distance d(c0 , c1 ) is the infimum of Len(γ) between all
γ connecting co , c1 ∈ M .
Definition 3.11. If there is a curve γ ∗ providing the minimum of Len(γ) be-
tween all curves connecting co , c1 ∈ M , then γ ∗ is called a (minimal) geodesic.
71. Properties of geodesics R1
The (minimal) geodesic is also the minimum of the action E(γ) = 0 |γ̇(v)|2 dv
(up to reparameterization).
Proposition 3.12. Let ξ ∗ provide the minimum minγ E(γ) in the class of all
paths γ in M connecting x to y. Then ξ ∗ is a (minimal) geodesic and its speed
|ξ˙∗ | is constant.
Viceversa, let γ ∗ be a minimal geodesic, then there is a reparameterization
ξ = γ ∗ ◦ φ s.t. ξ ∗ provides the minimum minγ E(γ).
∗

The action E is a smooth quadratic integral, and we can compute the Euler–
Lagrange equations; its minima are more regular, since they are guaranteed to
have constant speed; consequently, when trying to find geodesics, we will try to
minimize the action and not the lengh.
72. Exponential map
Definition 3.13 (exp map). Let γ̈ + Γ(γ̇, γ) = 0 be the Euler–Lagrange ODE
R1
of E(γ) = 0 |γ̇(v)|2 dv. Any solution of this ODE is a critical geodesic. (Note
that a minimal geodesic is a critical geodesic).
Define the exponential map expc : Tc M → M as expc (η) = γ(1), the
solution of
n
γ̈(v) + Γ(γ̇(v), γ(v)) = 0, γ(0) = c, γ̇(0) = η (11)

Solving the above ODE (11) is informally known as shooting geodesics.

73. Christoffel symbols
A curious remark.
Remark 3.14. In local coordinates we can write
Z 1
E(γ) = γ̇(v)i gi j (γ(v))γ̇(v)j dv
0

(assuming repeated indexes summation). If we compute the Euler-Lagrange

equation, and write it as
γ̈k + Γkij γ̇i γ̇j = 0
and rearrange the formula so that Γkij is symmetric in i, j, then we obtain that
Γkij are the Christoffel symbols associated to the metric gi j .

27
 74. exp as a local linear chart
Proposition 3.15. Suppose that M is a smooth differentiable manifold modeled
on a Hilbert space with a smooth Riemannian metric. The derivative of the
exponential map expc : Tc M → M at the origin is an isometry, hence expc is a
local diffeomorphism between a neighborhood of 0 ∈ Tc M and a neighborhood of
c ∈ M.
The exponential map can then “linearize” small portions of M , and so it
will enable us to use linear methods such as the principal component analysis.
Unfortunately, the above result does not hold if M is modeled on a Fréchet
space.
75. Hopf–Rinow theorem
Theorem 3.16 (Hopf–Rinow). Suppose M is finite dimensional Riemannian
or Finsler manifold. The following are equivalent
• (M, d) is complete ⇐⇒
• (11) can be solved for all c ∈ M and v ∈ lR ⇐⇒
• the map expc (η) is surjective;
and all those =⇒ ∀x, y ∈ M there exist a minimal geodesic connecting x to y.

3.6 Drawbacks in infinite dimensions

In a certain sense, infinite dimensional manifolds are simpler than their corre-
sponding finite-dimensional counterparts; indeed
Theorem 3.17 ([Eells and Elworthy, 1970]). Any smooth differentiable man-
ifold M modeled on an infinite dimensional separable Hilbert space H may be
embedded as an open subset of that Hilbert space.
In other words, it is always possible to express M using one single chart.
(But note that this may not be the best way for computations/applications).
77. no Hopf–Rinow theorem
When M is infinite dimensional Riemannian manifold, though, only a small
part of the Hopf–Rinow theorem still holds.
Proposition 3.18. Suppose M is infinite dimensional (model: Hilbert), and
(M, d) is complete, then the O.D.E. (11) of critical geodesics can be solved for
all v ∈ lR.
Example 3.19 ([Atkin, 1975]). There exists an infinite dimensional complete
Hilbert smooth manifold M and x, y ∈ M such that there is no critical geodesic
connecting x to y.
That is,
• (M, d) is complete 6⇒ expc is surjective,
• and (M, d) is complete 6⇒ minimal geodesics exist.

78. Grossman example

28
Example 3.20 ([Grossman, 1965] example (or, see sec. VIII.§6 in [Lang, 1999])).

e0
N Build an ellipsoid
( )
X
γ1 e1 M= x ∈ `2 | x20 + x2i /(1 + 1/i)2 = 1
e3 e2 i=1

γ2 in Hilbert space `2 (lN).

γ3 It is complete. It has radii
ρ0 = 1, ρ1 = 2, ρ2 = 3/2, ρ3 = 4/3, . . ..
Let N = e0 = (1, 0, 0, . . .), S = −e0 = (−1, 0, 0, . . .).
Let γi be the geodetic starting from γi (0) = N and with starting speed
γ̇i (0) = ei , such that γi (λi ) = S.
Then Len(γi ) → π, but γi 6→
( We may think of using weak convergence: but γi * to the diameter; and
ei * 0.)
79. Possible workarounds.
It is then, in general, quite difficult to prove that an infinite dimensional
manifold admits minimal geodesics (even when it is known to be metrically
complete).
There are though some positive results, such as[Ekeland, 1978].
Or the following.
Theorem 3.21 (Cartan-Hadamard). Suppose that M is connected, simply con-
nected and has seminegative curvature; then these are equivalent:
• (M, d) is complete
• for a c, the map η → expc (η) is well defined
and then there exists an unique minimal geodesic connecting any two points.
For the proof, see Corollary 3.9 and 3.11 in sec. IX.§3 in [Lang, 1999].

3.7 Gradient in abstract differentiable manifold

Let M be a differentiable manifold, c ∈ M , and φ1 : U1 → V1 be a chart with
c ∈ V1 .
Definition 3.22. Let E : M → lR be a functional; we say that it is Gâteaux
differentiable at c if for all h ∈ Tc M there exists the directional derivative
at c in direction h
E(φ1 (x + tk)) − E(c)
DE(c; h) := lim . (12)
t→0 t
where c = φ1 (x), h = Dφ1 (x; k).
We suppose that E is C 1
near c; this is expressed using the chart, imposing
that DE φ1 (x1 ), Dφ1 (x1 ; h) is continuous, for x1 ∈ U1 and h ∈ U . By theorem
3.5, this implies that DE(c; ·) is a linear function from Tc M into lR; for this
reason it is considered an element of the cotangent space Tc∗ M (that is the
dual space of Tc M — and we will not discuss here further).

29
 81. Gradient
Suppose now that we add a Riemannian geometry to M ; this defines an
inner product h·, ·ic on Tc M , so we can then define the gradient.
Definition 3.23 (Gradient). The gradient ∇E(c) of E at c is the unique
vector v ∈ Tc M such that
hv, hic = DE(c; h) ∀h ∈ Tc M .
If M is modeled on a Hilbert space H, and the inner product h·, ·ic used on
Tc M is equivalent to the inner product in H then the above equation uniquely
defines what the gradient is. When M is modeled on a Fréchet space, though,
in some cases the gradient does not exist.
Example 3.24. Let M = C ∞ ([−1, 1] → lR); since M is a vector space, then it is
trivially a manifold, with a single identity chart; and Tc M = M . Let E : M → lR
be the evaluation functional E(f ) = f (0), then DE(f ; h) = h(0); let
Z 1
hf, gi = f (t)g(t) dt ;
−1

then the gradient of E would be δ0 (the Dirac’s delta), that is a distribution (or
more simply a probability measure) but not an element of M .

3.8 Group actions

83. Group action
Definition 3.25. Let G be a group. We say that G acts on M if there is a
map
G×M → M
g, m 7→ g · m
that respects the group operations, that is, such that, if e is the identity element
then e · m = m, and for any g, h ∈ G, m ∈ M
h · (g · m) = (h · g) · m .
If G is topological group, that is, a group with a topology such that the group
operation is continuous, and M has a topology as well, we will require that the
action be continuous. Similarly for smooth actions between smooth manifolds.
84. Group action and quotient
The action of a group G on M induces an equivalence relation ∼, defined
by
m1 ∼ m2 iff there exists g such that m1 = g · m2 .
The orbit [m] of m is the family of all m1 equivalent to m.
The quotient M/G is the space M/ ∼ of equivalence classes.
There is a projection π : M → M/G, sending each m to its orbit π(m) =
[m].
85. Distances and groups
Let dM (x, y) be a distance on a space M , and G a group acting on M ;
suppose dM is invariant w.r.to the action of the group G, that is
dM (g · x, g · y) = dM (x, y) ∀g ∈ G .

30
Definition 3.26. A (semi)distance may be defined on B = M/G by

dB ([x], [y]) := inf dM (x, y) = inf dM (g · x, h · y) = inf dM (g · x, y)

x∈[x],y∈[y] g,h∈G g∈G

that is the lowest distance between two orbits.

We write dB (x, y) for simplicity.

Remark 3.27. There is a main problem: is dB ([x], [y]) > 0 when [x] 6= [y]?

4 Spaces and metrics of curves

In this section we review the mathematical definitions regarding the space of
curves in full detail, and express a set of mathematical goals for the theory.

4.1 Classes of curves

86. Curves
Remember that S 1 = {x ∈ lR2 | |x| = 1} is the circle in the plane.
We will often associate lR2 = C,
I for convenience. Consequently,

• sometimes S 1 will be identified with lR/(2π) (that is lR modulus 2π trans-

lations).
• but other times (and in particular if the curve is planar) we will associate
S 1 = {eit , t ∈ lR} ⊂ C.
I

A curve is a map c : S 1 → lRn . The image of the curve is c(S 1 ).

Definition 4.1 (Classes of Curves).

Imm(S 1 , lRn ) is the class of immersed curve c, such that c0 6= 0 at

•
all points.
• Immf (S 1 , lRn ) is the class of freely immersed curve, the immersed c
such that, moreover, if φ : S 1 → S 1 is a diffeomorphism and c(φ(t)) = c(t)
for all t, then φ =Id. So the curve is “completely” characterized by its
image.
Emb(S 1 , lRn ) are the embedded curves, maps c that are diffeomor-
• phic onto their image c(S 1 ); and the image is an embedded submanifold
of lRn of dimension 1.
Each class contains the one following it.
Example 4.2 (of a non-freely immersed curve).
The doubly traversed circle, defined as
c(z) = z 2 for z ∈ S 1 ⊂ C,
I a.k.a.
c(θ) = (cos(2θ), sin(2θ)) for θ ∈ lR/(2π) ∼ S 1 .
Setting φ(t) = t + π, we have that c = c ◦ φ, so c is not freely immersed.

Proposition 4.3 ([Michor and Mumford, 2006]). If c is immersed and there is

a x ∈ lRn s.t. c(t) = x for one and only one t, then c is freely immersed.

31
Proposition 4.4. Imm(S 1 , lRn ) , Immf (S 1 , lRn ) , Emb(S 1 , lRn ) , are open
subset of the Banach space C r (S 1 → lRn ) when r ≥ 1
Remark 4.5. • Imm(S 1 , lRn ) , Immf (S 1 , lRn ) , Emb(S 1 , lRn ) are (triv-
ially) submanifolds. (In particular, the charts are identity maps).
• They are closed w.r.t. reparameterizations.
• They are connected iff n ≥ 3; whereas in the case n = 2 of planar curves,
they are divided in connected components each containing curves with the
same rotation index (see Prop. 8.1 for the definition).

4.1.1 Gradients

90. Gradient in the space M of immersed curves

In the case of the space M of immersed curves, we already noted that charts
are trivial, so
E(c + th) − E(c)
DE(c; h) = lim . (13)
t→0 t
Still, though, to define the gradient we need an inner product! This inner
product may possibly be dependent on c – so in the end we need a Riemannian
metric on M .

4.2 Group actions on curves.

91. Group actions on curves.
Let M again be the manifold of immersed curves.
All groups that act on lRn also act on M . Many are of interest in Computer
Vision. The action of these groups on curves is always of the form c, A 7→ A ◦ c,
where A : lRn → lRn is the group action on lRn .

• O(n) is the group of rotations in lRn . It is represented as

O(n) := {A ∈ lRn×n | AAt = At A = Id}

The action on v ∈ lRn is the matrix·vector multiplication Av.

• SO(n) is the group of special rotations in lRn with det(A) = 1.
• lRn is the group of translations in lRn . The action on v ∈ lRn is the
vector sum v 7→ v + T .

• E(n) is the Euclidean group, containing rotations and translations.

• lR+ is the group of rescalings.

The quotient spaces M/G are the spaces of curves up to rotation and/or
translations and/or scaling (. . . ).
92. Diffeomorphisms and reparameterization
Definition 4.6. Let Diff(S 1 ) be the family of diffeomorphisms of S 1 : all the
maps φ : S 1 → S 1 that are C 1 and invertible, and the inverse φ−1 is C 1 .

32
 Diff(S 1 ) enjoys some important mathematical properties:
• Diff(S 1 ) is a group, and its group operation is the composition φ, ψ 7→
φ ◦ ψ.
• Diff(S 1 ) acts on M , the action is the right composition c, φ 7→ c◦φ. This
action is a reparameterization of c.
Note that Diff(S 1 ) has two connected components

Diff+ (S 1 ) = Diff+ (S 1 ) ∪ Diff− (S 1 )

where
• Diff+ (S 1 ) is the family of diffeomorphisms with φ0 > 0;
• Diff− (S 1 ) is the family of diffeomorphisms with φ0 < 0.
Diff+ (S 1 ) is a group, its action does not change the orientation of a curve.

4.3 Geometric curves, and functionals

94. Geometric curves
Shapes are usually considered to be geometric objects. Representing a curve
using c : S 1 → lRn forces a choice of parameterization, that is not really part of
the concept of “shape”. We present a first simplified definition.

Definition 4.7. The quotient space B = M/Diff(S 1 ) is the space of curves

up to reparameterization, also called geometric curves in the following. Two
parametric curves c1 , c2 ∈ M such that c1 = c2 ◦ φ are the same geometric curve
inside B.
B is mathematically defined as the set B = {[c]} of all equivalence classes
[c] of curves that are equal but for reparameterization,

[c] := {c ◦ φ for φ ∈ Diff(S 1 )}.

We may also consider the quotient w.r.t Diff+ (S 1 ). The quotient space
M/Diff+ (S 1 ) is the space of geometric oriented curves.
95. Geometric functionals
Definition 4.8. A functional F (c) defined on curves will be called geometric
if it is invariant w.r.t. reparameterization of c; this may take two forms,
• F (c) = F (c ◦ φ) for all curves c and for all φ ∈ Diff(S 1 ).
• For all curves c and all φ, if φ ∈ Diff+ (S 1 ) then F (c) = F (c ◦ φ), whereas
if φ ∈ Diff− (S 1 ) then F (c) = −F (c ◦ φ)
In the first case, F may be “projected” to B = M/Diff(S 1 ), that is, it may
be considered as a function F : B → lR.
“Geometrical” theories have often provided the best results in Computer
Vision.
96. Gâteaux differential of rep. invariant operator

33
 Consider the case of operators when O(c) is a curve or a vector field, and O
is reparameterization invariant i.e. ∀ϕ ∈ Diff+ (S 1 )

O(c ◦ φ)(θ) = O(c)(φ(θ)) . (14)

−
(whereas for ϕ ∈ Diff (S 1 ) there is a possible sign choice).
Proposition 4.9. Let a : S 1 → lR be an infinitesimal perturbation of Id in
Diff(S 1 ),

Dc,h O(c)(θ) = a(θ) . ∂θ O(c)(θ) for any h(θ) = a(θ).∂θ c(θ) (15)

and this last may be rewritten in arc parameter

Dc,h O(c)(s) = a(s) . Ds O(c)(s) for any h(s) = a(s).Ds c(s) . (16)
Proof. We define ϕ : lR × S 1 → S 1 by

ϕ̇(t, θ) = a(ϕ(t, θ))
ϕ(0, θ) = θ

(ϕ(t, ·) = exp(ta), the exponential map in Diff(S 1 )) We define γ(t, θ) = c(t, ϕ(t, θ)).
To compute Dc,h O(c) we compute ∂t O(γ)|t=0 but by invariance of O O(γ)(θ) =
O(c)(ϕ(t, θ)) so ∂t O(γ) = aO(c)(ϕ(t, θ)) hence we set t = 0.

Example 4.10. If O = T = Ds c then by (3) in Prop. 1.9

Dc,h O = Ds h − hDs h, T iT = πN (Ds h)

so when h = aT we get

Dc,h O = πN ((Ds a)T ) + πN (aDs T ) = aDs T

Similarly
Proposition 4.11. if E : M → lR is a reparameterization invariant energy,
then
Dc,h E = 0 for any h = a.Ds c . (17)

4.4 The two categories of “Shape Spaces”

98. Two categories of “Shape Spaces”
We should distinguish two different ideas of shape spaces of curves.
geometric curves geometric curves up to pose
Shapes = curves up to reparameter-
(pre)Shapes = curves up to repa-
ization, rotation, translation, scal-
rameterization,
ing,
is modeled as is modeled as
M/Diff(S 1 ) , M/Diff(S 1 )/E(n)/lR+ ,
and is well-suited to and is well-suited to
Shape Optimization. Shape Analysis.
References:
References:
[Michor-Mumford],
[Mio-Srivastava et al], [Younes],
[Yezzi-M.-Sundaramoorthi],
[Younes et al., 2008]
[Glauneès-Trouvé-Younes et al]

34
4.5 Geometric curves
99. Main definition: “geometric curves” [Michor and Mumford, 2006]
Unfortunately the quotient

Bi = Imm(S 1 , lR2 )/Diff(S 1 )

of immersed curves up to reparameterization is not a manifold.

We (re)define the space of geometrical curves.
Definition 4.12.

Bi,f (S 1 , lR2 ) = Immf (S 1 , lR2 )/Diff(S 1 )

is the quotient of Immf (S 1 , lR2 ) (free immersions) by the diffeos Diff(S 1 ) (repa-
rameterization).
The good news is that
Theorem 4.13 (§2.4.3 in [Michor and Mumford, 2006]). If Immf and Diff(S 1 )
have the topology of the Fréchet space of C ∞ functions, then Bi,f is a Fréchet
manifold modeled on C ∞ .
The bad news is that
• when we add a simple Riemannian metric to Bi,f , the resulting metric
space is usually not metrically complete;
• Some calculus is lost (Cauchy thm, exponential map).
Remark 4.14. It seems that this is the only way to properly define the man-
ifold. If otherwise we choose M = C k (S 1 → lR2 ) to be the manifold of curves,
then if c ∈ M , c0 6∈ Tc M . Hence we (must?) use M = C ∞ functions.

4.5.1 Bi,f is a Fréchet manifold

101. Charts in Bi,f

We now prove Thm. 4.13, namely, Bi,f is a Fréchet manifold
Let Π be the projection from Immf (S 1 , lR2 ) to the quotient Bi,f .
Let [c] ∈ Bi,f : we pick a curve c such that Π(c) = [c]. We represent the tan-
gent space T[c] Bi,f as the space of all k : S 1 → lRn such that k(s) is orthogonal
to c0 (s).
We can define a simple natural chart Φ[c] by

Φ[c] (k) := Π(c(·) + k(·))

that is, it moves c(u) in direction k(u); and it is easily seen that the chart does
not depend on the choice of c such that Π(c) = [c].
We will prove that the family Φ[c] is a smooth atlas for Bi,f .
102. Lemmas for proof of the theorem 4.13
Let c be a C 2 immersed planar curve. We recall these properties.
Lemma 4.15. A curve c ∈ C 2 can be reparameterized to c̃ = c ◦ ϕ with
|c̃0 | =constant.

35
 Note that, since the curve is closed, then κ cannot be identically zero.
Proposition 4.16 (Local embedding). Let δ := π/(4 max |κ|). For any a, b ∈
S 1 with L := len c|[a,b] ≤ δ; then |c(b) − c(a)| ≥ L/2; so c|[a,b] is embedded.
Proposition 4.17 (tubolar neighbourhood). Let δ := π/(4 max |κ|), τ :=
1/(2 max |κ|). Fix a, b ∈ S 1 with len c|[a,b] ≤ δ. Let φ(s, t) = c(s) + tN (s),
for s ∈ [a, b] and t ∈ [−τ, τ ]. Then φ is a diffeomorphism.
As a corollary, there is a neighbourhood of the image of c where the “projection
to c” is a C 2 multivalued map (with finitely many projections).

Lemma 4.18. Let c be a C 2 immersed planar curve. There exists a r > 0 such
that, if
c(s) + a(s)N (s) = c(φ(s))
with φ ∈ Diff(S 1 ) and a : S 1 → lR continuous with kak ≤ r, then a ≡ 0.
(Note that if c is a circle, then c(s) + 2N (s) = c(s + φ)).
Lemma 4.19. Immf is open in Imm

104. Proof of the theorem 4.13

Let i = 1, 2, ci ∈ Immf , ai : S 1 → lR, and define the charts Φi (a) = [ci +aNi ].
• we prove that Φ1 is injective on U1 = {kak ≤ r1 , ka0 k ≤ r1 }
• If a1 , a2 are in U1 , U2 , then, given a1 we can solve the equation

c1 (θ) + a1 (θ)N1 (θ) = c2 (ϕ(θ)) + a2 (ϕ(θ))N2 (ϕ(θ))

for ϕ ∈ Diff(S 1 ) and a2 , and the “solving” is a smooth operation

4.6 Goals (revisited)

105. Research path
Following [Michor and Mumford, 2006] we so obtained a possible program of
math research.
• define
Bi,f (S 1 , lR2 ) = Immf (S 1 , lR2 )/Diff(S 1 )
and consider B as a Fréchet manifold modeled on C ∞ ,
• define a Riemann/Finsler geometry on it, study its properties,
• metrically complete the space.
In the last step, we would hope to obtain a differentiable manifold; unfortu-
nately, this is sometimes not true, as we will see in the overview of the litera-
ture.
106. Abstract requirements
We formulate an abstract set of goal properties on a metric hh1 , h2 iG|c on
spaces of curves.
1. [rescaling] λ > 0, rescale c to λc, then hh1 , h2 iG|λc = λa hh1 , h2 iG|c

36
 2. [euclidean invariance] apply euclidean transformation A to c (with R
is its rotational part) then
hRh1 , Rh2 iG|Ac = hh1 , h2 iG|c
Implies that distance is invariant w.r.t. euclidean transformation,
d(Ax, Ay) = d(x, y) ∀A roto-traslation

3. [parameterization invariance] two versions:

• curve-wise p.i. (weaker): when the metric does not depend on
the parameterization of the curve, that is kh̃kc̃ = khkc when c̃(t) =
c(ϕ(t)) and h̃(t) = h(ϕ(t))
• homotopy-wise (stronger): for any ϕ : [0, 1] × S 1 → S 1 smooth,
ϕ(t, ·) a diffeo of S 1 for all fixed t, let C̃(t, θ) = C(t, ϕ(t, θ)), then
E(C̃) = E(C).
If a metric satisfies the above three properties (3rd in weaker form), we say
that it is a geometrical metric.
107. Well posedness
Do not forget
0. [well-posedness and existence of minimal geodesics]
• The metric induces a good distance d: that is, the distance between dif-
ferent curves is positive, and d generates the same topology that the atlas
of the manifold M induces
• (M, d) is complete
• for any two curves in M , there exists a minimal geodesic connecting them.

108. Are the goal properties consistent?

So we state the abstract problem:
Problem 4.20. Consider the space of curves M , and the family G of all Rie-
mannian (or, regular Finsler) Geometries F on M .
Does there exist a metric F ∈ G, satisfying the above properties 0,1,2,3?
Does there exist a metric F ∈ G that may be written in integral form
Z
F (c, h) = f c(s), ds c(s), . . . , djsj c(s), h(s), . . . , disi h(s) ds


c

and what is the relationship between the degrees i, j and the properties in this
section?
Also said: can we design metrics?.
Start of lecture: III. Literature Overview

Part III
Literature Overview
In this third lecture we will present some current literature, and discuss it

37
5 Representation/embedding/quotient
110. The representation/embedding/quotient paradigm
A common way to model shapes is by representation/embedding:
• we represent the shape A by a function uA
• and then we embed this representation in a space E, so that we can
operate on the shapes A by operating on the representations uA .
Most often, representation/embedding alone do not directly provide a sat-
isfactory Shape Space. In particular, in many cases it happens that the repre-
sentation is “redundant”, that is, the same shape has many different possible
representations. An appropriate quotient is then introduced.
There are many examples of Shape Spaces in the literature that are studied
by means of the representation/embedding/quotient scheme. Understanding the
basic math properties of these 3 operations is then a key step in understanding
Shape Spaces and designing/improving them.
We now present a rapid overview of how this scheme is exploited in the
current literature of Shape Spaces; then we will come back to some of them to
explain more in depth.
Remark 5.1. When we started talking of geometric curves, we proposed the
quotient B = M/Diff(S 1 ) (the space of curves up to reparameterization); and
this had to be followed by other quotients w.r.to Euclidean motions and/or
scaling, to obtain M/Diff(S 1 )/E(n)/lR+ . In practice, though, the space B
happens to be more difficult to study; hence most shape space theories that deal
with curves prefer a different order: the quotient M/E(n)/lR+ is modeled and
studied first; then the quotient M/E(n)/lR+ /Diff(S 1 ) is performed (often, only
numerically).

112. Compact sets and templates

Example 5.2 (The family of all non empty compact subsets of lRN ). A standard
representation is obtained by associating each closed subset A to the distance
function uA or the signed distance function bA (that were defined in slide 33).
We may then define a topology of shapes by deciding that An → A when
uAn → uA uniformly on compact sets. This convergence coincides with the
Kuratowski topology of closed sets; if we limit shapes to be compact sets, the
Kuratowski topology is induced by the Hausdorff distance. See section 6.2.
Example 5.3. Trouvé–Younes et al (see [Glaunès et al., 2005], [Trouvé and
Younes, 2005] and references therein) modeled the motion of shapes by studying
a left invariant Riemannian metric on the diffeomorphisms of the space lRN ; to
recover a true metric of shapes, a quotient is then performed.
113. Curves
But the representation/embedding/quotient scheme is also found when deal-
ing with spaces of curves:
Example 5.4 (Representation by angle function). In the work of [Klassen et al.,
2004; Srivastava et al., 2005; Mio and Srivastava, 2004b], smooth planar closed
curves c : S 1 → lR2 of length 2π are represented by a pair of angle-velocity func-
tions c0 (u) = exp(φ(u) + iα(u)) (identifying lR2 = C)
I then (φ, α) are embedded

38
as a subset N in L2 (0, 2π) or W 1,2 (0, 2π). Since the goal is to obtain a Shape
Space representation for Shape Analysis purposes, a quotient is then introduced
on N .
Example 5.5. Another representation of planar curves for Shape Analysis is
found in [Younes, 1998b]. In this case the angle function is considered mod(π).
This representation is both simple and very powerful at the same time. In-
deed, it is possible to prove that geodesics do exist and to explicitly show
examples of geodesics: see [Younes et al., 2008].
114. More examples
Example 5.6 (Harmonic representation). A. Duci et al (see [Duci et al., 2003,
2006]) represent a closed planar contour as the zero level of a harmonic function.
This novel representation for contours is explicitly designed to possess a linear
structure, which greatly simplifies linear operations such as averaging, principal
component analysis or differentiation in the space of shapes.
And, of course, we have in this list the spaces of embedded curves.
Example 5.7. When studying embedded curves, usually, for the sake of math-
ematical analysis, the curves are modeled as immersed parametric curves; a
quotient w.r.t the group of possible reparameterizations of the curve c (that
coincides with the group of diffeomorphisms Diff(S 1 )) is applied afterward to
all the mathematical structures that are defined (such as the manifold of curves,
the Riemannian metric, the induced distance, etc.).

6 Metrics of sets
We now present two example of Shape Theories where a shape may be a generic
subset of the plane; with particular attention to how they behave wrt curves.

6.1 Some more maths on distance and geodesics

We start by reviewing some basic results in abstract metric spaces theory.
117. Length induced by a distance
In this subsection (M, d) is a generic metric space.
Definition 6.1 (Length by total variation). Define the length of a continuous
curve γ : [α, β] → M , by using the total variation
t3 t6
X n t1 t4
d


Len γ := sup d γ(ti−1 ), γ(ti )
T i=1
t0 t2 t 5
where the sup is for all T = {t0 , · · · , tn } ⊂ [α, β] and t0 ≤ · · · ≤ tn .
Definition 6.2 ( the induced geodesic distance).

dg (x, y) := inf Lend γ

γ

the inf is in the class of all continuous γ in M connecting x to y.

Note that dg (x, y) < ∞ iff x, y may be connected by a Lipschitz path.
118. Minimal geodesics

39
Definition 6.3. If in the definition dg (x, y) := inf γ Lend γ there is a curve γ ∗
providing the minimum, then γ ∗ is called a (minimal) geodesic connecting x
to y.
Proposition 6.4. In Riemann and Finsler manifolds, the integral length Len(γ)
that we defined in slide 68 coincides with the total variation length Lend γ that
we defined in 6.1. As a consequence, d = dg : d is path-metric.
In general, though, it is easy to devise examples where d 6= dg .
Example 6.5 (for M = S 1 ⊂ lR2 ).
if d(x, y) = |x − y|,
then dg (x, y) = | arg(x) − arg(y)|.

119. Existence of geodesics

Proposition 6.6. If for a ρ > 0 the closed ball

Dg (x, ρ) := {y | dg (x, y) ≤ ρ}

is compact, then x and any y ∈ Dg may be connected by a geodesic.

Proposition 6.7. If for a x ∈ M and all ρ > 0, Dg is compact, then the Fréchet
distance based average (that was defined in slide 36) exists.

120. Geodesic rays

More in general,
Definition 6.8. a continuous curve γ : I → M (where I ⊂ lR is an interval) is
a geodesic ray if for each t ∈ I there is a ε > 0 s.t. J = [t − ε, t + ε] ⊂ I and
γ restricted to J is the geodesic between γ(t − ε) and γ(t + ε).
A critical geodesic in a Riemann smooth manifold (modeled on a Hilbert
space) is always a geodesic ray, as in this example.
Example 6.9. The multiply traversed full circle γ(t) = (0, cos(t), sin(t)) with
t ∈ lR is a geodesic ray in the sphere S 2 .
121. Hopf–Rinow , Cohn–Vossen Theorem
Definition 6.10. (M, d) is path-metric if d = dg .
(M, d) is locally compact if small closed balls are compact.
Theorem 6.11 (Hopf–Rinow , Cohn–Vossen). Suppose (M, d) is path-metric
and locally compact: the following are equivalent
• for all x ∈ M, ρ > 0, the closed ball D(x, ρ) := {y | d(x, y) ≤ ρ} is compact,
• (M, d) is complete,
• any geodesic ray γ : [0, 1) → M may be extended on [0, 1];
and all imply that any two points may be connected by a minimal geodesic.
Note that the above theorem cannot be used in infinite dimensional smooth
differentiable manifolds (modeled on Banach spaces), since the the closed balls
are never compact in that case.

40
6.2 Hausdorff metric
122. Hausdorff distance
Let Ξ be the collection of all compact subsets of lRN . (This is sometimes called
the“topological hyperspace” of lRN ). Let A, B ⊂ lRN compact non-empty. We
defined the distance function uA (x) := inf y∈A |x − y|. We defined the
Hausdorff distance of two sets A, B as
   
dH (A, B) := sup uB (x) ∨ sup uA (x) .
x∈A x∈B

Properties. • dH (A, B) = sup |uA (x) − uB (x)| .

x∈lRN

• (Ξ, dH ) is path–metric;
• each family of compact sets that are contained in a fixed large closed ball
in lRN is compact in (Ξ, dH ); so
• any closed ball D(A, ρ) := {B | dH (A, B) ≤ ρ} is compact in (Ξ, dH ), and
moreover
• by Theorem 6.11, minimal geodesics exist in (Ξ, dH ).
123. An alternative definition
Let Dr (x) be the closed ball of center x and radius r > 0 in lRN , and
Dr = Dr (0). We define the fattened set to be
[
A + Dr := {x + y | x ∈ A, |y| ≤ r} = Dr (x) = {y | uA (y) ≤ r}.
x∈A

Examples of a set A fattened to r = 1, 2; the set

A is the black polygon (and is filled in), whereas
111111111111
000000000000
the dashed lines in the drawing are the contour of 0110 0110
a fattened sets, (that are not drawn filled – other-
100
wise they would cover A). 110
Note that the fattened set is always closed, (since
the distance function uA (x) is continuous). 2
1
1010
We then can state the following equivalent defini- 1010
10
tion.

dH (A, B) = min{δ > 0 | A ⊂ (B+Dδ ), B ⊂ (A+Dδ )} .

1111
0000
124. Uncountable geodesics
Unfortunately dH is quite “unsmooth”. There are choices of A, B compact
that may be joined by an uncountable number of geodesics. In fact we can
consider this simple example.
Example 6.12 ([Duci and Mennucci, 2007]).
A = {x = 0, 0 ≤ y ≤ 2}
A Ct
B = {x = 2, 0 ≤ y ≤ 1}
B
3

Ct = x = 1, 0 ≤ y ≤ 2 ∪ {y = 0, 1 ≤ x ≤ t}
√
with 1 ≤ t ≤ 5/2;

41
 and in the picture we represent (dashed) the fattened sets A + D√5/2 and
√ √
B + D√5/2 . Note that dH (A, B) = 5 while dH (A, Ct ) = dH (B, Ct ) = 5/2:
so Ct are all midpoints that are on different geodesics between A and B.
125. Curves and connected sets
Let Ξc be the class of compact connected A ⊂ lRN ; Ξc is a closed subset
of (Ξ, dH ); Ξc is the closure in (Ξ, dH ) of the class of (images of) all embedded
curves. Ξc is Lipschitz-path-connected 1 ; for all above reasons, it is possible
to connect any two A, B ∈ Ξc by a minimal geodesic moving in Ξc . So, if we
try to find a minimal geodesic connecting two curves using the metric dH , we
will end up finding a geodesic in (Ξc , dH ); and similarly if we try to optimize
an energy defined on curves. But note that Ξc is not geodesically convex in Ξ,
that is, there exist two points A, B ∈ Ξc and a minimal geodesics ξ connecting
A to B in the metric space (Ξ, d), such that the image of ξ is not contained
inside Ξc . We don’t know if (Ξc , dH ) is path-metric.
126. Applications in Computer Vision
[Charpiat et al., 2004] proposed an approximation method to compute lenH (ξ)
by means of a family of energies defined using a smooth integrand; the approx-
imation is mainly based on the property kf kLp →p kf kL∞ , for any measurable
function f defined on a bounded domain; they successively devise a method to
find approximation of geodesics.

6.3 A Hausdorff like distance of compact sets

127. Distance from embedding. [Duci and Mennucci, 2007]
We fix ϕ : [0, ∞) → (0, ∞) , decreasing, C 1 , ϕ(|x|) ∈ Lp .
Define vA (x) := ϕ(uA (x)).
We eventually define the distance

d(A, B) := kvA − vB kLp . (18)

That is, our Shape Space is represented as Nc = {vA | A compact} and

embedded in Lp . Given v ∈ Nc , we recover the shape A = {v = ϕ−1 (0)} (that
is a level set of v).
Example: ϕ(t) = e−t , vA (x) = exp (−uA (x)), A = {v = 1}.

Properties (of this metric d). • it is Euclidean–invariant;

• it is locally compact but not path–metric.
• the topology induced is the same as that induced by dH
• minimal geodesics do exist, since Dg := {B | dg (A, B) ≤ ρ} is compact

128. Analogy with the Hausdorff metric, Lp vs L∞

We recall that dH (A, B) = kuA (x) − uB (x)kL∞ ; instead now d(A, B) :=
kvA − vB kLp . This distance of compact sets is modeled on Lp , whereas the
Hausdorff distance is “modeled” on L∞ . (Note that the Hausdorff distance is
not really obtained by embedding, since uA 6∈ L∞ ). Lp is more regular than
L∞ , as shown by this remark.
1 That is, any A, B ∈ Ξc can be connected by a Lipschitz arc γ : [0, 1] → Ξc

42
Remark 6.13. Given any f, g ∈ Lp with p ∈ (1, ∞), the segment connecting
f to g is the unique minimal geodesic connecting them. Suppose now that the
dimension of L∞ (Ω, A, µ) is greater than 1. Given generically f, g ∈ L∞ , there
is an uncountable number of minimal geodesics connecting them. (“Generically”
is meant in the Baire sense).
(For the proof see 2.11 & 2.13 in [Duci and Mennucci, 2007]). The above
result suggests that it may be possible to shoot geodesics in the “Riemannian
metric” associated to this distance.
129. Contingent cone
Given a v ∈ Nc , let Tv Nc ⊂ Lp be the contingent cone

Tv Nc := {lim tn (vn − v) | tn > 0, vn ∈ Nc , vn → v}

n
 
vn − v
= λ lim | λ ≥ 0, vn → v ,
n kvn − vk p
L

where it is intended that the above limits are in the sense of strong convergence
in Lp . Tv Nc contains all directions in which it is possible “in the Lp sense” to
infinitesimally deform a compact set. For example, if Φ(x, t) : lRN × (−ε, ε) →
lRN is smooth diffeomorphical motion of lRN , and At := Φ(A, t), then vAt is
(locally) Lipschitz, so it is almost everywhere differentiable, and the derivative
is in T Nc . This includes all perspective deformations. Unfortunately the
contingent cone is not capable of expressing some shape deformations.
Example 6.14. We consider the removing motion; let A be compact, and suppose
that x is in the internal part of A; let At := A \ B(x, t) be the removal of a
small ball from A. The motion vAt inside Nc is Lipschitz, but the limit
vA t − vA
lim
t→0+ kvAt − vA k Lp

does not exist in Lp (morally, if p = 1 it is the measure δx ).

131. Riemannian metric
Let now p = 2. The set Nc may fail to be a smooth submanifold of L2 ;
yet we will, as much as possible, pretend that it is, in order to induce a sort of
“Riemannian metric” on Nc from the standard L2 metric.

Definition 6.15. We define the “Riemannian metric” on Nc simply by

hh, ki := hh, kiL2

for h, k ∈ Tv Nc and correspondingly a norm by

p
|h| := hh, hi

where Tv Nc is the contingent cone.

132. Riemannian metric, properties

Proposition 6.16. the distance induced by this “Riemannian metric” coincides
with the geodesically induce distance dg .

43
 The proof is in 3.22 in [Duci and Mennucci, 2007].
To conclude, we propose an explicit computation of the Riemannian Metric
for the case of compact sets in the plane with smooth boundaries; we then pull
back the metric to obtain a metric of closed embedded planar curves. We start
with the case of convex sets.We fix p = 2, N = 2.
133. Polar coordinates of smooth convex sets
Let Ω ⊂ lR2 be a convex set with smooth boundary; let y(θ) : [0, L] → ∂Ω
be a parameterization of the boundary (by arc-parameter), ν(θ) the unit vector
normal to ∂Ω and pointing external to Ω: then the following “polar” change
of coordinates ψ holds:

ψ : lR+ × [0, L] → lR \ Ω , ψ(ρ, θ) = y(θ) + ρν(θ)

y(θ)

ν(θ) y(θ) + ρν(θ)

We suppose that y(θ) moves on ∂Ω in anticlockwise direction; so ν = J∂s y,

∂ss y = −κν; where J is the rotation matrix (of angle −π/2), κ is the curvature,
and ∂s y is the tangent vector.
We can then express a generic integral through this change of coordinates as
Z Z Z
f (x) dx = f (ψ(ρ, s))|1 + ρκ(s)| dρds
lR2 \Ω lR+ ∂Ω

where s is arc parameter, and ds is integration in arc parameter.

135. Smooth deformations of a convex set
We want to study a smooth deformation of Ω, that we call Ωt ; then the
border y(θ, t) depends on a time parameter t. Suppose also that κ(θ) > 0, that
is, that the set is strictly convex: then for small smooth deformations, the set
Ωt will still be strictly convex.
We suppose that the border of Ωt moves with orthogonal speed α; more
precisely, we assume that (∂t y) ⊥ ∂s y, that is, (∂t y) = αν with α = α(t, θ) ∈ lR.
Since this deformation is smooth, we expect that it will be associated to a vector
hα ∈ Tv Nc , defined by hα := ∂t vΩt .
We now show briefly how to explicitly compute it. By deriving

∂t ∂s y = ∂s (∂t y) − ∂s yh∂s y, ∂s (∂t y)i = πν (∂s (∂t y))

where
πν (w) := νhν, wi = w − ∂s yh∂s y, wi
is the projection of w parallel to ν. Supposing now that ρ = ρ(t) as well, we
can express the point ψ(ρ, y) in a first order approximation as
   
dψ = (∂t y) + ρ0 ν + ρJπν (∂s (∂t y)) dt + ∂θ y + ρ∂θ ν dθ

where moreover
     
∂θ y + ρ∂θ ν dθ = ∂s y + ρ∂s ν ds = 1 + ρκ ∂s y ds .

44
 If y(θ, t), ρ(t) are expressing a constant point x = ψ(ρ, y), then dψ = 0; we
apply scalar products w.r.t. ν and ∂s y to the above relations

hν, (∂t y)i + ρ0 = 0 , h∂s y, (∂t y)i − ρhν, ∂s (∂t y)idt + (1 + ρκ) ds = 0 .

Assuming that (∂t y) ⊥ ∂s y, that is, (∂t y) = αν with α = α(t, θ) ∈ lR, we

obtain the relationships
ds ρhν, ∂s (αν)i ρ ∂s α
ρ0 = −α , = = .
dt (1 + ρκ) (1 + ρκ)
Now, for x 6∈ Ωt , uΩt (x) = ρ(t) hence

hα := ∂t vΩt (x) = −ϕ0 (uΩt (x))α

whereas hα (x) = 0 for x ∈ Ω̊t ; so hα is the vector in Tv Nc that is associated

to α.
138. Pullback of the metric on convex boundaries
Let us then fix two orthogonal smooth vector fields α(s)ν(s), β(s)ν(s), that
represent two possible deformations of ∂Ω; those correspond to two vectors
hα , hβ ∈ Tv Nc ; so the Riemannian Metric that we defined in 6.15 can be pulled
back on ∂Ω, to provide the metric
Z Z
hα, βi := hα (x)hβ (x)dx = hα (x)hβ (x)dx =
lR2 lR2 \Ω
Z Z 
= (ϕ0 (ρ))2 (1 + ρκ(s)) dρ α(s)β(s)ds
∂Ω lR+

that is, Z
hα, βi = (a + bκ(s))α(s)β(s)ds (19)
∂Ω
with Z Z
0
a= (ϕ (ρ)) dρ 2
, b= (ϕ0 (ρ))2 ρ dρ .
lR+ lR+

139. Pullback of the metric on smooth contours

If Ω is smooth but not convex, then the above formula holds up to the
cutlocus (a.k.a. the external skeleton). We define a function R(s) : [0, L] → lR+
that spans the cutlocus Cut, that is,

Cut = {ψ(R(s), s), s ∈ [0, L]} .

ψ is a diffeomorphism between the sets

{(ρ, s) s ∈ [0, L], 0 < ρ < R(s)} ↔ lR2 \ (Ω ∪ Cut)

moreover R(s) is Lipschitz (by results in [Itoh and Tanaka, 2000],[Li and Niren-
berg, 2005]).

Cut

45
 The pullback metric for deformations of the contour ∂Ω is
Z "Z R(s)
#
hh, ki = (ϕ0 (ρ))2 (1 + ρκ(s)) dρ α(s)β(s) ds
∂Ω 0

We present a numerical computation of a minimal geodesic (by A. Duci).

Figure 2: Example of a minimal geodesic

142. Conclusion
In this case we have found (aposteriori) a Riemannian metric of closed em-
bedded planar curves, and we know the structure of the completion, and the
completion admits minimal geodesics. On the down side, the completion is
not really “a smooth Riemannian manifold”. For example, it is difficult to study
the minimal geodesics, and to prove any property about them (for the reasons
already explained in slide 125). Anyway, the fact that Nc is locally compact
and the regularity property 6.13 of Lp spaces suggest that it may be possible
to “shoot geodesics” (in some weak form). Unfortunately the topology has too
few open sets to be used for Shape Optimization: many simple functionals are
not continuous.

7 Finsler geometries of curves

We present two examples of Finsler geometries of curves that have been used
(sometimes covertly) in the literature.

46
7.1 L∞ metric and Fréchet distance
144. Tangent and normal
Let v, T ∈ lRN , let N = T ⊥ .2 We define the projection onto the normal space
N = T⊥
πN : lRN → lRN , πN v = v − hv, T iT
and on the tangent T

πT : lRN → lRN , πT v = hv, T iT

so πN v + πT v = v and |πN v|2 + |πT v|2 = |v|2

145. The L∞ metric of curves
If we wish to define a norm on Tc M that is modeled on the norm of the
Banach space L∞ (S 1 → lRN ), we define

F ∞ (c, h) = kπN hkL∞ = sup |πN h(θ)|

θ

this metric weights for the maximum normal motion of the curve.
This Finsler metric is (homotopy-wise) geometric. The length of a homotopy
is then Z 1
Len(C) = sup |πN ∂t C(t, θ)| dv
0 θ

(The rôle of πN will be properly explained in the next lecture).

146. an auxiliary result in Finsler metrics
Definition 7.1 (Fréchet distance).

df (c0 , c1 ) := inf sup |c1 (φ(u)) − c0 (u)|

φ u

where u ∈ S 1 and φ is chosen in the class of diffeomorphisms of S 1 .

This distance is induced by the Finsler metric F ∞ (c, h) that is,

Theorem 7.2 (theorem 15 in [Mennucci et al., 2008]). if the length of a homo-

topy C(t, θ) is
Z 1
∂C
Len∞ (C) = sup πN dt
0 θ∈S 1 ∂t
then df (c0 , c1 ) coincides with the infimum of that length Len∞ (C) for C con-
necting c0 to (a reparameterization of) c1 .

7.2 L1 metric and Plateau problem

If we wish to define a geometric norm on Tc M that is modeled on the norm of
the Banach space L1 (S 1 → lRN ), we may define the metric
Z
F (c, h) = kπN hkL1 = |πN h(θ)||c0 (θ)|dθ
1

2 There is a slight abuse of notation here, since in the definition N = T ⊥ given for planar

curves in 1.7, we defined N to be a “vector” and not a “vector space”.

47
 The length of a homotopy is then
ZZ
Len(C) = |πN ∂t C(t, θ)||C 0 (t, θ)| dt dθ

which coincides with

ZZ
Len(C) = |∂t C(t, θ) × ∂θ C(t, θ)| dθ dt

This last is easily recognizable as the surface area of the homotopy (up to
multiplicity); the problem of finding a minimal geodesic connecting c0 and c1 in
the F 1 metric may be reconducted to the Plateau problem of finding a surface
which is an immersion of I = S 1 × [0, 1] and which has fixed borders to the
curves c0 and c1 . The Plateau problem is a wide and well studied subject upon
which Fomenko expounds in the monograph [Fomenko, 1990].

8 Riemannian metrics of curves

148. Why?
It is quite clear why we research metrics of curves in Shape Analysis.
This research can be very useful, though, for Shape Optimization as well.
Consider an Active Contour Energy E(c) used in Shape Optimization.
Changing the metric does not change the (global) minima of E.
Changing the metric changes the minimization flow of E.
So we may hope to better solve Shape Optimization problems (without
adding regularization/penalizations terms to E).

8.1 Curves up to pose, as a submanifold of a Hilbert space

8.1.1 Shape Representation using Direction Functions [Klassen et al.,
2004]
A particular approach to the study of shapes is to define a shape to be a curve
up to reparameterization, rotation, translation, scaling.
Mio, Srivastava et al in [Klassen et al., 2004; Mio and Srivastava, 2004b]
use a different choice of curve representation: they represent a planar curve c
by a pair of velocity-angle functions (φ, α) through the identity

c0 (u) = exp(φ(u) + iα(u))

(identifying lR2 = C),

I and then defining a metric on the velocity-angle function
space. They propose models of spaces of curves where the metrics involve higher
order derivatives in [Klassen et al., 2004].
We review here the simplest such model, where φ = 0. We consider in
the following planar curves ξ : S 1 → lR2 of length 2π and parameterized by
arclength. (Such curves are automatically Lipschitz continuous).
150. Lifting, rotation number
We consider in the following planar curves ξ : S 1 → lR2 of length 2π and
parameterized by arc length.

48
Proposition 8.1 (Continuous lifting, rotation index). If ξ ∈ C 1 and by arc-
parameter, then ξ 0 is continuous and |ξ 0 | = 1, so there exists a continuous
function α : lR → lR satisfying

ξ 0 (s) = cos(α(s)), sin(α(s))

(20)

and α(s) is unique, up to adding the constant k2π with k ∈ Z.

Moreover α(s + 2π) − α(s) = 2πI, where I is an integer, known as rotation
index of ξ. This number is unaltered if ξ is deformed in a smooth way.
The use of the angle function α for representing (possibly non-closed) planar
curves dates back to [Zahn and Roskies, 1972].
Example 8.2 (of curves of different rotation index).

3
1 −2
0
152. Shape Representation using Direction Functions
The addition of a generic constant to α is equivalent to a rotation of ξ.
We then understand that we may represent arc-parameterized curves ξ(s), up
to translation, scaling, and rotations, by considering a suitable class of liftings
α(s) for s ∈ [0, 2π].
Two spaces are defined in [Klassen et al., 2004]; we present the case of
“Shape Representation using Direction Functions”; let L2 := L2 ([0, 2π] → lR);
Φ : L2 → lR3 is defined by
Z 2π Z 2π Z 2π
Φ1 (α) := α(s) ds , Φ2 (α) := cos α(s) ds , Φ3 (α) := sin α(s) ds .
0 0 0

and the space of (pre)-shapes is defined as the closed subset S of L2 ,

S := α ∈ L2 | Φ(α) = (2π 2 , 0, 0) ;


the condition Φ1 (α) = 2π 2 forces a choice of rotation, while Φ2 = 0, Φ3 = 0

ensure that α represents a closed curve.
153. Preshape and shape
For any α ∈ S, it is possible to reconstruct the curve by integrating
Z s


ξ(s) = cos(α(t)), sin(α(t)) dt (21)
0

This means that α identifies an unique curve (of length 2π, and arc-parameterized)
up to rotations and translations, and to the choice of the base point ξ(0); for
this last reason, S is called in [Klassen et al., 2004] a preshape space.
Inside the family of arc-parameterized curves, the reparameterization of ξ is
ˆ
restricted to the transformation ξ(u) = ξ(u − s0 ), that is a different choice of
basepoint for the parameterization along the curve. Inside the pre shape space
S, the same operation is encoded by the relation α̂(s) = α(s − s0 ). To obtain
a model of immersed curves up to reparameterization, rotation, translation,
scaling (recalling slide 98), we need to quotient S by the relation α ∼ α̂, for all
possible s0 ∈ lR. We do not discuss this quotient here.
154. S \ Z is a manifold

49
Definition 8.3 (Flat curves). Let Z be the set of all α ∈ L2 ([0,R2π]) such that
α(s) = a + k(s)π where k(s) ∈ Z and k is measurable, a = 2π − k ∈ lR, and

|{k(s) = 0 mod 2}| = |{k(s) = 1 mod 2}| = π

Z is closed (by thm. 4.9 in [Brezis, 1986]). We see that Z contains the
(representations α of) flat curves ξ, that is, curves ξ whose image is contained
in a line;
Example 8.4. one such curve is
( √ (
s/ 2 s ∈ [0, π] π/2 s ∈ [0, π]
ξ1 (s) = ξ2 (s) = √ , α=
(2π − s)/ 2 s ∈ (π, 2π] 3π/2 s ∈ (π, 2π]

Proposition 8.5. By the implicit function theorem, S \Z is a smooth immersed

submanifold of codimension 3 in L2 .
Proof. Indeed, suppose by contradiction that ∇Φ1 , ∇Φ2 , ∇Φ3 are linearly de-
pendant at α ∈ S, that is, there exists a ∈ lR3 , a 6= 0 s.t.

a1 cos(α(s)) + a2 sin(α(s)) + a3 = 0

for almost all s; then, by integrating, a3 = 0, therefore a1 cos(α(s))+a2 sin(α(s)) =

0 that means that α ∈ Z. See also §3.1 in [Klassen et al., 2004].

156. Geodesics in the manifold of curves

Note that S \ Z contains the (representation by continuous lifting of) all
smooth immersed curves.
The manifold S \ Z inherits a Riemannian structure, induced by the scalar
product of L2 ; (critical) geodesics may be prolonged smoothly as long as they
do not meet Z.
Even if S may not be a manifold at Z, we may define the geodesic distance
dg (x, y) in S as the infimum of the length of Lipschitz paths γ : [0, 1] → L2 going
from x to y and whose image is contained in S;3 since dg (x, y) ≥ kx − ykL2 , and
S is closed in L2 , then the metric space (S, dg ) is metrically complete.
We don’t know if (S, dg ) admits minimal geodesics, or if it falls in the same
category of the Grossman example 3.20.
157. Multiple measurable representations
We may represent any Lipschitz closed arc-parameterized curve ξ using a
measurable α ∈ S:
Definition 8.6 (Measurable lifting). let arc : S 1 → [0, 2π) be the inverse of
α 7→ (cos(α), sin(α)); arc() is a Borel function; then ((arc ◦ ξ 0 )(s) + a) ∈ S, for
a choice of a ∈ lR.
We remark that the measurable representation is never unique: for example,
given any measurable A, B ⊂ [0, 2π] with |A| = |B|,

α(s) + 2π1A (s) − 2π1B (s)

will as well represent ξ in S.

3 It seems that S is Lipschitz-arc-connected, so dg (x, y) < ∞; but we did not carry on a

detailed proof

50
 This implies that the family Aξ of measurable α ∈ S that represent the same
curve ξ is infinite. It may be then advisable to define a quotient distance dˆ
(see Defn. 3.26) as follows:

ˆ ξ 0 ) :=
d(ξ, inf d(α, α0 ) (22)
α∈Aξ ,α0 ∈Aξ0

where d(α, α0 ) = kα − α0 kL2 , or alternatively d = dg is the geodesic distance

on S.
159. Continuous vs measurable lifting – no rotation index
If ξ ∈ C 1 , we have an unique 4 continuous representation α ∈ S; but note
that, even if ξ, ξ2 ∈ C 1 , the infimum (22) may not be given by the continuous
representations α, α0 of ξ, ξ2 . Moreover there are rectifiable curves ξ that do not
admit a continuous representation α, as for example the polygons.
A problem similar to the above is expressed by this proposition

Proposition 8.7. For any h ∈ Z, the set of closed smooth curves ξ with rotation
index h, when represented in S using the continuous lifting (20), is dense in
S \ Z.
It implies that we cannot properly extend the concept of rotation index to
S.
The proof is based on this lemma.

Lemma 8.8. Suppose that ξ is not flat, let τ be one of the measurable liftings
of ξ. There exists a smooth projection π : V → S defined in a neighborhood
V ⊂ L2 of τ such that, if f ∈ L2 ∩ C ∞ then π(f ) is in L2 ∩ C ∞ .
This is the proof of both the lemma and the proposition.Fix α0 ∈ S \ Z.
Let T = Tα0 S be the tangent at α0 . T is the vector space orthogonal to ∇φi (α0 )
for i = 1, 2, 3. Let ei = ei (s) ∈ L2 ∩ Cc∞ be near ∇φi (α0 ) in L2 , so that the map
(x, y) : T × lR3 → L2
3
X
(x, y) 7→ α = α0 + x + ei yi (23)
i=1

is an isomorphism. Let S 0 be S in these coordinates; by the Implicit Function Theorem

(5.9 in [Lang, 1999]), there exists an open set U 0 ⊂ T , 0 ∈ U 0 , an open V 0 ⊂ lR3 , 0 ∈ V 0 ,
and a smooth function F : U → lR3 such that the local part S 0 ∩ (U 0 × V 0 ) of the
manifold S 0 is the graph of y = F (x).
We immediatly define a smooth projection π : U 0 × V 0 → S 0 by setting π 0 (x, y) =
(x, F (x)); this may be expressed in the original L2 space; let (x(α), y(α)) be the inverse
of (23) and U = x−1 (U 0 ); we define the projection π : U → S by setting
3
X
π(α) = α0 + x + ei Fi (x(α))
i=1

Then
3
X
π(α)(s) − α(s) = ei (s)ai , ai := (Fi (x(α)) − yi (α)) ∈ lR (24)
i=1

so if α(s) is smooth, then π(α)(s) is smooth. This ends the proof of the lemma.
4 Indeed, the continuous lifting is unique up to addition of a constant to α(s), which is

equivalent to a rotation of ξ; and the constant is decided by Φ1 (α) = 2π 2

51
 Let αn be smooth functions such that αn → α in L2 , then π(αn ) → α0 ; if we choose
them to satisfy αn (2π)−αn (0) = 2πh, then, by the formula (24), π(α)(2π)−π(α)(0) =
2πh so that π(αn ) ∈ S and it represents a smooth curve with the assigned rotation
index h.

8.1.2 A metric with explicit geodesics [Younes et al., 2008]

161. A metric with explicit geodesics [Younes et al., 2008]

A similar method has been proposed recently in [Younes et al., 2008].
We consider immersed planar curves and again we identify lR2 = C.I
Proposition 8.9 (Continuous lifting of square root). If ξ is immersed, then
ξ 0 is continuous and ξ 0 6= 0, so there exists a continuous function α : lR → CI
satisfying
ξ 0 (θ) = α(θ)2 (25)
and α(θ) is unique, up to multiplying by ±1.
Let e, f be the real and imaginary part of (1/2 of) the square root lifting so
that
1
ξ 0 = (e + if )2
2
R 2π
The condition that ξ is a closed curve translates into 0 (e + if )2 dθ = 0 (where
equality is in C),
I hence we have two equalities (for real and imaginary part)
Z 2π Z 2π
2 2
e − f dθ = 0, ef dθ = 0
0 0

while the condition that len(ξ) = 1 translates into

Z 2π
1 2π 1 2π 2
Z Z
0 2
|ξ | dθ = |e + if | dθ = (e + f 2 ) dθ = 1 .
0 2 0 2 0
With some algebra we obtain that the above conditions are equivalent to
Z 2π Z 2π Z 2π
e2 dθ = 1, f 2 dθ = 1, ef dθ = 0.
0 0 0

Let L2 = L2 ([0, 2π] → lR). Let St(L2 , 2) ⊂ L2 × L2 ,

 Z 2π Z 2π Z 2π 
2
St(L , 2) := (e, f ) | e2 dθ = 1 = f 2 dθ, ef dθ = 0.
0 0 0

be the Stiefel manifold of orthonormal pairs in L2 . St(L2 , 2) is a smooth

manifold, and inherits the (flat) metric of L2 × L2 . What is most surprising is
that
Theorem 8.10 (2.2 in [Younes et al., 2008]). Let δξ be a small deformation of
ξ, and δe, δf be the corresponding small deformation of the representation e, f .
Then Z Z 2π
|Ds δξ|2 ds = 2 δe2 + δf 2 dθ
ξ 0

that is (geometrical) Riemannian metric in M is mapped into the (flat and

parametric) metric in St(L2 , 2).

52
Proof. By ξ 0 = (e+if )2 /2 we obtain |ξ 0 | = |e+if |2 /2 and δξ 0 = (e+if )(δe+iδf )
so that Z 2π
|δξ 0 |2
Z Z
2
|Ds δξ| ds = 0
dθ = 2 |δe + iδf |2 dθ
ξ S 1 |ξ | 0

It is then natural to “embed” closed planar curves (up to translation and

scaling) into St(L2 , 2).
165. Curves up to rotation ↔ Grassmanian
We note that rotation of ξ by an angle τ is equivalent to rotation of the
frame (e, f ) by an angle τ /2. So the orbit of all rotations of ξ is associate
to the plane generated by (e, f ) in L2 . So the space of curves up to rota-
tion/translation/scaling is represented by the Grassmanian of 2-planes in L2 .
A theorem by Neretin then applies, that provides a closed form formula for
(critical) geodesics. See section 4 in [Younes et al., 2008].
166. Completions and quotients
Each smooth curve ξ is represented by two smooth frames (e, f ) and (−e, −f ),
(so the square root lifting inverts to a double covering of St on the immersed
smooth curves). Hence the space of smooth curves is isometric to St(C ∞ , 2)/{±1}
(a Fréchet manifold).
The Stiefel manifold St(L2 , 2) is a complete smooth Riemannian manifold
inside L2 × L2 ; and it contains the (representation of) all closed rectifiable
parametric planar curves, up to scaling and translation.
So with this choice of metric and representation, we “obtain” that the com-
pletion of the Fréchet manifold of smooth curves is a Riemannian smooth man-
ifold. 5
But there are two problems left.
• The quotient w.r.t reparameterization. This is studied in [Younes et al.,
2008], where it is numerically proven that unfortunately geodesics in the
quotient space may develop singularities in the reparameterizations at the
end points.
• But there is also the quotient wrt representation.
167. Quotient wrt representation
As in the case before (see Defn. 8.6), we may talk of a measurable lifting.
Again, there are many measurable liftings (even when c ∈ C 1 ): let (e, f ) be a
square root representation of ξ, that is ξ 0 = (e + if )2 /2; choose a : S 1 → {−1, 1}
arbitrarily (measurable); then (ae, af ) represents the same curve.
Moreover there are rectifiable curves ξ that do not admit a continuous rep-
resentation (e, f ), as for example the polygons.
So again we may define a quotient metric (see Defn. 3.26)
 
ˆ 1 , ξ2 ) :=
d(ξ inf d St (e 1 , f1 ), (ae 2 , af2 ) (26)
1 a:S →{−1,1}

that is the quotient w.r.t. different representations in St(L2 , 2).

If ξ ∈ C 1 , we have two continuous representations ±(e, f ); but note that,
even if ξ1 , ξ2 ∈ C 1 , the infimum (26) may not be given by the continuous
representations of ξ1 , ξ2 .
5 This subject has been studied in a recent thesis by S. Deiala.

53
8.2 Riemannian spaces of immersed curves
Metrics of “geometric” curves have been studied by Michor–Mumford [Michor
and Mumford, 2006, 2007, 2005] and Yezzi–Mennucci [Yezzi and Mennucci,
2005b, 2004a,b]; more recently, Yezzi–Mennucci–Sundaramoorthi [Sundaramoor-
thi et al., 2007b, 2006; Mennucci et al., 2008; Sundaramoorthi et al., 2005,
2008a, 2007a] have studied Sobolev–like metrics of curves and shown many good
properties for applications to Shape Optimization; similar results have also been
shown independently by Charpiat et al [Charpiat et al., 2004].
We now discuss some Riemannian metrics on immersed curves.

• H 0 (the simplest)
Z
hh1 , h2 iH 0 = hh1 (s), h2 (s)i ds
c

• [Michor and Mumford, 2006]’s metric

Z
hh1 , h2 iH A = (1 + A|κc |2 )hh1 , h2 i ds
c

• [Yezzi and Mennucci, 2004b] Conformal metric

Z
hh1 , h2 iHφ0 = φ(c) hh1 , h2 i ds
c

• [Sundaramoorthi et al., 2005] Sobolev type metrics

Z Z
hh1 , h2 iH n = hh1 , h2 i ds + len(c)2n h∂sn h1 , ∂sn h2 i ds
c c
DZ Z E Z
hh1 , h2 iH̃ n = h1 ds, h2 ds + len(c)2n h∂sn h1 , ∂sn h2 i ds
c c c

8.2.1 H 0 (traditional)

170. The H 0 Inner Product

D E Z
h1 , h2 := h1 · h2 ds
H0 c

• We already noted that the minimizing flows implemented in traditional

Active Contour methods are “gradient flows” only if we use the H 0 inner
product.

• The H 0 distance is identically zero (We wil prove this in sec. 14.1.)
• In section 14.6 we will see a possible result of existence of geodesics for
this metric, when restricted to curves with bounded curvature.

54
8.2.2 H A (Michor-Mumford)

171. H A Metric
[Michor and Mumford, 2006] propose the metric H A
Z L
hh1 , h2 iH A = (1 + A|κc |2 )hh1 , h2 i ds
|c
0

where κc (u) is the curvature of c(u), and A > 0 is fixed.

Properties. • The induced distance is non degenerate
• The completion M (intended in the metric sense) is

BV 2 ⊂ M ⊂ Lip

where BV 2 are the curves that admit curvature as a measure (see section
14.5) and Lip are the rectifiable curves.
• Compactness bounds (prevents the pulley example).

8.2.3 Conformal metrics

172. Conformal metrics [Yezzi and Mennucci, 2004b]

Suppose we change the metric, from H 0 to a conformal metric
Z
hh1 , h2 iHψ0 = ψ(c) hh1 , h2 i ds
c

where ψ(c) associates to each curve c a positive number.

Then the gradient descent flow of an energy E defined on curves C(·, t) is

∂C 1
= −∇ψ E(C) = − ∇E(C)
∂t ψ(C)

where ∇E(C) is the gradient for the H 0 metric.

This is equivalent in a change of time variable t in the gradient descent
flow. So all properties of the flows are unaffected if we switch from a H 0 to a
conformal– H 0 metric.
173. Properties of Length-Dependent Conformal Factors
Consider a conformal metric where ψ(c) depends (monotonically) on the
length L(c) of the curve:
• If ψ(c) ≥ L(c), then the generated metric is non degenerate
• unfortunately, according to a result by J. Shah, when ψ(c) ≡ L(c) very
few (minimal) geodesics do exist (only “grassfire” geodesics, moving by
constant normal speed)

55
8.3 Sobolev type Riemannian metrics
174. Sobolev Metrics [Sundaramoorthi et al., 2005]
Let Ds := |c10 | ∂θ be “the derivative with respect to arclength”. Let
Z
hh1 , h2 iH j := hDsj h1 , Dsj h2 i ds
0
S1

[Sundaramoorthi et al., 2005] consider Sobolev type metrics

Z
hh1 , h2 iH j = hh1 , h2 i ds + λL2j hh1 , h2 iH j
0
S1
DZ Z E
hh1 , h2 iH̃ j = h1 ds, h2 ds + λL2j hh1 , h2 iH j
0
S1 S1

(with λ > 0); where L := len(c).

Notice that we use integration by arc parameter. Notice also the scale
invariance of the norms, by means of L = len(c).
Pros: Some nice properties (later on).
Cons: Horizontality is a mess (needs pseudodifferential operators)
Notice that these metrics are geometric:
• they are easily seen to be invariant w.r.to rotations and translations, (in
a stronger sense than in frame 106, indeed in this case

hh1 , h2 iAc = hh1 , h2 ic , hRh1 , Rh2 ic = hh1 , h2 ic

for any Euclidean transformation A and rotation R);

• they are reparameterization invariant due to the usage of the arc param-
eter in derivatives and integrals;
• they are scale invariant, since the normalizing factors L2j make them
0-homogeneous w.r.to rescaling of the curve.
For this last reason, we redefine the H 0 metric to be
D E Z
h1 , h2 0 := h1 · h2 ds (27)
H c

so that it is again 0-homogeneous.This is a conformal version of the H 0 metric

defined previously so a gradient descent flow is a time reparameterization of the
flow for the original H 0 metric; and the induced distance is again degenerate.

8.3.1 Properties of H j metrics

In [Sundaramoorthi et al., 2005] it was shown experimentally that Sobolev flows
are smooth in the space of curves, are not as dependent on local image informa-
tion as H 0 flows, are global motions which deform locally after moving globally,
and do not require derivatives of the curve to be defined for region-based and
edge-based energies. We will see some numerical experiments in the following
lecture
178. H 1 -gradients

56
 Let again L = len(c). Recall the definition 3.23 of gradient ∇E by means of
the identity
h∇E, hic = DE(c)(h) ∀h ∈ Tc M .
Let f = ∇H 0 E, g = ∇H 1 E be the gradients wrt the inner products H 0 and H 1 ;
by the definition of gradient, we obtain that

hf, hiH 0 ,c = hg, hiH 1 ,c ∀h ∈ Tc M

that is Z Z
h · f ds = h · g + λL(Ds h · Ds g) ds ∀h
c c
with some integration by parts this becomes
Z
h · (f − g + λLDs2 g) ds = 0 ∀h
c

then we conclude that

∇H 0 E = ∇H 1 E − λLDs2 (∇H 1 E) .

With similar computation, for H̃ 1 we obtain that

Z
∇H 0 E = ∇H̃ 1 E ds − λLDs2 (∇H̃ 1 E)
c

Given ∇H 0 E, both equation can be solved for ∇H 1 E and ∇H̃ 1 E, by means of

suitable convolution kernels K̃λ , Kλ , that is, we have the formulas

∇H 1 E = ∇H 0 E ∗ Kλ , ∇H̃ 1 E = ∇H 0 E ∗ K̃λ .

The above idea extends to degree j SAC: it is possible to obtain ∇H j E and

∇H̃ j E from ∇H 0 E, by convolution.
The bad news is that, when Shape Optimization is implemented using Level
Sets Methods (as we do), we need to trace out the curve to compute gradients.
K̃λ , Kλ are known in closed form:
 L
s− 2
cosh √λL
Kλ (s) = √   , for s ∈ [0, L], (28)
2L λ sinh 2√1 λ
!
2
1 (s/L) − (s/L) + 1/6
K̃λ (s) = 1+ , s ∈ [0, L]. (29)
L 2λ

and Kλ , K̃λ are periodically extended to all of lR. See figure 3.

181. Smoothing of gradients
In [Sundaramoorthi et al., 2006] and [Sundaramoorthi et al., 2007a] it was
noted that the regularizing properties may be explained in the Fourier domain:
indeed, if we calculate Sobolev gradients ∇H n E of an arbitrary energy E in the
frequency domain, then

∇
\H 0 E(l)
∇
\H n E(l) = for l ∈ Z (30)
1 + λ(2πl)2n

57
 6 10
λ=0.01 λ=0.01
λ=0.04 λ=0.04
5 λ=0.08 8 λ=0.08

6
4

4
3
2

2
0

1
−2

0 −4
−0.5 0 0.5 −0.5 0 0.5

Figure 3: Plots of Kλ (left) and K̃λ (right) for various λ with L = 1. The plots
show the kernels over one period.

and
∇
\H 0 E(l)
∇
\
H̃ n E(0) = ∇H 0 E(0),
\ ∇
\
H̃ n E(l) = for l ∈ Z\{0}, (31)
λ(2πl)2n
It is clear from the previous expressions that high frequency components of
∇H 0 E(c) are increasingly less pronounced in the various forms of the H n gradi-
ents.
Note that we are smoothing gradients, so we will obtain smoother minimiza-
tion flows. But we are not smoothing the curves themselves.
182. Flow regularization
It is possible to show mathematically that the Sobolev–type gradients regu-
larize the flows of well known energies, by reducing the degree of the P.D.E., as
shown in this example:
Example 8.11. in the case of the elastic energy E(c) = c κ2 ds = c |Ds2 c|2 ds,
R R
(3)
the H 0 gradient is ∇H 0 E = LDs (2Ds c+3|Ds2 c|2 Ds c) that includes fourth order
derivatives; whereas the H̃ 1 -gradient is

Lcss + L(|Dss c|2 Ds c) ∗ K̃λ,1 (32)

that is an integro-differential second order P.D.E.

(Other examples will be shown in the following lecture.)
183. Existence of flow for geodesic active contour
We consider the geodesic active contour model [Caselles et al., 1995][Kichenas-
samy et al., 1995].
We consider once again the energy E(c) = c φ(c(s)) ds where φ : R2 → R+ .
R

Then the gradient with respect to H 0 is

∇H 0 E = L(∇φ(c) · N )N − Lφ(c)κN

where N is the unit inward normal, and κ is the curvature.

We omit the argument in φ(c), for simplicity, and write c0 for arc parameter
derivation. Let us first note that ∇H 0 E = L∇φ − L(φc0 )0 . Integrating by parts
we find that
1
∇1 E = ∇φ ∗ K − (φc0 )0 ∗ K = ∇φ ∗ K − (φs c) ∗ K 0 − (φc) ∗ K 00 ,
L

58
where φs := (φ◦c)0 . Since K̃λ is the
fundamental solution of the gradient prob-
lem, that is, K 00 (s) = L1 L1 − δ(s) (that is the eqn. (18) in [Sundaramoorthi
et al., 2007a]), we find that
φc − avgc (φc)
∇H̃ 1 E = − L(φs c) ∗ K̃λ0 + L∇φ ∗ K̃λ . (33)
λL
The above does not require that the curve be twice differentiable, and then we
may prove that
1,1
Proposition 8.12. Suppose that φ ∈ Cloc (that is, φ ∈ C 1 and its derivative is
n
Lipschitz on compact subsets of lR ), let the initial curve be c0 ∈ C 1 ; then the
gradient flow
dc
= −∇H̃ 1 E(c)
dt
exists and is unique in C 1 for small times.
If lim inf x→∞ φ(x)|x| > 0, then it exists for all (positive) times.
Of particular interest is when φ = 1, that is E = L, the length of the curve.
In this case, we already know that the H 0 flow is the geometric heat flow.
We see that
c − avgc (c)
∇H̃ 1 L = .
λL
so H̃ 1 gradient flow constitutes a simple rescaling of the curve about its centroid
(!).
It is interesting to notice that the H 1 and H̃ 1 gradient flows are well-posed
for both ascent and descent while the H 0 gradient flow is only well-posed for
descent.
While the H 0 gradient descent flow smooths the curve, the H̃ 1 gradient
descent (or ascent) has neither a beneficial nor a detrimental effect on the reg-
ularity of the curve.

8.3.2 Proof of local ∃! of flows, (prop. 8.12)

Definition 8.13. Given a vector field h along the curve, let k = Pc h be the
primitive of h, that is the unique vector field with avgc (k) = 0 and such that
Ds k = h − avgc (h). If we express k = h ? KcP , then
s 1
KcP (s) := − + for s ∈ [0, L) (34)
L 2
and KcP (s) is extended periodically to s ∈ lR (note that KcP (s) jumps at all
points of the form s = nL, n ∈ Z).
In particular, the kernels defined in eqn. (28), (34) are related by
1 1
K̃λ = − K P ? KcP
L λL2 c
so we obtain the relations
1
Ds K̃λ = − KP
λL2 c
1 1
Dss K̃λ = − 2
δ0 + .
λL λL3

59
Proof. We rapidly sketch the proof of local existence and uniqueness (for more
details, look into section 10.6 in [Mennucci, 2009]).
We show that ∇H̃ 1 E(c) is locally Lipschitz in the C 1 Banach space (of para-
1,1
metric curves). Indeed, since φ ∈ Cloc , the maps

c 7→ ∇φ(c)
c 7→ φ(c)Ds c

are locally Lipschitz as maps from C 1 to C 0 , so (see Lemmas 10.25 and 10.26
in [Mennucci, 2009]) we obtain that

c 7→ avgc (∇φ(c))

is loc.Lip. from C 1 to lRn , and

c 7→ Pc Pc ∇φ(c)


c 7→ Pc φ(c)Ds c

are locally Lipschitz as maps from C 1 to C 1 ; combining all together (and using
the Lemma 10.25 again)
1
1

∇H̃ 1 E(c) = Lavgc (∇φ(c)) − Pc Pc ∇φ(c) + Pc φ(c)Ds c
λL λL
is locally Lipschitz as maps from C 1 to C 1 ; so by the Cauchy–Lipschitz theorem
we know that the gradient flow exists for small times.
When φ ∈ C 1 , φ > 0, it is nonetheless possible to prove that the flow exists
for small (positive) times.

Proof. The proof is based on the inequality

1 1
|∇H̃ 1 E(s)| ≤ Mr (1 + r) + LMr r + LMr
λ 2λ
and
1 1
E(c) ≤ L ≤ E
Mr mr
where
Mr := sup (max{φ(x), |∇φ(x)|}) and mr := inf φ(x) ;
|x|<r |x|<r

the above (by using the Euler method) implies existence, in the class of curve
evolutions c(θ, t) that are Lipschitz in both variables.
We propose the trace of the proof, in case λ > 24, for H̃ 1 flow.
Let B(0, r) ⊂ Rd be the ball centered in 0 of radious r > 0.
Let Mr > 0 be a supremum to φ and |∇φ| in B(0, r).
Since
L inf φ(x) ≤ E(c) ≤ L sup φ(x)
|x|<r |x|<r

then
1 1
E(c) ≤ L ≤ E
Mr mr

60
 By using the Poincaré–type inequality (8) in [Sundaramoorthi et al., 2007a]
we prove that
Z Z Z
φc − avgc (φc) ≤ |(φc) | ds ≤ (∇φ · c )|c| ds + φ|c0 | ds ≤ Mr (1 + r)L
0 0

moreover Z L
1
|K̃λ0 | ds =
0 2λ
then, for all s,
1 1
|∇H̃ 1 E(s)| ≤ α(r, L) := Mr (1 + r) + LMr r + LMr
λ 2λ
Suppose now that the curve c(·, t) is contained in B(0, R) with R = R(t):
then R0 (t) ≤ α(R(t), L);
We use the Euler method, in the class of curve evolutions c(θ, t) that are
Lipschitz in both variables. We pick a N ; let tj := j2−N ; let so that, defining

c(θ, tj+1 ) := c(θ, tj ) − 2−N ∇H̃ 1 E(c(·, tj ))(θ)

for N big enough, we will have that the energy E(c(θ, tj ) is decreasing in j; so
L will be bounded from above, and then

c(θ, t) − c(θ, s) ≤ α

We let N go to infinity, and apply Ascoli-Arzelà, to obtain a solution defined

for all t = j2−N .

8.3.3 Mathematical properties

186. Mathematical properties [Mennucci et al., 2008]

Proposition 8.14. The H j and H̃ j distances are equivalent:
s
1 + (2π)2j λ
dH̃ j ≤ dH j ≤ dH̃ j
(2π)2j λ

whereas dH̃ j ≤ dH k for j < k.

Proposition 8.15. The H j and H̃ j distances are lower bounded (with appro-
priate constants depending on λ) by the Fréchet distance (defined in 7.1).
Proposition 8.16. c 7→ len(c) is Lipschitz in M with H j metric, that is,

| len(c0 ) − len(c1 )| ≤ dH j (c0 , c1 )

Theorem 8.17 (Completion of Bi w.r.t. H 1 ). let dH 1 be the distance induced

by H 1 ; the metric completion of the space of curves is equal to the space of all
rectifiable curves.
Theorem 8.18 (Completion of Bi w.r.t. H 2 ). Let E(c) := |Ds2 c|2 ds
R

Then E is locally Lipschitz in M w.r.t. dH 2

As a corollary the completion of C ∞ according to the metric H 2 is the
space of curves that admit curvature κ ∈ L2 (S 1 )

61
 188. Related works
A family of metrics similar to H j above (but for the length dependent scale
factors) was studied in [Michor and Mumford, 2007]: the Sobolev-type weak
Riemannian metric on Imm(S 1 , R2 )
Z X n Z
hh, kiGnc = hDsi h, Dsi ki ds = hAn (h), ki ds where
S 1 i=0 S1
n
X
An (h) = An,c (h) = (−1)i Ds2i (h)
i=0

in that paper the geodesic equation, horizontality, conserved momenta, lower

and upper bounds on the induced distance, and scalar curvatures are computed.
Note that this norm is locally equivalent to the norms in slide 174.

8.4 A new Sobolev-type Riemannian metric

We will now see a model that was proposed and studied in [Sundaramoorthi
et al., 2011]. It is a Sobolev-type Riemannian Geometry of immersed curves.
190. Gâteaux derivative
The Gâteaux derivatives of the centroid avgc (c) and of the length L(c) of
a curve are
Z
D(avgc (c))(c; h) = h + (c − avgc (c))hDs h, Ds ci ds ∈ R2
Zc
DL(c)(c; h) = Ds h · Ds c ds .
c

(By Prop. 1.9).

191. tangent decomposition
To define the metric H, we first define the following decomposition for c ∈ M
and h ∈ Tc M :
h = ht + hl (c − avgc (c)) + hd (35)
The components ht and hl of h are defined as
Z
ht = D(avgc (c))(c; h) = h + (c − avgc (c))hDs h, Ds ci ds ∈ R2 (36)
c
Z
l
h = D(log L(c))(c; h) = Ds h · Ds c ds ∈ R . (37)
c

and then hd = h − ht − hl (c − avgc (c)).

The component hd deforms the curve without scaling or translating, since
D(L(c))(c; hd ) = 0 , Davgc (c)(c; hd ) = 0 .

192. Metric
Definition 8.19. If h, k ∈ Tc M are decomposed as above, then we define the
Riemannian metric H as
Z
. t t
hh, kiH = h · k + λl h k + λd Ds hd · Ds k d ds,
l l
(38)
c

62
where the first two products are the Euclidean dot products, the last term is a
normalized geometric Sobolev metric, and λl , λd > 0 are (constant) weights.
The third term of the metric may be rewritten directly as a function of
h ∈ Tc M by using the identity
Z Z Z Z
d d
Ds h · Ds k ds = Ds h · Ds k ds − Ds h · Ds c ds Ds c · Ds k ds . (39)
c c c c

194. Md
We will find it useful to define a submanifold Md of M .
Definition 8.20. Let

Md = {c ∈ M : L(c) = 1, avgc (c) = 0} , (40)

that is the space of all smooth immersed curves with unit length and with
centroid at the origin.
This is a smooth submanifold of M , when modeled on C ∞ (the proof follows
from a corollary of Nash–Moser theorem, see [Hamilton, 1982]).
Its tangent space at ξ ∈ M is

Tξ Md = {h ∈ Tξ M | D(avgc (c))(c; h) = 0, DL(c)(c; h) = 0} . (41)

195. Isometry
Theorem 8.21. Let λl = λd = 1 for simplicity. We define a map and its
inverse
 
c − avgc (c)
c∈M 7→ v = avgc (c) , l = log L(c) , ξ =
L(c)
with (v, l, ξ) ∈ R2 × R × Md
2
(v, l, ξ) ∈ lR × lR × Md 7→ v + el ξ ∈ M

This map is a Riemannian isometry, when we associate to Md the metric

Z
hh, kiMd = Ds h · Ds k ds (42)
ξ

196. Computing Geodesics in M

The isometry in Thm. 8.21 implies that to compute a minimal geodesic in
M between c0 and c1 , we apply the following procedure:
1. define
. c0 − avgc (c0 ) . c1 − avgc (c1 )
c̃0 = , c̃1 = ;
L(c0 ) L(c1 )

2. compute a geodesic C̃ between c̃0 and c̃1 in the space Md

3. rebuild the geodesic in M

C(t, ·) = L1−t (c0 )Lt (c1 )C̃(t, ·) + (1 − t)avgc (c0 ) + tavgc (c1 ) .

63
Corollary: along a geodesic of curves, the centroid translates uni-
formly, and the scale/length changes exponentially.
So we have to compute geodesics in Md , that is, in St(L2 , 2).
197. Square root lifting
Since the space Md is the same studied in [Younes et al., 2008] (and with
the same metric) we again use the square root lifting 8.9.
Let e, f be real and imaginary part of the square root lifting, so that
1
ξ 0 (θ) = (e + if )2 .
2
We map ξ ∈ Md to (e, f ) ∈ St(L2 , 2) the Stiefel manifold of orthonormal
pairs in L2 = L2 ([0, 2π] → lR).
 Z 2π Z 2π Z 2π 
St(L2 , 2) := (e, f ) | e2 dθ = 1 = f 2 dθ, ef dθ = 0
0 0 0

2 2
with the induced Riemannian metric from L × L ; and again by Prop. 8.10 this
is an isometry.
198. Computing Geodesics in St(L2 , 2)
n
St(R , p) is the set of n × p matrixes with orthonormal columns.
Proposition 8.22 (Lippert in [Edelman et al., 1998]). Let Y : [0, 1] → St(Rn , p)
be a path, then the geodesic equation is

Ÿ + Y (Ẏ T Ẏ ) = 0.

The solution is
 
A −S
(Y (t)eAt , Ẏ (t)eAt ) = (Y (0), Ẏ (0)) exp t
I A

where A = Y T Ẏ , S = Ẏ T Ẏ , and I is the p × p identity matrix.

Same result holds for St(L2 , p)!

199. A complete Riemannian manifold of “curves”

The Stiefel manifold St(L2 , 2) is a complete smooth Riemannian manifold; it
contains the (representation of) all closed rectifiable parametric planar curves,
up to scaling and translation. Moreover
Theorem 8.23 ([Harms and Mennucci, 2010]). Any two points in the Stiefel
manifold St(L2 , p) can be connected by a minimal geodesic.
(and by similar tecniques, the Karcher mean of shapes does exist).
201. Quotients
But there are two problems left.
• The quotient w.r.t reparameterization. This is exactly as in [Younes et al.,
2008], since, by Thm. 8.21, reparameterizations act only on the “deforma-
tion component” Md .
• But there is also the quotient wrt representation (exactly as we saw in
slide 167).

64
 202. Shape Space(s)
This is the scheme of Shape spaces (of smooth curves) and their quotients.
lR ⊗ lR2 ⊗ Md o / lR ⊗ lR2 ⊗ St(C ∞ , 2)/{±1} immersed curves
 
Md o / St(C ∞ , 2)/{±1} immersed curves up to scale,
translation
  immersed curves up to scale,
Md /Rot o / Gras(C ∞ , 2)/{±1}
translation, rotation
  geometric curves up to scale,
Md /Rot/Diff(S 1 ) o / Gras(C ∞ , 2)/{±1}/?
translation, rotation
(Note that, to ensure that all quotients are manifolds, only freely immersed
curves should be considered)

8.5 Some final remark (trade-off; quotient order)

203. Trade-off
What is good for Shape Optimization may be bad for Analysis, and viceversa.
To explain the rationale, suppose H, L are two metrics, H being stronger than
L.
• When evolving the shape A, the L–flow will move to- L
B
wards a shape B with small scale deformations (such
as those induced by noise), since B is nearer to A in
A
the the L–induced distance; whereas the H–flow will
move towards the shape C with large scale deforma- H C

tions, since C is nearer to A w.r.t its related distance.

• Suppose now though that a dataset contains a template shape A; an algo-

rithm is given a version B of A that was corrupted by noise, and different
shape C, and it has to decide what is the best match to A. In this case,
the weaker metric L would associate A to the correct shape B, whereas
an algorithm using the metric H would fail to associate A to B.

204. Quotient order

It is interesting to note this fact.
When we started talking of geometric curves, we proposed the quotient B =
M/Diff(S 1 ) (the space of curves up to reparameterization); and this had to be
followed by other quotients wrt euclidean motions and/or scaling, to obtain
M/Diff(S 1 )/E(n)/lR+ .
In the end, though, the space B happens to be more difficult to study; hence,
in practice, most Shape Space theories prefer a different order: the quotient
M/E(n)/lR+ is modeled and studied first; then the quotient M/E(n)/lR+ /Diff(S 1 )
is performed (often numerically).
Start of lecture: V. Applications & Experiments:
Sobolev Active Contours

65
Part IV
Sobolev Active Contours
9 Sobolev Active Contours
Most of what follows was originally presented in [Sundaramoorthi et al., 2007a]
In this part we present the Sobolev Active Contours, with Applications &
Experiments.
206. The Important Properties of Sobolev Active Contours

• PDE order reducing property: Sobolev gradient flows are lower order than
H 0 (we already saw an example, and more to come)
• Coarse-to-fine deformation property
– automatically favors coarse-scale deformations before moving to fine-
scale deformations
– ideal for visual tracking
• Sobolev-type norms induce well-defined Riemannian metrics on space of
curves [Michor and Mumford, 2007],[Mennucci et al., 2008]

– First work to offer consistent theory of shape optimization and shape

analysis

9.1 Gradients
We now add some more comments on the gradient ∇E(c) that was defined in
3.23
207. Interpretation of the Gradient

• Property of the gradient:

Proposition 9.1. If DE(c) 6= 0, the gradient ∇E(c) is the vector v in

Tc M that provides the maximum in

DE(c)(v) |DE(c)(h)|
= sup . (43)
kvkc h∈Tc M \{0} khkc

• Thus, gradient is the most efficient perturbation, i.e., maximizes

change in energy by moving in direction h
cost of moving in direction h

• By constructing k·kc to penalize “unfavorable” directions, we have control

over the path taken to minimize E without changing E.

In a sense, the focus of research for 20 years has been on the numerator in
eqn. (43) – whereas we now focus on the denominator.

66
9.2 Why H̃ j is better than H j
208. H̃ 1 : Easier gradients
Let E(c) be an energy of curves. (Let λ = 1, len(c) = 1 for simplicity.)
Gradients are implicitely define by:

DE(c)(h) = hh, ∇H 0 EiH 0 = hh, ∇H 1 EiH 1 = hh, ∇H̃ 1 EiH̃ 1 ∀h

As we saw in slide 178, the above leads to the ODEs

Norm ODE
H1 ∇H 0 E = ∇H 1 E − Ds2 (∇H 1 E)
H̃ 1 ∇H 0 E = avgc (∇H̃ 1 E) − Ds2 (∇H̃ 1 E)
The second is much easier to solve.
We can decompose any h ∈ Tc M as h = avgc (h) + h̃, that is

Tc M = lR2 ⊕ Dc M

with n o
Dc M := h : S 1 → lR2 | avgc (h) = 0 .

Let f = ∇H 0 E, g = ∇H̃ 1 E; decompose f = avgc (f ) + f˜, g = avgc (g) + g̃.

To solve
f = avgc (g) − Ds2 g
you solve two equations:
avgc (g) = avgc (f ) Ds2 g̃ = −f˜
lR2 ⊕ Dc M

To solve the rightmost, simply integrate twice! (see next slide).

210. Close form formula for H̃ 1 gradient

• H̃ 1 gradient need not be computed as a convolution:

Z s
1
g(s)= g(0) + sg 0 (0)− 2 (s − ŝ)(f (ŝ) − avgc (f )) dŝ
λL 0
Z L
1
g 0 (0) = − 3 s(f (s) − avgc (f )) ds
λL 0
Z L
g(0) = f (s)K̃λ (s) ds
0

where g = ∇H 1 E and f = ∇H 0 E
• nearly same computational speed as H 0 gradient; moreover the result-
ing flow is more stable, so it works fine with larger choice of time step
discretization. For this reason, H̃ 1 Sobolev Active Contours are very fast.

211. Regularization of energy vs regularization of flow/metric

Imagine an energy E minimized on curves (numerically sampled to n points).
Suppose it is not satisfactory in applications (ill posed, or not robust to noise).
Two solutions

67
 • Add a regularization term E(c) + εR(c) and minimize this by H 0 gradient
descent. The numerical complexity of the ∇H 0 R is of order n.

• Minimize E(c) and minimize this by H̃ 1 gradient descent. The numerical

complexity of ∇H̃ 1 E is in solving the Ds2 g̃ = −f˜, that is of order n.
The numerical complexity is similar; but in the second case we are minimizing
the original energy.

9.3 Experiments
212. Sobolev Norms Robust to Noise and Local Minima
Region-Based
Z Segmentation:
Z E =Chan-Vese, + regularization for H 0 flows.
E(c) = (I − u)2 dA + (I − v)2 dA
cin cout

−∇H 0 E + ακN

−∇H 0 E + βκN
(β >> α)

−∇H̃ 1 E

213. Experimental Demonstration of “Local”

1. Initialize contour in noisy image. 1.

2. Run H 0 gradient flow on energy

E(c) = Ecv (c) + L(c),

where Ecv is the Chan-Vese energy. Call

converged contour c0 . 2.

3. Adjust c0 at one sample point by one

pixel; call modification ĉ0 . Note: ĉ0 is an
H 0 (also “rigidified”) local perturbation of
c0 .
3.
4. Run and compare H 0 , “rigidified,” and
Sobolev gradient flows initialized with ĉ0 .
Results are the following:

68
 . . . . . . . . . . . . . . . . . . . . . . . . time → . . . . . . . . . . . . . . . . . . . . . . . .

H0

. . . . . . . . . . . . . . . . . . . . . . . . time → . . . . . . . . . . . . . . . . . . . . . . . .
“Rigidified”
(Translation favored)

. . . . . . . . . . . . . . . . . . . . . . . . time → . . . . . . . . . . . . . . . . . . . . . . . .

Sobolev

Note: The same result holds for Sobolev even without perturbing the local
minimum (due to numerical noise).
Numerically, these local minima no longer exist!!!
215. Experiment for Tracking Benefits
Region-based energy using means and variances.
H 0 active contour

Sobolev active contour

216. Another Experiment for Tracking Benefits

H 0 active contour

“Rigidified” active contour

Sobolev active contour

69
10 New Applications With Sobolev Active Con-
tours
Most of what follows was originally presented in [Sundaramoorthi et al., 2007b]
and [Sundaramoorthi et al., 2008a]

10.1 A Few Useful Energies Precluded By H 0

217. Energies Precluded by H 0 Optimization
• There are many useful active contour energies that cannot be optimized
using H 0 gradient flow.
– some energies result in ill-posed H 0 flows
– some energies result in high-order PDE and are difficult to implement
numerically
• In many cases, we can optimize these energies using Sobolev active con-
tours and avoid ill-posed problems and reduce order of PDE.

218. Examples of Energies Precluded by H 0

Average Weighted Length:
Z
1
E1 (c) = φ(c(s)) ds = avgc (φ)
L c
∇H 0 E1 (c) = [∇φ · N − κ(φ − avgc (φ))]N

Scaled Weighted Area:

Z
1 Aφ
E2 (c) = 2 φ(x) dA(x) =
L c
L2
 in 
Aφ φN κN
∇H 0 E2 (c) = +2
L2 Aφ L

• Both energies have ill-posed H 0 gradient descents (backward heat flow compo-
nents arising from length increasing terms).
• Increasing length is well-posed with Sobolev gradient descent.
Notation: φ : R2 → R, κ = signed curvature, N = inward normal vector.

R219. Analysis of Average Weighted R Length

Let Lφ = c
φ ds and L = len(c). Let E 1 (c) = c
φ ds = Lφ /L, then the
gradient is
1 Lφ
∇E1 = ∇Lφ − 2 ∇L
L L
we already computed H 0 and H̃ 1 gradients of both L and Lφ ; in the case of H 0
gradient descent flow, the second term is ill posed, since Lφ /L2 > 0 and ∇H 0 L
is the driving term in the geometric heat flow.
To calculate Sobolev gradient descent, plug the gradients in ∇H̃ 1 E1 = L1 ∇H̃ 1 Lφ −
Lφ
L2 ∇H̃ 1 L to obtain

c − avgc (c)
∇H̃ 1 E1 = E1 − (φcs ) ∗ Ks − ∇φ ∗ K
L

70
(where cs = Ds c, css = Dss c).
The second term, which was the cause of ill-posedness of the H 0 gradient
c−avgc (c)
descent, is converted to E1 L , which is a stable rescaling of curve.
220. Another Example of Energy Precluded by H 0
Weighted Scale-Invariant Type Elastic Energy:
Z
E(c) = L φ(c(s))κ2 (s) ds
c
∇H 0 E(c) = −Ecss + 2L2 ∂ss (φcss ) + 3L2 ∂s (φκ2 cs ) + L2 κ2 ∇φ
E 2
∇H̃ 1 E(c) = − 2 (avgc (c) − c) + (avgc ((φcss )) − φcss )−
λL λ
− 3L2 K 0 ∗ (φκ2 cs ) + L2 K ∗ (κ2 ∇φ).
Notation: φ : R2 → R, κ = signed curvature, cs = Ds c, css = Dss c.
• If φ = constant, then the energy is scale-invariant.
• Fourth order H 0 gradient descent (numerically difficult to implement; level
set methods?)
• Sobolev gradient descent is second order; numerically easier to implement

10.2 Segmentation Applications for New Energies

221. Elastic Regularization
Typically, in active contour energies one adds a length penalty to obtain regu-
larity of the active contour:
E(c) = Edata (c) + L(c).

• Regularity obtained since the H 0 gradient flow of length is curvature flow.

• We are partly relying on the H 0 norm’s property to obtain regularity.
• In addition to smoothing, length of curve is restricted.
Consider an alternative approach for regularity of curve:
Z
E(c) = Edata (c) + L(c) κ2 (s) ds
c

• Energy favors regularity of curve; not relying on metric properties

• Regularization is scale-invariant

222. Elastic Regularization

. . . . . . . . minima for α increasing → . . . . . . . .

E(c) = Edata (c) + αL(c)

(H 0 gradient descent)

71
 E(c) = Edata (c)
Z
+ αL(c) κ2 (s) ds
c

(Sobolev gradient descent;

cannot use H 0 .)
223. New Edge-Based Active Contour Models R
Traditional edge-based active contour energy is Eold (c) = c φ ds: has a
length shrinking effect, which leads to undesirable local minimum.

• Consider a non-shrinking edge-based model:

Z
Enew (c) = φ(c(s)) L−1 + αLκ2 (s) ds


c

– first term is edge-detection without shrinking bias (nor regularity)

– second term is elastic regularity; relaxed near edges

• Consider a length increasing edge-based model. Maximize:

Z Z
Einc (c) = φ(c(s)) ds − α κ2 (s) ds
c c

where φ is high near edges.

224. Non-Shrinking Edge-Based Model

Initial Eold , H 0 Eold , Sobolev Enew , Sobolev

Z Z
φ(c(s)) L−1 + αLκ2 (s) ds


Eold (c) = φ(c(s)) ds, Enew (c) =
c c

225. Length-Increasing
Z Edge-Based
Z Model
Length Increasing Model: Einc (c) = φ(c(s)) ds − α κ2 (s) ds Typical Edge-
Z c Z c

based balloon model: Ebal (c) = φ(c(s)) ds − α φ dA

c R

72
Left to right: initial contour, minimizing Ebal α = 0.2, 0.25, 0.4 using H 0 , and increas-
ing weighted Einc α = 0.1 using Sobolev. The contour expands to enclose the entire
image (fifth image).

11 A new Sobolev-type Riemannian metric

11.1 Dynamical models
226. Review of Linear Filtering in Rn
Mesurement
System Noise Noise

Input x y
System Mesurement

e
Error

System x̂ Measurement ŷ
Model Model

x state, u input, η system noise, y measurament, ξ measurament noise.

Model (
xk = Axk−1 + Buk + ηk system
yk = Cxk + Duk + ξk measurament
Luenberger observer
 −

 x̂k = Ax̂k−1 + Buk predictor
ŷ = C x̂− + Du

measurament
k k k


 e k = y k − ŷk error
x̂k = x̂− Le estimator

k + k

If dk := xk − x̂k then dk = (I − LC)Adk−1 + ηk + L(ξk + µk ).

228. Defining a Simple Dynamical Model
We generalize the usual model in Rn to a dynamical model on M .
The state is x = (p, v) with p ∈ M position, and v ∈ Tp M velocity.

 
pk = pk−1 + vk−1 pk = exppk−1 (vk−1 ) ∈ M
→
vk = vk−1 + ηk−1 vk = Pk−1 (vk−1 + ηk−1 ) ∈ Tpk M

73
(noise affect only velocity; no input, that is, u = 0; Pk−1 is parallel transport
along the geodesic related to exppk−1 (vk−1 )).
The measurement model becomes

yk = p k + ξk → yk = exppk (ξk ) ∈ M

(we measure only position).

The model and measurement noise is now ξk , ηk ∈ Tpk M .
229. Defining the Predictor/Estimator
The predictor becomes
 − ( −
p̂k = p̂k−1 + v̂k−1 p̂k = expv̂k−1 (p̂k−1 )
→
v̂k− = v̂k−1 v̂k− = P̂k−1 (v̂k−1 ) ∈ Tp̂− M
k

The estimator becomes


 ek = yk − p̂−  ek = logp̂− (yk )


k k
−
p̂ = p̂k + Kp ek → p̂k = expp̂− (Kp ek )
 k k
v̂k = v̂k− + Kv ek = v̂k− + Kv ek

 v̂
k

(in the estimator we only update the velocity). The goal is to choose Kp , Kv > 0
so that the observer tracks the system.
230.
Comparison between frame-by-frame segmentation, i.e., no dynamics (left),
dynamical model only on the motion (scales and translation) in the middle col-
umn and dynamical model on both the motion and deformation (right column).
Red: p̂k , blue: v̂k , and green: yk .

74
 Frame=36 frame34 frame34
0 0 0

50 50 50

100 100 100

150 150 150

Frame=40 frame38 frame38

0
200 0
200 0
200

50 0 50 100 150 200 250 50 0 300 50 100 150 200 250 50 0 300 50 100 150 200 250 300

100 100 100

150 150 150

Frame=44 frame42 frame42

0
200 0
200 0
200

50 0 50 100 150 200 250 50 0 300 50 100 150 200 250 50 0 300 50 100 150 200 250 300

100 100 100

150 150 150

Frame=48 frame46 frame46

0
200 0
200 0
200

50 0 50 100 150 200 250 50 0 300 50 100 150 200 250 50 0 300 50 100 150 200 250 300

100 100 100

150 150 150

Frame=52 frame50 frame50

0
200 0
200 0
200

50 0 50 100 150 200 250 50 0 300 50 100 150 200 250 50 0 300 50 100 150 200 250 300

100 100 100

150 150 150

Frame=56 frame54 frame54

0
200 0
200 0
200

50 0 50 100 150 200 250 50 0 300 50 100 150 200 250 50 0 300 50 100 150 200 250 300

100 100 100

150 150 150

200 200 200

0 50 100 150 200 250 0 300 50 100 150 200 250 0 300 50 100 150 200 250 300

244. Stochastic minimization

75
Part V
Riemannian manifolds of curves
In this part we study some mathematical properties of the Riemannian mani-
fold of (freely)immersed curves, when it is endowed with the metrics that were
presented previously.
We recall the definitions: the projection onto the normal πN v = v − hv, T iT
and on the tangent πT v = hv, T iT seen in slide 144 on page 47
247. Reparameterization to normal motion
Proposition 11.1. Given any smooth homotopy C of immersed curves, there
exists a reparameterization given by a parameterized family of diffeomorphisms
Φ : [0, 1] × S 1 → S 1 , so that setting

C̃(v, θ) := C(v, Φ(v, θ))

we have that ∂v C̃ is orthogonal to ∂θ C̃ at all points; more precisely,

πT̃ ∂v C̃ = 0 , πÑ ∂v C̃ = πN ∂t C

(where the last equality is up to reparameterization Φ).

Proof. (taken from [Yezzi and Mennucci, 2004b, Thm. 3.10]). By direct compu-
tation,
πN
e ∂v C(v, θ) = πN ∂v C(v, τ )
e (44)
πTe∂v C(v,
e θ) = πT ∂v C(v, τ ) + Ċ(v, τ )∂v ϕ(v, θ) (45)
where τ := ϕ(v, θ). Both πT ∂v C and Ċ∂v ϕ are parallel to T : so, if ∂v ϕ = −h∂v C, T i/|Ċ|,
then πTe∂v C
e = 0. Let
h∂v C(v, τ ), T (v, τ )i
M (v, τ ) = −
|Ċ(v, τ )|
The O.D.E. 
∂v ϕ(v, θ) = M (v, ϕ(v, θ))
(46)
ϕ(θ, 0) = θ, θ ∈ S 1
can be solved for v ∈ [0, 1] (since M is bounded). Defining ψ = ∂θ ϕ = ϕ0 , ψ solves

∂v ψ(v, θ) = ψ(v, θ) ddτ M (v, φ(v, θ))
(47)
ψ(θ, 0) = 1, θ ∈ S 1

and then ∂θ ϕ > 0 at all times v.

249. Example
Suppose that c is an immersed C 1 closed curve. Let C(t, θ) = c(θ) + vt an
uniform translation of c. Let C̃ be as in the previous proposition.

C C’

76
 The dotted line represent the trajectory of a point on the curve.
The two motions C and C̃ coincide in B but are represented differently in
M.
Which one is best? Both “translation” and “orthogonal motions” seem a
good idea at first glance.

12 Horizontality
250. Recall: parameterization invariance
Let M be the manifold of (freely)immersed curves.p Let hh1 , h2 ic be a Rie-
mannian metric G, for h1 , h2 ∈ Tc M ; let khkc = hh, hic be its associated
R1
norm. Let E(γ) := 0 kγ̇(v)k2γ(v) dv be the action. Let C : [0, 1] × S 1 → S 1 be
a smooth homotopy, and γ(t) = C(t, ·) be the associate path. We recall these
two definition that we saw in slide 106. The metric G is
• curve-wise p.i. when the metric does not depend on the parameteri-
zation of the curve, that is kh̃kc̃ = khkc when c̃(t) = c(ϕ(t)) and h̃(t) =
h(ϕ(t));

• homotopy-wise p.i. for any ϕ : [0, 1] × S 1 → S 1 smooth, ϕ(t, ·) a diffeo

of S 1 for all fixed t, let C̃(t, θ) = C(t, ϕ(t, θ)), then E(C̃) = E(C).

251. Remark: on distance of curves

The standard distance d(c0 , c1 ) in M is the infimum of Len(γ) where γ
connects co , c1 ∈ M . — But we are interested in studying metrics and distances
in the quotient space B := M/Diff(S 1 ).
We suppose that the metric G is “curve-wise parameterization invari-
ant”; then G may be projected to B := M/Diff(S 1 ).
c0 Oc
γ’ Consider two geometrical curves [c0 ], [c1 ] ∈ B, and a path γ :
c1oφ’ [0, 1] → B connecting them.
Associate γ to a homotopy C : [0, 1] × S 1 → lR2 : C connects a
c0 to c1 ◦ φ1 .

Definition 12.1 (geometric distance). dB is the infimum of the length

Len(C) in the class of all homotopies C connecting the curve c0 to a repa-
rameterization c1 ◦ φ of c1 .
This implements the quotienting formula dB (x, y) = inf dM (gx, y) that we
g∈G
saw in slide 85; here G = Diff(S 1 ).
We are abusing notation: dB is not a distance in the space M (it is a
semidistance, the distance between c and a reparameterization c ◦ φ is zero.)

12.1 Horizontal and vertical space

253. Horizontal and vertical space
Consider a metric G (curve-wise p.i) on M . Let Π projection from M :=
Immf (S 1 ) to the quotient B = Bi,f = M/Diff+ (S 1 ).

77
 The orbit is Oc = [c] = {c ◦ φ | φ} = Π−1 ({c}).
(Orbits Oc are dotted, The vertical space Vc is the tangent to Oc :
spaces Wc and Vc are dashed.)

M Vc := Tc Oc ⊂ Tc M
Oc
that can be explicitly written as

Vc := {h = b(s)c0 (s) | b : S 1 → lR}

Wc i.e. all the vector fields h where h(s) is tangent to
c c.
The horizontal space Wc is the orthogonal com-
plement inside Tc M
Vc
Wc := Vc⊥ .
Note that Wc depends on G, but Vc does not.

M Oc

Wc
c

Vc

Π
B

[c]

Figure 4: The vertical and horizontal spaces, and the projection

The figure 4 may help in understanding. The whole orbit Oc is projected

to [c]. The vertical space Vc is the kernel of DΠc , so it is projected to 0. The
horizontal space Wc is isometric (thru DΠc ) with the tangent space T[c] B.
255. Gradients are horizontal
We noted in Prop. 4.11 that if E : M → lR is a reparameterization invariant
energy, then
DE(c; h) = 0 for any h = a.Ds c . (48)
This implies that the gradient is always horizontal.
256. Homotopy wise par. invariance =⇒ not a metric
Suppose that a Riemannian metric is [homotopy-wise invariance] (that
is, if C̃ is a reparameterization of C, then E(C̃) = E(C)).
Apply to: let C(t, u) = c(ϕ(t, u)) with ϕ(t, ·) be a time-varying family of
diffeos of S 1 . So E(C) = E(c) = 0, that is
Z 1
kc0 ∂t φk2 dt = 0
0
0
so kc ∂t φk = 0; but note that, by choosing appropriately φ, we may represent
any h ∈ Tc M that is tangent to c as h = c0 ∂t φ.

78
 Summarizing, the homotopy-wise invariance implies that, if h1 , h2 are tan-
gent to c at all points, then hh1 , h2 i = 0. p
So, the Riemannian metric is not a metric on M . ( hh1 , h2 i is a
seminorm).

12.2 From curve-wise p.i to homotopy-wise p.i

257. horizontally projected metric
The horizontally projected metric G⊥ is defined by

hh, kiG⊥ ,c := hh̃, k̃iG,c

where h̃, k̃ are the projections of h, k to Wc . Equivalently

khkG⊥ = inf kh + bc0 kG

b

1
for all choices of b : S → lR.
Proposition 12.2. G⊥ is homotopy-wise p.i.
Corollary 12.3. G homotopy-wise p.i. ⇔

khkG = kh + bc0 kG

for all b.
So we understand/study/design the metric G on B by:
1. choose a metric G on M that is curve-wise p.i
2. G generates the horizontal space W = V ⊥
3. project G on W to define G⊥ that is homotopy-wise p.i.

259. Horizontal G⊥ as length minimizer

To define the distance, we minimize the length of the paths γ : [0, 1] → M
that connect a curve c0 to a reparameterization c1 ◦ φ of the curve c1 .

c0 Oc
γ’
Wc is orthogonal to the orbits, Vc is tan-
c1oφ ’ gent to the orbits.
The path γ (that moves in some tracts
Wc γ
along the orbits, with γ̇ ∈ Vγ ) is longer
c
than the path γ 0 .
c1o φ

Vc
Lemma 12.4. 1. Let C̃(t, θ) = C(t, ϕ(t, θ)) where ϕ(t, ·) is a diffeomor-
phism for all fixed t. Minimize

min EG (C̃)
ϕ

79
 Then at the minimum C̃ ∗ , ∂t C̃ ∗ is horizontal at every point, and EG (C̃ ∗ ) =
EG⊥ (C̃ ∗ )
2. So the distance can be equivalently computed as the infimum of LenG⊥ or
of LenG ; and distances are equal, dG⊥ = dG .
(A more general Lemma can be proven for Finsler metrics). So the hori-
zontal G⊥ is inevitable.
The proof of the above lemma is based on the validity of “the lifting Lemma”:
Lemma 12.5 (lifting Lemma). Given any smooth homotopy C of immersed
curves, there exists a reparameterization given by a parameterized family of
diffeomorphisms Φ : [0, 1] × S 1 → S 1 , so that setting
C̃(t, θ) := C(t, Φ(t, θ))
we have that ∂t C̃ is in the horizontal space WC at all times t.
For the metric H 0 , this reduces to lemma 11.1 in this paper; for Sobolev-type
metrics, the proof is in §4.6 in [Michor and Mumford, 2007].
262. Horizontality according to H 0
The horizontal space Wc w.r.t. H 0 is
Wc := {h : h(s) ⊥ c0 (s)∀s}
The horizontal version of H 0 is
D E Z
h1 , h2 0,⊥ := πN h1 · πN h2 ds
H S1

where πN h1 (s) is the part of h1 (s) that is ⊥ c0 .

R
h · h2 ds
S1 1
on M
Good paradigm: think of R
π h
S1 N 1
· πN h2 ds on B
The same holds for the Finsler metrics F 1 and F ∞ that we saw in last lecture.
(In that case, we already presented the horizontally projected version.)
263. Practical consequences
Note that, if we move a curve along its tangent direction, we end up repa-
rameterizing it; and this part of the motion can be eliminated, as we saw in
Prop. 11.1. For this reason, in traditional Active Contours methods, the tan-
gent part of the flow is ignored and zeroed – and this couples well with Level
Set Methods.
“a good movement of a curve is
Effect on people:
a movement that is orthogonal to the curve.”
By saying this, you are “thinking H 0 ”. When using a different metric, the
above is not true anymore.
264. Horizontality according to H j ...
... is a mess. We need to solve
inf kh + bc0 kH j
b

in closed form, and this needs (pseudo)differential operators. See [Michor and
Mumford, 2007] for details.
But remember: in this case,

80
 “a good movement of a curve is not necessarily
a movement that is orthogonal to the curve.”

265. Horizontality is for any G

In the above we studied the horizontality properties of the quotient B =
M/Diff(S 1 ). But the theory works in general for any other quotient. So we may
apply the same study and ideas to further quotients (such as B/E(n)).

13 Momenta
266. The method of momenta
What follows comes from §2.5 in [Michor and Mumford, 2007].
Consider again the action of a group G on a Riemannian manifold M . Most
groups that act on curves are smooth differentiable manifolds themselves, and
the group operations are smooth. So they are Lie groups.
In our cases of interest, the actions are smooth as well.
We recall that the action satisfies the properties: e · m = m (with e ∈ G the
identity), and for any g, h ∈ G, m ∈ M

h · (g · m) = (h · g) · m . (49)

We want to study some differentiable properties of the action g · m; using the

rule (49), we see that we can restrict to studying e · m.
267. Conservation of momenta
Given a c ∈ M , and ξ ∈ Te G, we derive the action, for fixed c and e “moving”
in direction ξ; the result of this derivative is a ζ = ζξ,c ∈ Tc M (depending
linearly on ξ). This direction ζ is intuitively “the infinitesimal motion of c,
when we infinitesimally act in direction ξ on c”. Do not forget also that, since
ζ ∈ Tc M , then ζ = ζ(θ) : S 1 → lRn is a vector field.
We then recall a very interesting condition by Emmy Noether’s
Theorem 13.1. Suppose that the metric on M is invariant wrt the action of
G: this is equivalent to saying that the action is an isometry. Let γ(t) be a
critical geodesic. Then hζγ(t) , γ̇(t)i is constant in t.

268. Horizontal geodesics

As an alternative interpretation, note that the vectors ζ that you obtain by
deriving the action are exactly all the vectors in the vertical spaces Ve of the
corresponding action G (since ζ are infinitesimal motions inside the orbit).
Recall that h ∈ Tc M is horizontal (that is, h ∈ Wc ) iff it is orthogonal to Vc
hζ, ḣi = 0 for all ζ ∈ Vc .
So, as corollary of Emmy Noether’s theorem, we obtain that if a geodesic is
shot in a horizontal direction c0 (0), then the geodesic will be horizontal for al
subsequent times.
269. Examples of momenta

• The scaling is represented by lR+ , that is one dimensional, so there is only

one tangent direction ξ = 1 in lR+ ; the action is just l, c 7→ lc; so there is
only one ζ, that is ζ = c.

81
• The translation is represented by lRn , that is a vector space, so the tangent
directions are ξ ∈ lRn ; the action is just v, c 7→ v + c; then ζ = ξ (and is
constant in θ).

• The rotation group is represented by orthogonal matrixes, so we set G =

O(n); the group identity e is the identity matrix; the action is matrix-
vector multiplication A, c(θ) 7→ Ac(θ); the tangent Te G is the set of the
antisymmetric matrixes B ∈ lRn×n ; then ζ = Bc.
• The reparameterizations are φ ∈ Diff(S 1 ); a tangent vector is just a scalar
field ξ : S 1 → lR; the action is φ, c 7→ c ◦ φ; we in the end have that

ζ = ξ(θ)c0 (θ)

is a generic vector field parallel to the curve.

270. Momenta for the metric H̃ 1

1
For the H̃ metric, this implies the following properties of a geodesic path γ(t).
• Scaling momentum:
Z
hγ, γ̇iH̃ 1 = avgc (γ) · avgc (γ̇) + λL2 Ds γ · Ds γ̇ ds = constant.

• Translations momentum: avgc (γ̇) is constant, since, for all ξ ∈ lRn ,

hξ, γ̇iH̃ 1 = ξ · avgc (γ̇) = constant;

since ξ is arbitrary, this means that avgc (γ̇) is constant in t.

• Angular momentum: for any antisymmetric matrix B ∈ lRn×n ,

Z
hBγ, γ̇iH̃ 1 = (Bavgc (γ)) · avgc (γ̇) + λL2 (BDs γ) · (Ds γ̇) ds = constant.

• Reparameterization momentum: for any scalar field ξ : S 1 → lR, setting

ζ(θ, t) = ξ(θ)γ 0 (θ, t)

we get
Z
0
hζ, γ̇iH̃ 1 = avgc (ξγ ) · avgc (γ̇) + λL 2
Ds (ξγ 0 ) · Ds γ̇ = constant;

integrating by parts,
Z
avgc (ξγ 0 ) · avgc (γ̇) − λL2 (ξγ 0 ) · Dss γ̇ = constant;

since ξ is arbitrary, this means that

γ 0 · avgc (γ̇) − λL2 γ 0 · Dss γ̇

is constant in t, for any θ ∈ S 1 .

82
 272. Momenta for the metric H
Proposition 13.2 (Momenta, Prop.3.6 in [Sundaramoorthi et al., 2011]). Sup-
pose that C is a geodesic, then the following quantities are conserved:
• [translation] ∂t avgc (C) is constant;
• [scaling] ∂t (log L(C)) is constant;
• [rotation] the angular momentum (that may be expressed in two equal
ways) Z Z
(ADs C) · (∂t Ds C) ds = (ADs C) · (Ds ∂t C) ds (50)
C C
is constant, for any antisymmetric matrix A; when the curve is planar,
this means that Z
κ T · ∂t C ds
C
is constant;

• [reparametrization] (Ds C)·(Ds2 ∂t C) is constant in t, for any fixed θ ∈ S 1 .

14 Existence of Geodesics
We present some results on how the direct method in Calculus of Variations
may be applied to the problem of existence of geodesics.

14.1 The H 0 distance is degenerate

275. H 0 distance is degenerate.
It is possible to connect two segments c0 , c1 with a family of “zigzag” homo-
topies C k (t, θ) such that E(C k ) →k 0.

∂C
∂C ∂v
∂θ

Len(C) ∼ sin(α)

consequently the distance between any two curves is 0.RR

0
RR that the2H distance is degenerate. Let E(C) =
We now prove |∂t C|2 ds dt
and E⊥ (C) = |πN ∂t C| ds dt.
5 M.Mumford, Slides of the Gibbs Lectures [Mumford]

83
 We start with a zizzag homothopy C; now
t = 1/2
∂C
t = 1/4 ∂C ∂v
∂θ
t = 1/8

t=0 α

case t ∈ [0, 1/2]

1/k
t=1
|πN ∂t C|2 |∂θ C| =
t = 7/8 = |∂t C|2 |∂θ C| sin2 (α) ∼ sin(α)
t = 3/4
t = 1/2
If the angle α is small then E⊥ (C) < ε.
We smooth C just a bit to obtain Ĉ, so that E⊥ (Ĉ) <
case t ∈ [1/2, 1] 2ε.
We then apply Prop. 11.1 to Ĉ to finally obtain C̃, that moves only in the
normal direction, so E(C̃) < 2ε.

14.2 (Critical) geodesics for H j

Theorem 14.1 (4.3 in [Michor and Mumford, 2007]). Consider the Sobolev
type metric Z
hh, kiGc =
n hh, ki + hDsn h, Dsn ki ds
S1

with n ≥ 1. Let k ≥ 2n + 1; suppose c, h are in the (usual and parametric) H k

space (that is, c, h admit k-th (weak)derivative, that is square integrable). Then
it is possible to shoot the geodesic, starting from c and in direction h, for short
time.

14.3 The direct method

To prove that a Riemannian manifold admits minimal geodesics, we will study
the action E(γ) by means of methods in Calculus of Variations; we review some
basic ideas.
Let X be a topological space, endowed with a topology τ , and Ω ⊂ X, and
f : Ω → lR. A function f is lower semi continuous if, for all x ∈ X,

lim inf f (y) ≥ f (x) .

y→x

Note that, if the topology τ is not defined by a metric, then we can introduce a
different condition:

Definition 14.2. f is sequentially lower semi continuous if, for all x ∈ X,

for all xn → x,
lim inf f (xn ) ≥ f (x)
n

Definition 14.3. We define the sequential relaxation Γf of f on Ω to be

the function Γf : Ω → lR that is the supremum of all f 0 : Ω → lR that are
sequentially lower semicontinuous on Ω, and f 0 ≤ f in Ω. We have that, for all
x ∈ Ω,
ΓΩ f (x) = min {lim inf f (xn )}
(xn ),xn →x n

where the minimum is taken on all sequences xn converging to x, with xn , x ∈ Ω.

84
 f is sequentially lower semicontinuous, iff Γf |Ω = f .
If X is a metric space, then “sequentially lower semicontinuous functions” are
“lower semicontinuous functions”, and viceversa; so we may drop the adjective
“sequentially”.
279. The direct method
Definition 14.4. Consider again a function f : X → lR: it is called coercive
in X if ∀M ∈ lR,
{x ∈ X : f (x) ≤ M }
is contained in a compact set.
Proposition 14.5. If f is coercive in X and sequentally lower semi continuous
then it admits a minimum on any closed set, that is, for all C ⊂ X closed there
exists x ∈ C s.t.
f (x) = min f (y) .
y∈C

This result is one of the pillars of the modern Calculus of Variations. We will
see that unfortunately this result may not be applied directly to the problem at
hand.

14.4 Winding and pulley

280. Homotopy wise par. invariance =⇒ not coercive
Unfortunately, if the metric is independent of parameterization then the action
cannot be coercive.
Example 14.6 (Winding effect). Indeed, let Ck (t, θ) = c(θ + kt): the curve c is
winding upon itself; then Ck is not moving in the space of geometric curves, but
at the same time the family Ck is not compact in the usual spaces of immersed
curves.
It is then difficult to prove existence of geodesics, or of minima of energies
in active contours. By the above example, any such proof must start by some
unwinding.
281. Example: the pulley (la carrucola)
A similar example is the following pulley. It transforms the normal motion
of a curve into tangent motion (and viceversa).
D
C

F
E The thick line is the curve C5 (u, 0). The
thick arrows represent the normal part of
the velocity, while the thin dashed arrows
A
represent the tangent part of the velocity.
B
In the part DE (respectively F A) of the curve, vertical segments move apart
(resp., together) as the thick arrows indicate, with horizontal velocity with norm
1; as the curve unravels, it is forced to move also parallel to itself.
The generic curve movement Cj (θ, t) has 2j wheels. Let l = l(t, j) be the
distance from feature point D to feature point E. Then ∂t l → ∞ if j → ∞.

85
14.5 Curvature(s) as measures
282. Curvature H as a measure
Definition 14.7 (H as a measure). In general, we will say that a curve c ∈
W 1,1 (S 1 ) admits mean curvature in the measure sense if there exists a
vector valued measure H on S 1 such that
Z Z
T (s)∂s φ(s) ds = − φ(s) dH(s) ∀φ ∈ C ∞ (S 1 )
I I

that is Z Z
T (θ)∂θ φ(θ) dθ = − φ(θ) |c0 | dH(θ) ∀φ ∈ C ∞ (S 1 )
I I

See also [Simon, 1983], §7 and §16.

Note that this new definition is related to the usual definition we saw in 1.6:
when c ∈ C 2 , the measure is H = ∂s T · ds.
283. Curvature κ as a measure
If H is a measure, then k is a real valued measure defined by
n Z
X
κ(A) = Ni (θ) dHi (θ) .
i=1 A

Note that |H| = |κ|.

284. Example: polygons
This is a curvilinear polygon, oriented anticlockwise.

The curvature H is a measure; it has an absolutely continuous part and a

jump part. The absolutely continuous part is a vector valued function, that
is represented by the blue dashed arrows. The jump part is present at each
corner, where the tangent vector has a jump discontinuity, and the “derivative”
of the tangent vector is given by a delta vector, represented as a dash–dotted
red arrow.

14.6 Existence of geodesics for H 0

285. Semicontinuity
Let V, W ∈ lRn . Let W × V be the vector of all n(n − 1)/2 determinants of all
2 by 2 minors of the matrix having W, V as columns.

86
Definition 14.8. Consider a continuous function f : lRn×2 → lR such that

f (W, V ) = g(W, V, W × V )

where g : lRn(n+3)/2 → lR is convex: then f is polyconvex.6

Proposition 14.9. Let p ≥ 2, suppose f ≥ 0 and is polyconvex: by theorem
4.1.5 and remark 4.1.6 in [Buttazzo, 1989] then
ZZ
f (∂θ C, ∂t C) dθ dt

is W 1,p -weakly-lower-semi-continuous. 7

We immediatly remark the identity

|πW ⊥ V |2 |W |2 = |V |2 |W |2 − hV, W i2 = |V × W |2

that is easily checked. (In lR3 we have hV, W i = |V ||W | cos α and |V × W | =
|V ||W | sin α, where α is the angle between the two vectors).
So |πW ⊥ V |2 |W |2 is polyconvex; then
ZZ
|π∂θ C ⊥ ∂t C|2 |∂θ C|2 dθ dt

is weakly–lower–semicontinuous. This is similar to the horizontal version of the

action of H 0 , that is ZZ
|π∂θ C ⊥ ∂t C|2 |∂θ C| dθ dt .

Given an energy and two continuous curves c0 and c1 , we will try to find a
homotopy that minimizes this energy, searching in the class C.
287. Classes C and F
Definition 14.10 (class C). Let I = S 1 × [0, 1]. Let C be the class of all
homotopies C : I → lRn , continuous on I and locally Lipschitz in S 1 × (0, 1),
such that c0 = C(·, 0) and c1 = C(·, 1).
Definition 14.11 (class F of prescribed-parameter homotopies). We also fix
a measurable positive function l : [0, 1] → lR+ and restrict our attention to a
family F of homotopie such that

|∂θ C(θ, t)| = l(t)

for all θ, t.
In particular, len(C) = 2πl(t).
288. Semicontinuity in F
6 The more general definition and the properties may be found sec. 4.1 in [Buttazzo, 1989],

or in 2.5 and 5.4 in [Dacorogna, 1982]

7 Sequentially weakly-* if p = ∞

87
Proposition 14.12. Let p ≥ 2. The action E is W 1,p -weakly-lower-semi-
continuous in the class F.
More precisely: let
C ∈ F, (Ch )h ⊂ F
if Ch * C weakly in W 1,p , that is,
Ch → C, ∂v Ch * ∂v C ∂θ Ch * ∂θ C
in Lp , and
l(v) = |Ċ(θ, v)| = |Ċh (θ, v)|
then
lim inf E(Ch ) ≥ E(C)
h

Proof. We prove a more general statement. Let α > 0, β > 0, V, W ∈ lRn ,

define
e(W, V ) = |πW ⊥ V |α |W |β
and define Z
Eα,β (C) := e(∂θ C, ∂v C)
I
we prove that Eα,β is lsc. Note that E⊥ (C) is obtained by choosing α = 2, β = 1.
In general if β = 1 then Eα,β (C) is a geometric energy.
We prove the theorem in steps:
• Let λ > 0. Suppose |Ċ| ≡ |Ċh | ≡ λ; then
Z
β−α
Eα,β (C) = λ f (∂θ C, ∂v C)α

is l.s.c.
• for any homotopy C and any continuous g : [0, 1] → lR+ , define eg (C)(v) :
[0, 1] → lR+ ,

g(v)β−α S 1 f (∂θ C, ∂v C)α dθ if g(v) > 0
R
eg (C) :=
0 if g(v) = 0

If C ∈ F (that is l(v) = |Ċ(θ, v)|) then

Z 1
Eα,β (C) = el (C)dv
0

• Consider a piecewise function g ≥ 0 defined

m
X
g= gi χ[ai ,ai+1 ) (51)
i=1

and let Z 1 m Z
X ai+1
Ê(C) := eg (C)dv = egi (C)dv
0 i=1 ai

then we apply the previous reasoning to all addends and conclude that
lim inf Ê(Ch ) ≥ Ê(C)
h

88
 • Suppose l is continuous and l ≥ 0. Let τ be the class of piecewise functions
g ≥ 0 defined as in (51), such that on any interval [ai , ai+1 ), either gi = 0
or
gi ≥ sup l .
[ai ,ai+1 )

Then for any such g and C ∈ F,

Z 1
eg (C)dv ≤ Eα,β (C) .
0

• choose g in the class τ ; then

Z 1 Z 1
lim inf Eα,β (Ch ) ≥ lim inf eg (Ch )dv ≥ eg (C)dv
h h 0 0

• Fix C: it is possible to find a sequence (gj ) ⊂ τ such that

egj (C)(v) →j el (C)(v)
for almost all points v, monotonically increasing: indeed, let
Aj,i = [i2−j , (i + 1)2−j ), Ij,i = inf l(v), Sj,i = sup l(v)
v∈Aj,i v∈Aj,i

and 
0 if v ∈ Aj,i and Ij,i = 0
gj (v) =
Sj,i if v ∈ Aj,i and Ij,i > 0
Then Z 1
Eα,β (C) = sup egj (C)dv
j 0

289. Existence of geodesics

The above semicontinuity result may be coupled with a compactness results,
that we will not present for lack of space. The compacteness result enforces a
bound on the curvature, and this is aimed at controlling ∂θ C. Note that a
bound on the curvature limits the pulley effect (since each of the pulley wheel
is indeed adding to the curvature!).
In the most general form, those results may be used to prove existence of
geodesics when we restrict the geodesic to use curves with an upper bounded
total mass of the curvature. (But the math has still room for development – you
are welcome to join).
We present a simplified result.
290. Existence of geodesics
Let b > 0. Let S be the space of unit length curves s.t. the curvature κ of c
is bounded by b, i.e.
S := {c ∈ C 1,1 : |κ| ≤ b} .
Theorem 14.13 ([Yezzi and Mennucci, 2004b]). The Riemannian metric H 0
defines a positive geodesic distance in S: this distance admits minimal geodesics.
Unfortunately, since S has a border, then the minimal geodesic will not
satisfy the Euler-Lagrange ODE defined by E.
Probably geodesics travel on border {c : |κ| ≡ b}.

89
Afterword
291. Acknowledgments
I gratefully acknowledge that some part of the theory and all experiments ex-
posited were discussed and developed in joint work with Prof. Yezzi (GaTech)
and Prof. Sundaramoorthi (KAUST), and Dr. Duci.

References
C. J. Atkin. The Hopf-Rinow theorem is false in infinite dimensions. Bull.
London Math. Soc., 7(3):261–266, 1975. doi: 10.1112/blms/7.3.261.
A. Blake and M. Isard. Active Contours. Springer Verlag, 1998.

H. Brezis. Analisi Funzionale. Liguori Editore, Napoli, 1986. (italian translation

of Analyse fonctionelle, Masson, 1983, Paris).
G. Buttazzo. Semicontinuity, relaxation and integral representation in the cal-
culus of variation. Number 207 in scientific & technical. Longman, 1989.
V. Caselles, F. Catte, T. Coll, and F. Dibos. A geometric model for edge
detection. Num. Mathematik, 66:1–31, 1993.
V. Caselles, R. Kimmel, and G. Sapiro. Geodesic active contours. In Proceedings
of the IEEE Int. Conf. on Computer Vision, pages 694–699, Cambridge, MA,
USA, June 1995.

T. Chan and L. Vese. Active contours without edges. IEEE Transactions on

Image Processing, 10(2):266–277, February 2001.
G. Charpiat, O. Faugeras, and R. Keriven. Approximations of shape metrics
and application to shape warping and empirical shape statistics. Foundations
of Computational Mathematics, 2004. doi: 10.1007/s10208-003-0094-xgg819.
INRIA report 4820 (2003).
Guillaume Charpiat, Olivier D. Faugeras, and Renaud Keriven. Approxima-
tions of shape metrics and application to shape warping and empirical shape
statistics. Foundations of Computational Mathematics, 5(1):1–58, 2005.
Y. Chen, H. Tagare, S. Thiruvenkadam, F. Huang, D. Wilson, K. Gopinath,
R. Briggs, and E. Geiser. Using prior shapes in geometric active contours in
a variational framework. International Journal of Computer Vision, 50(3):
315–328, Dec 2002.
David L. Chopp and J. A. Sethian. Motion by intrinsic laplacian of curvature.
Interfaces and Free Boundaries, 1:1–18, 1999.

L. D. Cohen. On active contour models and ballons. Computer Vision, Graphics,

and Image Processing: Image Processing, 53(2), 1991.
D. Cremers and S. Soatto. A pseuso distance for shape priors in level set
segmentation. In IEEE Int. Workshop on Variational, Geometric and Level
Set Methods, pages 169–176, 2003.

90
Bernard Dacorogna. Weak Continuity and Weak Lower Semicontinuity of Non-
Linear Functionals, volume 922 of Lecture Notes in Mathematics. Springer–
Verlag, 1982.
Alessandro Duci and Andrea Mennucci. Banach-like metrics and metrics of
compact sets. 2007.
Alessandro Duci, Anthony J. Yezzi, Sanjoy K. Mitter, and Stefano Soatto.
Shape representation via harmonic embedding. In International Confer-
ence on Computer Vision (ICCV03), volume 1, pages 656 – 662, Washing-
ton, DC, USA, 2003. IEEE Computer Society. ISBN 0-7695-1950-4. doi:
10.1109/ICCV.2003.1238410.
Alessandro Duci, Anthony J. Yezzi, Stefano Soatto, and Kelvin Rocha. Har-
monic embeddings for linear shape. J. Math Imaging Vis, 25:341–352, 2006.
doi: 10.1007/s10851-006-7249-8.
A. Edelman, T. Arias, and S. Smith. The geometry of algorithms with or-
thogonality constraints. SIAM J Matrix Analy Appl, 20:303–353, 1998. doi:
10.1137/S0895479895290954. arXiv:physics/9806030v1 (1998).
J. Eells and K. D. Elworthy. Open embeddings of certain Banach manifolds.
Ann. of Math. (2), 91(3):465–485, May 1970. URL http://www.jstor.org/
stable/1970634.
Ivar Ekeland. The Hopf-Rinow theorem in infinite dimension. J. Differential
Geom, 13(2):287–301, 1978.
A. T. Fomenko. The Plateau problem. Studies in the development of modern
mathematics. Gordon and Breach, 1990.
M. Gage and R. S. Hamilton. The heat equation shrinking convex plane curves.
Journal of Differential Geometry, (23):69–96, 1986.
Joan Glaunès, Alain Trouvé, and Laurent Younes. Modeling planar shape vari-
ation via Hamiltonian flows of curves. In Anthony Yezzi and Hamid Krim,
editors, Analysis and Statistics of Shapes, Modeling and Simulation in Sci-
ence, Engineering and Technology, chapter 14. Birkhäuser - Springer - Verlag,
2005.
M. Grayson. The heat equation shrinks embedded planes curves to round points.
J. Differential Geometry, (26):285–314, 1987.
Nathaniel Grossman. Hilbert manifolds without epiconjugate points. Proc.
Amer. Math. Soc., 16:1365–1371, 1965. ISSN 0002-9939. doi: 10.2307/
2035933.
Richard S. Hamilton. The inverse function theorem of Nash and Moser. Bull.
Amer. Math. Soc. (N.S.), 7(1):65–222, 1982. ISSN 0273-0979.
Philipp Harms and Andrea Mennucci. Geodesics in infinite dimensional Stiefel
and Grassmann manifolds. 2010. http://cvgmt.sns.it/papers/harmen10/.
J. Itoh and M. Tanaka. The Lipschitz continuity of the distance function to the
cut locus. Trans. A.M.S., 353(1):21–40, 2000.

91
H. Karcher. Riemannian center of mass and mollifier smoothing. Communica-
tions on Pure and Applied Mathematics, 30:509–541, 1977.

M. Kass, A. Witkin, and D. Terzopoulos. Snakes: Active contour models. In-

ternational Journal of Computer Vision, 1:321–331, 1987.
S. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum, and A. Yezzi. Gradient
flows and geometric active contour models. In Proceedings of the IEEE Int.
Conf. on Computer Vision, pages 810–815, 1995.

Eric Klassen, Anuj Srivastava, Washington Mio, and Shantanu H. Joshi. Analy-
sis of planar shapes using geodesic paths on shape spaces. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 26:372–383, 2004. ISSN 0162-
8828. doi: 10.1109/TPAMI.2004.1262333.
Serge Lang. Fundamentals of differential geometry. Springer–Verlag, 1999.

M. Leventon, E. Grimson, and O. Faugeras. Statistical shape influence in

geodesic active contours. In IEEE Conf. on Comp. Vision and Patt. Recog.,
volume 1, pages 316–323, 2000.
YanYan Li and Louis Nirenberg. The distance function to the boundary, Finsler
geometry and the singular set of viscosity solutions of some Hamilton-Jacobi
equations. Comm. Pure Appl. Math., LVIII, 2005. (first received as a personal
communication in June 2003).

R. Malladi, J. Sethian, and B. Vemuri. Shape modeling with front propagation:

a level set approach. IEEE Transactions on Pattern Analysis and Machine
Intelligence, (17):158–175, 1995.
Andrea Mennucci. Metrics of curves in shape optimization and analysis. page
110. 2009.
Andrea Mennucci, Anthony Yezzi, and Ganesh Sundaramoorthi. Properties of
Sobolev Active Contours. Interf. Free Bound., 10:423–445, 2008.

P. W. Michor and D. Mumford. Riemannian geometries on the space of plane

curves. ESI Preprint 1425, arXiv:math.DG/0312384, Dec. 2003.
Peter W. Michor and David Mumford. Vanishing geodesic distance on spaces
of submanifolds and diffeomorphisms. Documenta Math., 10:217–245, 2005.
URL http://www.univie.ac.at/EMIS/journals/DMJDMV/vol-10/05.pdf.
Peter W. Michor and David Mumford. Riemannian geometris of space of plane
curves. J. Eur. Math. Soc. (JEMS), 8:1–48, 2006.
Peter W. Michor and David Mumford. An overview of the Riemannian metrics
on spaces of curves using the Hamiltonian approach. Applied and Computa-
tional Harmonic Analysis, 23:76–113, 2007. doi: 10.1016/j.acha.2006.07.004.
URL http://www.mat.univie.ac.at/~michor/curves-hamiltonian.pdf.
Washington Mio and Anuj Srivastava. Elastic-string models for representation
and analysis of planar shapes. In CVPR (2), pages 10–15, 2004a.

92
Washington Mio and Anuj Srivastava. Elastic-string models for representation
and analysis of planar shapes. In Conference on Computer Vision and Pattern
Recognition (CVPR), June 2004b. URL http://stat.fsu.edu/~anuj/pdf/
papers/CVPR_Paper_04.pdf.
D. Mumford. Slides of the gibbs lectures. URL http://www.dam.brown.edu/
people/mumford/Papers/Gibbs.pdf.
D. Mumford and J. Shah. Boundary detection by minimizing functionals. In
Proc. IEEE Conf. Computer Vision Pattern Recognition, 1985.
D. Mumford and J. Shah. Optimal approximations by piecewise smooth func-
tions and associated variational problems. Comm. Pure Appl. Math., 42:
577–685, 1989.
S. Osher and J. Sethian. Fronts propagating with curvature-dependent speed:
algorithms based on the Hamilton-Jacobi equations. J. Comp. Physics, 79:
12–49, 1988.
N. Paragios and R. Deriche. Geodesic active regions: a new paradigm to deal
with frame partition problems in computer vision. International Journal of
Visual Communication and Image Representation, Special Issue on Partial
Differential Equations in Image Processing, Computer Vision and Computer
Graphics, 13(2):249–268, June 2002a.
N. Paragios and R. Deriche. Geodesic active regions and level set methods for
supervised texture segmentation. International Journal of Computer Vision,
46(3):223, 2002b.
T. R. Raviv, N. Kiryati, and N. Sochen. Unlevel-set: Geometry and prior-based
segmentation. In Proc. European Conf. on Computer Vision, 2004.
Remi Ronfard. Region based strategies for active contour models. International
Journal of Computer Vision, 13(2):229–251, October 1994. URL http://
perception.inrialpes.fr/Publications/1994/Ron94.
M. Rousson and N. Paragios. Shape priors for level set representations. In Proc.
European Conf. Computer Vision, volume 2, pages 78–93, 2002.
Walter Rudin. Functional analysis. McGraw–Hill, 1973.
K. Siddiqi, Y. B. Lauzière, A. Tannenbaum, and S. Zucker. Area and length
minimizing flows for shape segmentation. IEEE Transactions on Image Pro-
cessing, 3(7):433–443, 1998.
L. Simon. Lectures on Geometric Measure Theory, volume 3 of Proc. Center for
Mathematical Analysis. Australian National University, Canberra, 1983.
Stefano Soatto and Anthony J. Yezzi. DEFORMOTION: Deforming motion,
shape average and the joint registration and segmentation of images. In ECCV
(3), pages 32–57, 2002.
Anuj Srivastava, Shantanu H. Joshi, Washington Mio, and Xiuwen Liu. Sta-
tistical shape analysis: Clustering, learning, and testing. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 27:590–602, 2005. ISSN 0162-
8828. doi: 10.1109/TPAMI.2005.86.

93
Ganesh Sundaramoorthi, Anthony Yezzi, and Andrea Mennucci. Sobolev active
contours. In Nikos Paragios, Olivier D. Faugeras, Tony Chan, and Christoph
Schnörr, editors, VLSM, volume 3752 of Lecture Notes in Computer Science,
pages 109–120. Springer, 2005. ISBN 3-540-29348-5. doi: 10.1007/11567646_
10.
Ganesh Sundaramoorthi, Jeremy D. Jackson, Anthony Yezzi, and Andrea Men-
nucci. Tracking with Sobolev active contours. In Conference on Computer
Vision and Pattern Recognition (CVPR06). IEEE Computer Society, 2006.
ISBN 0-7695-2372-2. doi: 10.1109/CVPR.2006.314.
Ganesh Sundaramoorthi, Anthony Yezzi, and Andrea Mennucci. Sobolev
active contours. Intn. Journ. Computer Vision, 2007a. doi: 10.1007/
s11263-006-0635-2.
Ganesh Sundaramoorthi, Anthony Yezzi, Andrea Mennucci, and Guillermo
Sapiro. New possibilities with Sobolev active contours. In Scale Space Vari-
ational Methods 07, 2007b. URL http://ssvm07.ciram.unibo.it/ssvm07_
public/index.html. "Best Numerical Paper-Project Award" ; also [Sun-
daramoorthi et al., 2008b].
Ganesh Sundaramoorthi, Anthony Yezzi, and Andrea Mennucci. Coarse-to-fine
segmentation and tracking using Sobolev Active Contours. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence (TPAMI), 2008a. doi:
10.1109/TPAMI.2007.70751.
Ganesh Sundaramoorthi, Anthony Yezzi, Andrea Mennucci, and Guillermo
Sapiro. New possibilities with Sobolev active contours. Intn. Journ. Computer
Vision, 2008b. doi: 10.1007/s11263-008-0133-9.
Ganesh Sundaramoorthi, Andrea Mennucci, Stefano Soatto, and Anthony Yezzi.
A new geometric metric in the space of curves, and applications to tracking
deforming objects by prediction and filtering. SIAM Journal on Imaging
Sciences, 4:109–145, 2011. doi: 10.1137/090781139.
Alain Trouvé and Laurent Younes. Local geometry of deformable templates.
SIAM J. Math. Anal., 37(1):17–59 (electronic), 2005. ISSN 0036-1410.
A. Tsai, A. Yezzi, W. Wells, C. Tempany, D. Tucker, A. Fan, E. Grimson, and
A. Willsky. Model-based curve evolution technique for image segmentation. In
Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, volume 1,
pages 463–468, Dec. 2001a. doi: 10.1109/CVPR.2001.990511.
A. Tsai, A. Yezzi, and A. S. Willsky. Curve evolution implementation of the
mumford-shah functional for image segmentation, denoising, interpolation,
and magnification. IEEE Transactions on Image Processing, 10(8):1169–1186,
Aug 2001b.
L. A. Vese and T. F. Chan. A multiphase level set framework for image seg-
mentation using the mumford and shah model. Int. J. Computer Vision, 50
(3):271–293, 2002.
C. Xu and J. L. Prince. Snakes, shapes, and gradient vector flow. IEEE Trans-
actions on Image Processing, 7(3):359–369, March 1998.

94
A. Yezzi and A. Mennucci. Metrics in the space of curves. Preprint,
arXiv:math.DG /0412454, May 2005a.

A. Yezzi, A. Tsai, and A. Willsky. A statistical approach to snakes for bimodal

and trimodal imagery. In Int. Conf. on Computer Vision, pages 898–903,
October 1999.
Anthony Yezzi and Andrea Mennucci. Geodesic homotopies. In EUSIPCO04,
2004a. URL http://www.eurasip.org/content/Eusipco/2004/defevent/
papers/cr1925.pdf.
Anthony Yezzi and Andrea Mennucci. Metrics in the space of curves. arXiv,
2004b.
Anthony Yezzi and Andrea Mennucci. Conformal metrics and true “gradient
flows” for curves. In International Conference on Computer Vision (ICCV05),
pages 913–919, 2005b. doi: 10.1109/ICCV.2005.60.
L. Younes. Computable elastic distances between shapes. SIAM J. Appl. Math.,
58(2):565–586, 1998a.
Laurent Younes. Computable elastic distances between shapes. SIAM
Journal of Applied Mathematics, 58(2):565–586, 1998b. doi: 10.1137/
S0036139995287685.
Laurent Younes, Peter W. Michor, Jayant Shah, and David Mumford. A metric
on shape space with explicit geodesics. Atti Accad. Naz. Lincei Cl. Sci. Fis.
Mat. Natur. Rend. Lincei (9) Mat. Appl., 19(1):25–57, 2008. ISSN 1120-6330.
doi: 10.4171/RLM/506.
Charles T. Zahn and Ralph Z. Roskies. Fourier descriptors for plane closed
curves. IEEE Trans. Comput., 21(3):269–281, 1972. ISSN 0018-9340. doi:
10.1109/TC.1972.5008949.
Song Chun Zhu, Tai Sing Lee, and Alan L. Yuille. Region competition: Unify-
ing snakes, region growing, energy/bayes/MDL for multi-band image seg-
mentation. In ICCV, pages 416–, 1995. URL citeseer.ist.psu.edu/
zhu95region.html.

95