See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/283500791

Thoughts on Shape

Chapter · January 2006

DOI: 10.1075/aicr.67.19ley

CITATIONS READS

2 188

1 author:

Frederic Fol Leymarie

Goldsmiths, University of London
113 PUBLICATIONS 2,146 CITATIONS

SEE PROFILE

All content following this page was uploaded by Frederic Fol Leymarie on 05 November 2015.

The user has requested enhancement of the downloaded file.

JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.1 (57-351)

Thoughts on shape

Frederic Fol Leymarie

. Introduction

What is shape?
Shape is the information, the structure induced by the presence of object el-
ements in a space. An object manifests its presence by bits of data we refer to as
“outline,” to be defined later. Such an outline can be obtained, for example, by a
sampling process, such as when a laser camera scans the surface of a 3D object.
Space itself is defined by its associated measuring sticks, or lack thereof, and the
number of dimensions needed to enumerate its extent.
Shape is the “glue” between object and space, revealed by an interrogation of
outlines. We interrogate an object by probing its outline. This inquisitive process
we call a “transform” as it not only is a process which attempts to identify and
characterise the object’s structure, but in doing so modifies, perturbs, the object
and space itself.
Complexity in forms emerges by the grouping, assembly, concatenation of ob-
ject elements. This complexity can be understood by defining a language for shape.
Such a language could include atomic structures, the alphabet or a code, which
when put together, via a grammar, create a word, and assembly of words create
sentences. The alphabet and grammar can then be thought of as the essence of
what the organisation of shape is.
In the following, I first define a few concepts to initiate our discussion by giv-
ing meanings to “space,” “object,” “outline,” “shape,” “form,” “structural field,”
“shape probing,” “representation.” I proceed to explore some of the most pop-
ular transforms which can reveal the nature of shape and then specify classical
geometries: the science of measure on space and its objects which leaves invari-
ant certain properties under specific transforms. I then depart from the tradition
and propose a dichotomy between two main classes of transform sequences, I call
“horizontal” and “vertical,” which do not require invariance to specify and classify
useful probes for shapes. While horizontal transform sequences refer to transforms
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.2 (351-411)

 Frederic Fol Leymarie

which operate along an outline, vertical ones act away from it. Along the way I pro-
vide a number of examples from the field of computer vision illustrating various
transforms and their belonging to one class or the other. Finally I consider a few
examples of “shape grammars” which permit to construct and analyse shape at
varying degrees of complexity.

. Preliminaries

Definition 2.1 (Space) Entity having extent spanning a number of dimensions; i.e.,
such that we can refer to any constitutive locus – positions – of that entity. Space can be
homogeneous, i.e., such that its dimension set is everywhere similar, or heterogeneous,
i.e., such that its dimension set may vary locally, regionally. For example, consider
the “classical” Euclidean space in 3D which is homogeneous, isotropic and of infinite
extent, versus Einstein’s relativistic space which is heterogeneous as a function of mass
(of objects).
Definition 2.2 (Object) A region of a given space, such as the Euclidean 3D space,
E3 , having some distinguishing characteristic. For example, a (homogeneous) region,
x of E3 , whose density, δ, is above some fiducial value, φ, i.e.: δ(x) > φ; refer to
the concept of solid shape and tolerances for shape by Koenderink (Koenderink
1990: Ch. 2).
Definition 2.3 (Outline) (i) Samples of an object delimiting its outer limits. For ex-
ample, consider the lines (curves) marking the edges of a pattern in an image. Also,
(ii) a sketch in which object features (not necessarily boundaries) are marked (without
shading; e.g., a line-drawing of a figure).
Definition 2.4 (Shape (in isolation)) Description of the outline of an object.
Such a “description” might simply consist in a quantitative study of the outline
itself; e.g., the ordered enumeration of the contour points.
Alternatively, a “description” of some of the distinctive features extracted from the
outline can define “shape;” e.g., the 3 corners of a triangle.
Thus, this description will depend on the method used to probe the object’s
outline. As such, one may say that “shape is operationally defined” (Koenderink
1990: 15).
Definition 2.5 (Form) Qualitative description of the shape of an object; e.g., the
“round form” of the sphere or of an ellipsoid.
Thus, “space” is a place-holder for content, while the notion of “object” corre-
sponds to (i) the physical body (assumed) present in the scene, or (ii) the region
of space under scrutiny. The notion of “shape” corresponds to the description of
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.3 (411-457)

Thoughts on shape 

the “available” outline of this object.1 For example, it may be derived from some
projection of the object in the scene onto the image (or retinal) plane. Finally the
notion of “form” is a more abstract concept. Many shapes may have the same form.
Shape is the “content” of form; i.e., it gives a more precise meaning or identity to
a form. I think of form as a kind of “meta-shape,” a concept which is useful for
object recognition.

Class of object outlines to be considered

In the following I will consider as “objects” sets made of points, curve segments,
surface patches and volumetric elements.
N.B.: I will not define at this point, in a formal manner, what I mean by a curve
or a surface or a volumetric object. I shall not commit myself (yet) to a particular
method, such as in differential geometry, where these objects are defined through
parametrisation under compatible diffeomorphisms (i.e., coordinate charts). Such
methods of definition tend to impose strong constraints on the way one looks at
the objects being scrutinised. Along the way, however, I will introduce and use
some of the terminology of differential geometry, e.g., to express relevant concepts
such as tangents and vector fields.

Shape can also be defined within a context or environment. For example, shape
can be extended to embody the influence of “gravity,” i.e., the orientation of an
object with respect to an horizon or reference frame, and its relative position with
respect to that horizon. Consider for example a “diamond” shape which is distin-
guished from a square or parallelogram only by its orientation with respect to an
horizon or ground plane. This is a famous “phenomenon” first identified by Ernst
Mach in the late 19th century and later “refined” by Goldmeier in the 1930’s (Ley-
ton 2001: Fig. 8.3). Refer also to the related notions of “visual weight” and “visual
direction” from the Gestalt theory of perception: a measure of perceptual forces
exercised by an object on and via its visual context (Arnheim 1974).

Definition 2.6 (Shape (within a context)) Description of the outline of an object

under the influence of its environment, or of other objects.

The study of the shape of an object in isolation or within a context, a scene, can be
unified via the concept of a “structural field.”

. Shape as the structural field of an object

Concept 2.1 (Shape as a structural field) Shape is the structure of the field existing
“around” an object. This field may consists of geometric entities – such as curvatures,
singularities (of some appropriate mappings), gauge figures (e.g., Dupin’s indica-
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.4 (457-507)

 Frederic Fol Leymarie

trix) – which may exist in association with each sample of the object being scrutinised.
In order to identify or measure such entities, we will have to “probe” the field.

Koenderink’s point of view

Paraphrasing Jan Koenderink (Koenderink 1990: §2.2):
One can say that objects induce a (geometrical) field around them which can
be probed operationally. The structure of this field is the shape. Shape is a property
of both the object and the probing method. Furthermore, one can say that shape
will depend on the perception, i.e., the mode of interaction (probing) and the ex-
pectations (models, theories).

Concept 2.2 (Probing shape and the uncertainty principle) Since “shape” will also
depend on the probing method, and not simply on the object’s structure, we have here
an example of a variation on the “uncertainty principle” of Heisenberg in action. We
will be as precise about our identification and measure of the structural field of an
object as the selected probing method permits to.

We will have to seek for probing methods which, although they may modify (and
not just identify or measure) the object’s field, preserve as much as possible this
structure, or emphasise certain of its features. Also, we will have to identify which
structures are most relevant for shape analysis. This selection can only be per-
formed by considering jointly the effects of probing the field. Note also that, if we
want to take into account the context for an object, say its orientation relative to
a ground plane, the probing method will need to be defined to also consider this
influence. Various probing methods will emphasise various features in the form of
a “representation.”

Definition 2.7 ((Shape) Representation) Mathematical abstraction of an object

making explicit special selected features.

Typically, a representation results from applying a probing method to the outline

of an object, revealing certain shape features, such as the corners of a polygon.
Note that we can further consider the representation of a representation and build
a hierarchy of structures for an object, i.e., make its shape available at various levels
of abstraction. For example, consider a “donut” object in Euclidean space (a torus).
An initial representation may capture the closed tubular surface structure from
original surface samples. This representation can be further probed to reveal the
central circular axis of the tore-like object and the associated hole it creates in
space.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.5 (507-555)

Thoughts on shape 

A final refinement in my definition of “shape” is necessary to capture the

notion of complexity, such as due to the concatenation of parts or due to defor-
mations and growths.

. Shape as the recoverable structural field of an object

Shape as a recoverable structural field

Paraphrasing Michael Leyton (Leyton 2001: §2.23, 75):
Shape is the recoverable “structural field” of a machine whose initial state
space is the object’s outline.2

A “machine” is a pair of functions acting on a given input set (of parameters, data)
(Leyton 2001: §1.15, 30): such that (i) one function specifies the states of the ma-
chine being transformed, and (ii) the other function specifies a new output set
produced. The concept of a “machine” emphasises that “shape” can be cast as a
dynamic phenomenon.3
In the following section, important examples of probes used to reveal the
shape of an object are first considered. Such probes are called “shape transforms”
or simply “transforms.” After some preliminary definitions, a number of impor-
tant (practical) properties for shape transforms are considered. This preliminary
study will set the stage to consider sequences of shape transforms and construct a
bi-partite classification (in §4) inspired by applications in computer vision used to
reveal the shape of objects.

. Transforms for shape

Definition 3.1 (Domain and range) Let us call the space (or region of space) con-
taining the outline or representation being probed, the domain, while the space (re-
gion) containing the resulting outline or representation is called range. Let us denote
the former by D and the latter by R.

Concept 3.1 (Shape transform) A mapping F from D to R with properties to be

specified latter (§ 3.1). Thus,
F : D→R.

Concept 3.2 (Shape operator) The mathematical (operational) object, ρ, which

permits to implement the transform F. Thus,
ρ
F : D→R.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.6 (555-612)

 Frederic Fol Leymarie

For example, the operator could be a convolution kernel in the context of a

smoothing transform.

Remark 3.1 (Spaces for domain and range) The domain and range of a shape
transform may lie in the same space or not. E.g., in the case of a “projection,” the
range lies in a subspace with respect to the original space of the domain.

Note, however, that in many practical situations, both the domain and range will
lie in the same space, i.e., the transform is an instance of a general deformation:
1–to–1 mappings of a space onto itself.

Question 3.1 (What is the purpose of a transform?) Typically, one uses the concept
of a transform to simplify, or break down, or map to a more explicit representation,
the problem being attacked. A simplification may have different meanings depending
what are the goals to be achieved. A transform may make more explicit some prop-
erty(ies) of the shape (or the domain); e.g., it could amplify the presence of linearities
of a planar object such as a square; a classical example here is the Hough transform.4
Alternatively, a transform may look more like a filter which removes some undesir-
able properties; e.g., smoothing a surface with convolution kernels, this to attenuate
noise or to interpolate data. A classical example here is the Fourier transform which
first maps a signal or image from the spatial domain to the frequency range where
certain spatial frequencies can be selected; e.g., in the context of low, high or band pass
filtering.

Remark 3.2 (Shape interrogation in CAD-CAM) In the field of CAD-CAM,5 the

term “interrogation” is used for an equivalent notion of shape transform (Patrikalakis
& Maekawa 2002: Ch. 8): “interrogation is the process of extraction of information
from a geometric model,” where the latter corresponds to the notion of outline.

. Examples of shape transforms

In the following I enumerate important shape transforms in terms of their distin-

guishing properties.

.. Invertible transforms

It is often argued that a transform for shape should be invertible, i.e., the mapping
F should have a unique inverse F –1 : R → D, such that F–1 ◦ F = I, where I denotes
the identity, and ◦ the composition of functions. Why so? Because it is believed that
if no information is lost (since we can then always recover the original data) the
transform F should be a powerful one in the sense that it is information–preserving.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.7 (612-669)

Thoughts on shape 

Involutive transform. An involutive transform is a special case of an invertible

transform, such that it is its own inverse. Apply it twice and you are back to the
original object, i.e., F–1 ≡ F, thus F ◦ F = I.
A famous example of such a transform is the Legendre transform which maps
functions on a vector space to functions on the dual space.
Definition 3.2 (Legendre tr. for functions of 1 variable) For a convex curve, y =
f (x), s.t., f "(x) > 0, consider the transformation of a curve from point coordinates
(x, f (x)) to (tangent) line coordinates (p, g(p)).
Example 3.1 (Legendre tr. of polygonal contours) Let f (x) be a convex polygon.
Consider the graphs of f and g, i.e., the traces of f and g in the 2D spaces (x, f )
and (p, g). Then g(p) is also a convex polygon, in which the vertices (edges) of f (x)
correspond to edges (vertices) of g(p).6 Thus a corner is transformed into a segment
(e.g., see Arnold 1989: Fig. 44, p. 62).
Definition 3.3 (Legendre tr. for functions of 2 varsiables) For a surface parame-
terised by a Monge patch, u = u(x, y), which is non-developable (i.e., its Gaussian
curvature is nowhere vanishing, and thus it cannot be spread flat on a plane), consider
the transformation of a surface from point coordinates (x, y, u(x, y)) to (tangent)
plane coordinates (ξ, η, ω(ξ, η)).
Remark 3.3 The Legendre Transform is different from a mere coordinate transform,
since, rather than assigning to a single point another point, it assigns to every surface
element (x, y, u, ux, uy ) a surface element (ξ, η, ω, ωξ , ωη ).

Remark 3.4 (From Lagrange’s to Hamilton’s mechanics) It is by means of a Leg-

endre Transform, that a Lagrangian system of n 2nd order differential equations is
converted into a remarkably symmetrical system of 2n 1st order differential equations,
the “Hamiltonian system of equations” (or “canonical equations”).
Note that, with respect to the previous comment on the “purpose of a shape
transform,” it appears that there is potential for some conflict of interest, when
considering that a “good” transform should be invertible. For example, if one de-
sires to “simplify” the shape representation (and eventually, description) it would
come as no surprise that some information or level of detail is lost in the pro-
cess. It might even be desirable to eliminate (or modify) some data which would
be “irrelevant” for further analysis in any case; e.g., details existing at a “too” fine
resolution to be perceptually significant.7

.. Idempotent transforms

An idempotent transform is such that if you apply it more than once and se-
quentially (i.e., the output of one transform becomes the input for the following
transform) you get the identity: nothing happens to the data anymore.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.8 (669-737)

 Frederic Fol Leymarie

In other words, the composition of the transform F with itself yields F again
(and forever):
F◦F ≡F,
or, in terms of the corresponding operator, ρ,
ρ2 = ρ .
Remark 3.5 (Fixed points of idempotent transforms) Note that, all the points in
the range of an idempotent transform, i.e. F(x) = y ∈ R, are fixed points.
Jean Lagarde put forward the following informal argument in favour of idempo-
tency (Lagarde 1990: Ch. 2):
If the goal of the operator is to recover a valid interpretation from initial data,
then applying the process one more time on the supposedly valid interpretation
should not change it.

A few examples of idempotent transforms are listed below.

Projection. The projection, F ≡ π, which maps a set U ∈ D onto a subspace R*

of R, i.e., π : → R* ⊂ R is idempotent. E.g., in Euclidean 3-D space, think of the
orthogonal projection onto the x–y plane: π : (x, y, z) ∈ E3 → (x, y) ∈ E2 ⊂ E3 .

Linearisation. A transform, F ≡ l, used in linear algebra of vector spaces, which is

such that v ∈ R(l) ⇒ l(v) = v. Thus, the corresponding matrix (operator), ρ ≡ A,
is such that A2 = A (Bamberg & Sternberg 1988: 38).
This linear operation is non–singular only if A = I, i.e., the identity, in which
case l is a trivial mapping. Otherwise, it is singular, i.e., such that Det(A) = 0. E.g.,
if l is such that l : R2 → R2 and A is singular but not the zero matrix, then l maps
the plane, R2 onto a line. If A = 0, then l maps the whole plane onto the origin.
An important application is the decomposition in a Linearly Independent (LI)
set, i.e., the projection on basis sets/functions. E.g., think of an orthonormal rep-
resentation such as the Fourier decomposition.

Mathematical morphology transforms. The closing and opening operators of

mathematical morphology are such that they fill-in or remove details of size
smaller or equal than a structural element B (Serra 1982; Soille 2004), which can
be seen as a “probe” or tool to sculpt the original dataset.8 Applying these opera-
tors again does not further change the data which is said to have been regularised –
up to the scale and layout defined by B.

Edge detection. Edge detection can be postulated as an idempotent transform

(Owens et al. 1989; Ronse 1993). The transform F acts on the space of gray level
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.9 (737-793)

Thoughts on shape 

images, D ≡ I , and has for its range a binary feature map corresponding to the
edges and lines originally in I(x, y) ⊂ I . Note that R ≡ F(I) is also a valid im-
age, i.e., F(I) ⊂ I . However, applying F once more should not change the result:
(F ◦ F)(I) = F(I).9

Remark 3.6 (Optimisation) Local energy minimisation can be seen as an idempo-

tent transform, a local minimum being a fixed point of the cost function (operator)
(Lagarde 1990: 15). Applying the transform, once a local minimum has been reached,
will not change the result.10

Essentially, any transform implying an iterative and optimising process, which is

defined as having to converge on some fixed point, can be cast as an idempotent
transform.

.. Separable transform

Transform in a coordinate system adapted to the boundaries11 of the shape being
transformed or to the domain acted upon. E.g., a transform acting on the real
plane, R2 , being such that:

F(x, y) = F(x) ◦ F(y) = F(y) ◦ F(x) .

Convolution kernels used in image processing often are selected because they share
this property. E.g., smoothing and gradient filters based on the Gaussian kernel
are more efficiently performed by using the separability of the Gaussian function
(Burt 1981, 1983).

.. Linear transforms

A transform such that

F(ax + by) = aF(x) + bF(y) , a, b ∈ R .

A linear transform permits one to use the principle of superposition, i.e., arbitrary
linear combinations of data can be transformed separately or altogether.
Note that a linear transform always possess a matrix representation, A, which
can be taken as the operator. In particular, the composition of two linear trans-
forms corresponds to matrix multiplication.

Affine transform. A linear transform without any preferred origin. Any affine
transformation can be written as a (non–singular) linear transform followed by
a translation.
An affine transform does not preserve distances or angles, but it does preserve
parallelism and ratios of lengths (i.e., ordering of points).
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.10 (793-860)

 Frederic Fol Leymarie

Conformal transform. A linear transform which preserves angles and orientation

(i.e., such that Det(A) > 0), but not necessarily distances (Bamberg & Sternberg
1988: §2.1).12

Orthogonal transform. A linear transform which preserves distances (and, as a

consequence, angles as well). It also preserves “ordinary” scalar products, (•, •),
i.e., such that they obey the following three properties: (a) bilinearity, (b) sym-
metry, and (c) positive definiteness.13 E.g., rotations, translations (and thus rigid
motions), and reflections through the origin are orthogonal transformations. An
Euclidean transform, i.e., equipped with the Euclidean metric and scalar product,
is an example of an orthogonal transform.

Lorentz transform. A linear transform which preserves a quadratic form, Q, whose

scalar product Q(•) =< •, • > is symmetric, billinear but not necessarily positive–
definite. Associated with Q is a symmetric matrix, A, i.e., such that A = AT , and
thus Q(v) = (Av, v) = (v, Av).
In the context of special relativity, the Lorentz transform must carry the light
cone (or null cone), {v : Q(v) = 0}, onto itself. The corresponding affine trans-
form is called the Poincaré transform and the corresponding metric is called the
Minkowski distance, where the Euclidean circles are replaced by hyperbolas (Bam-
berg & Sternberg 1988: 154–157).
The Lorentz transform is typically used to relate the space and time coordi-
nates (or in general any four-vector) of one inertial reference frame, A, into those
of another one, A , travelling at a relative speed of v to A along some axis, say the
x-axis. This is the natural set-up to study the position, brightness and shape of im-
ages in arbitrary general relativistic spacetime, e.g., when considering gravitational
lensing, i.e., how massive objects like stars bend light (Perlick 2004).

Symplectic transform. A linear transform which preserves the symplectic scalar

product, ω(•, •), a bilinear, not necessarily positive–definite, and anti–symmetric
(or skew–symmetric) product; thus, ω(v, w) = –ω(w, v). Note that ω is a 2–form,
i.e., it takes its (real) values on pairs of vectors, and is a skew–symmetric opera-
tor. In (differential) geometry, it represents the oriented area of the parallelogram
spanned by the two vectors it acts upon. In particular, in R3 , it corresponds to the
vector product of vector analysis.
Symplectic transforms are useful in geometric optics and Hamiltonian me-
chanics, and, in particular, are relevant to wavefront (or grassfire) propagations
where one propagates fronts, i.e., stacks of oriented one-forms, while preserving
their areas.14
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.11 (860-913)

Thoughts on shape 

.. Differential transforms

Transform which maps the outline to a differential representation. For example,
the most common such transform maps an outline to its tangent space, where the
best first order linear representation is obtained at each locus of the outline.

Lie group transforms and Hamiltonian flows. Lie group transforms are linear
differential ones which obey a particular constraint, i.e., the Jacobi identity. Essen-
tially, the latter states that flows (associated with tangent vector fields) commute,
i.e., loosely speaking, if one travels along two flows, L1 followed by L2 , on a man-
ifold, M, one ends at the same point than if one had travelled first along L2 and
then along L1 . More on this topic and its relation to vision problems can be found
in the works of William C. Hoffman (Hoffman 1966, 1978, 1984) and his followers
(Ferraro & Caelli 1988; Papathomas & Julesz 1989; Hansen 1992).
Furthermore, certain classes of global deformations (of a continuum) can be
expressed in terms of Lie derivatives, i.e., certain linear operators on vector fields
(associated to flows) which indicate how differential forms change along integral
curves on these vector fields. Lie transforms are instances of “horizontal transform
sequences,” as we shall see later (§4.1).

Interrogation methods in CAD-CAM. Shape interrogation methods in CAD-

CAM can be classified by the “order of the derivatives of the curve or surface
position vector” involved (Patrikalakis & Maekawa 2002: Ch. 8). “Ray tracing” is
an example of a first order method used to compute a shaded image of a surface
assuming the reflected light is a simple (linear) function of the angle θ between the
surface normal and the light direction vector (from a point light source). Examples
of second order methods include “curvature plots” and “inflection point” maps
for the interrogation of planar curves, “surface curvature” maps and “geodesics”
(curves of zero geodesic curvature) for the interrogation of surfaces. Third order
methods include “torsion” maps of space curves, and “stationary points (extrema)
of curvature” of curves and surfaces. “Stationary points of the torsion of space
curves” is an example of a fourth order method.

.. General deformation of a continuum

A non–necessarily–linear, non–necessarily–affine transform, where an object, O,
is deformed via a mapping, F, that carries each material point, p ∈ O, into a
point F(p) = x ∈ R. This is a natural set–up for kinematics in the context of
continuum mechanics.15 The shape transform, F, is called a deformation, and in
general it is assumed one–to–one and smooth (thus, differentiable). An important
special case is the homogeneous deformation, where the deformation gradient, ∇F,
is constant: thus, global and local motions are the same. It can be shown that any
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.12 (913-968)

 Frederic Fol Leymarie

homogeneous deformation, Fhomo , can be uniquely decomposed into a rotation

about a fixed point, G, and a stretch, Sr or Sl :
Fhomo = G ◦ Sr = Sl ◦ G .
Note that, in the special case of a rigid (and affine) transform/deformation, the
gradient of Fhomo reduces to a (constant) rotation with matrix representation  ,
i.e., ∇Fhomo = ∇G ≡ .
There are three main concepts used in the elementary study of general de-
formations: forces (operators), conservations laws (constraints), and constitutive
assumptions (simplifications).
– Forces A force, f, in this context can be equated with our notion of the operator
of the transform; i.e., ρ ≡ f. Three types of forces are considered in continuum
mechanics:
1. Contact force between separate parts of the object, O, denoted fc .
2. Contact force exerted by the environment on the boundary of O, denoted
fe .
3. Internal force exerted on the interior points of O, denoted fi .

The forces fc and fe acting on the boundary of O, i.e., on the surface ∂O, are
represented through the use of a spatial tensor field, T, called stress (or Cauchy
stress). Note that, forces like “pressure,” “tension” or “shear” are all special
instances of stresses.
In computer vision a simple model for deformable curves and surfaces un-
der tension, called “snakes,” has been used to represent outlines and track them
in dynamic imagery such as video sequences (Kass et al. 1988; Terzopoulos et
al. 1987; Leymarie 1990; Leymarie & Levine 1993), as well as to simulate wave
propagation constrained by a potential field (Leymarie & Levine 1992b).
– Conservation laws Various conservation laws are defined within the realm of
continuous mechanics. These are general laws which are usually helpful in
imposing constraints on the deformation F. Examples of such laws – which are
all, really, different instances of the same fundamental conservation principle –
are:
– Local and global conservation of mass.
– Conservation of energy (i.e., of the sum of potential and kinetic energies).
– Conservation of momentum, linear and angular, also called momentum bal-
ance laws.
– Conservation of power (i.e., the power expanded equals the sum of stress
power and rate of change of kinetic energy).

– Constitutive assumptions Conservation laws do not distinguish between dif-

ferent types of materials. To distinguish different types of material behaviour,
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.13 (968-1023)

Thoughts on shape 

additional hypotheses, called constitutive assumptions are made. These are of

three main types:
1. Constraints on the possible deformations the body may undergo. E.g., an
isochoric motion (i.e., such that the object’s volume remains constant).
2. Assumptions on the form of the stress tensor, T. E.g., a pure pressure or
a pure tension, like in the case of the snake (or “active contour”) model
of computer vision (Kass et al. 1988; Terzopoulos et al. 1987; Leymarie &
Levine 1993).
3. Constitutive equations relating the stress to the deformation in time, i.e., to
the motion; e.g., giving the pressure as a function of density as in the case
of an elastic fluid.

.. Wavefront propagation and shock transforms

Imagine another type of general deformation where no 1–to–1 mapping restriction
is imposed, i.e., many–to–1 or 1–to–many mappings are allowed or even desirable.
E.g., there will be loci in the range, R, where points in the domain, D, are mapped
to singletons – we say a shock occurred. In other words, no continuum restriction is
put on the material being deformed.

Continuous wave propagation in geometrical optics. Consider a point q0 in Eu-

clidean space E3 at time t0 , taken to be the generator of a disturbance transmitting
itself “locally” (Arnold 1989: 250). At time t0 + ∆t the disturbance is propagated
through space in the form of a surface, which we denote by φ(x, y, z, t) = L, repre-
senting the wavefront for each time t, where L is some constant. At any later time,
the new wavefront can be obtained as a collection of disturbances from a previ-
ous wavefront; this is the geometrical construction first discovered by Christiaan
Huygens in 1678.

Lemma 3.1 (Huygens’ principle) Let φq0 (t) be the wavefront of the point after time
t. For every point q of this front, consider the wavefront after time t1 : φq (t1 ). Then,
the wavefront of the point q0 after time t + t1 : φq0 (t + t1 ), will be the envelope16 of the
fronts φq (t1 ) such that q ∈ φq0 (t).

The wavefront is thus the boundary of the set of all points q to which “infor-
mation” from a given source can travel in time less than or equal to t (Arnold
1989: 249).
Huygens’ metaphor can be transfered to the discrete domain without diffi-
culty, e.g. by considering the Minkowski sum of small discrete spherical sets with
a discrete front at time t. This is precisely the realm of 3D Mathematical Mor-
phology, where the spherical sets are called structural elements (Serra 1982). The
problem is one of inefficiency: Minkowski sums with isotropic structural ele-
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.14 (1023-1080)

 Frederic Fol Leymarie

ments – required to ensure Euclidean results – have too much overlap. Consider
then the alternative viewpoint of Pierre de Fermat’s principle (circa 1660), which
states that:
Lemma 3.2 (Fermat’s principle) Information travels along rays from a point q0 to a
point q in the shortest possible time.
Rays are such that their gradient vector is normal to the wavefront, i.e., the direc-
tion of the ray p = ∇S, where S is the (optical) “path length” from a source q0 to
q, a function such that Sq0 (q) = t. The (level) set of all such ray paths thus coin-
cide with the wavefront φq0 (t). Rays and wavefronts are related precisely through
the refractive index ν, which provides the (local) metric structure of the medium:
dφ · ∇S = ν (Stavroudis 1972).17

Eikonal equation of geometric optics. Consider the special case where the index is
taken to be isotropic and homogeneous. Then, the direction of motion of the rays
and the wavefront coincide (Arnold 1989); thus
 
dφ · ∇S = dφ ∇S = 1
Note that, considering a(light)
 ray travelling at constant speed c, since the index is
fixed, we also have that p = ∇S = c, leading to the classical “eikonal” equation
for the wavefront (Stavroudis 1972):

2
2
2
 2
dφ = ∂φ + ∂φ + ∂φ = 1/c2 (1)
∂x ∂y ∂z
Without loss of generality, we can take the constant c = 1, that is, we propagate
wavefronts along straight rays with unit (relative) speed. Thus, the normal field to
any wavefront is identically ∇S and each wavefront moves along this field by a unit
distance step for each time step. We can therefore identify time with distance for an
accurate simulation of wavefront propagation. Given an initial set of wavefronts or
set of generators Q = {q0 : φq0 = 0} we can simulate wave propagation, according
to the eikonal equation (1), by computing point-to-set (Euclidean) distances:
 
L(q, Q) = min q – r (2)
r∈Q

Note that, we may consider the set of all wavefronts, Φ, from an initial time (dis-
tance) t = 0 to a final time of propagation t = T ≡ Lmax , where Lmax is the
maximal distance travelled by the wavefront. The set Φ then fills-in a portion of
space according to the distance function (2).
We call this volume of space filled with distance values by wave propagation,
“Distance Hypersurface,” following Goodman who introduced the term in 2D
(Goodman 1964).18 This is equivalent to an embedding of the wavefronts in a
4D space, where the fourth axis is taken for time or distance. The Distance Hyper-
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.15 (1080-1141)

Thoughts on shape 

surface is the graph of τ = Φ(x, y, z), τ being taken along the time (or distance)
axis.
Definition 3.4 (Regular distance hypersurface) A hypersurface, τ = Φ(x, y, z), is
called a Regular Distance Hypersurface over a region  of 3D space E3 , if at each
point of  it is a “local distance hypersurface.” An hypersurface is a local distance
hypersurface at a point q ∈ E3 if there is:
1. a spherical neighbourhood ℵ of q,
2. a smooth surface  which divides ℵ into two sub-regions, and
3. a constant l,

such that L = Φ(x, y, z) – l is the directed (or signed) distance from  for each point
in ℵ.
This graph takes single values almost everywhere, according to equation (2), ex-
cept at those points where wavefronts meet. This is where regularity is lost and
singularities in the distance mapping occur. The latter define the symmetry points
constitutive of a kind of 3D skeleton or symmetry set (§4.2). If we think in terms
of the physical metaphor of light rays, this is where such rays meet to define the
“burning points” or caustics of Archimedes.19
Another way to interpret the above concept is via the Implicit Function The-
orem. If the gradient ∇Φ = 0 at some point q, then it implies the regularity of
Φ in a neighbourhood of q. This only requires that Φ is one time continuously
differentiable on the respective open set. Equivalently, having a Lipschitz contin-
uous gradient ensures that the “symmetry points” stay away from the distance
hypersurface for some distance (Wolter 1985, 1992).20
The process of computing such a distance hyper-surface for a discrete dataset,
is commonly called “distance transform” or “map” in image processing. When us-
ing the Euclidean metric, such as in equation (2) above, it is called “Euclidean Dis-
tance Transform” (EDT) (Danielsson 1980; Leymarie & Levine 1992a; Leymarie &
Kimia 2000).

Shock transforms. As a somewhat concrete example, consider an outline taken as

an initial wavefront. Then, take the transform which moves this wavefront accord-
ing to the eikonal equation (1). This is a continuous deformation, also called an
evolution for the (wave)front. There will be times and places where material points
of the wavefronts collapse on, i.e., where a front self-intersects or becomes singu-
lar (non-smooth), or where two or more wavefronts meet each other. In terms of
a wavefront propagation, this is called a “shock.”
A hierarchy of shocks of different types can be defined depending on the na-
ture of this “collapse” or coalescence of material points. For example, in the case
of shocks occurring in the context of the planar curve evolution problem, refer to
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.16 (1141-1195)

 Frederic Fol Leymarie

the early works of Blum (Blum 1967, 1973), Montanari (Montanari 1969), and
more recent works of Kimia et al. (Kimia 1991; Siddiqi & Kimia 1996; Siddiqi et
al. 1999a, b; Kimia 2003). Shock transforms are instances of “vertical transform
sequences,” as we shall see later (§4.2).
Having described a number of commonly used shape transforms, I will now
consider a classical way of classifying them via “geometries.”

. Geometries for shape: Invariance under transforms

I follow here the classical definition of a geometry as the study of those properties of
subsets of a region, D, which are invariant under the application of some specified
transform. For example, one can consider affine, Euclidean, projective, or symplectic
geometries for shape.

Felix Klein’s Erlangen program

Hierarchies of geometries for shape were proposed by Felix Klein in a famous
speech at the University of Erlangen in 1872. Klein argued that different geome-
tries can be stratified by the properties of objects that are preserved by different types
of transforms (Klein 1893).
The category of transforms, Klein suggested to consider, was the group of
motions, i.e., 1–to–1 mappings of a space onto itself, that leave intact certain prop-
erties of objects in the considered space.

Norman and Todd have argued that (Norman & Todd 1992: 106):
There is a growing amount of evidence to suggest that human vision involves
a similar type of stratification in which the most perceptually salient aspects of
an object’s structure are those that remain invariant over the largest number of
possible transforms.

For an illustration of this concept of a hierarchy of geometries for a planar surface

refer to Lord and Wilson’s treatise (Lord & Wilson 1984: Fig. 7, p. 35).
Remark 3.7 (Differential geometry) When one studies geometry in a local neigh-
bourhood m of a smooth manifold M, one can reduce such local probing of differen-
tiable structure to infinitesimal interrogations in the tangent space to m ⊂ M, i.e.,
the first order, linear, approximation to M at m; this permits to “reduce” geometry
to linear algebra.

.. Riemannian geometries

A Riemannian geometry is a generalisation of the intrinsic geometry of a surface.21
The associated (metric) spaces are spaces of constant curvature. If the (Riemannian
or Gaussian) curvature of the space is zero, then we have a Euclidean space which is
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.17 (1195-1260)

Thoughts on shape 

homogeneous and isotropic. For strictly positive curvature, we have a spherical or

elliptic space (e.g., the space for “flat inhabitants” of the Earth). Finally, for strictly
negative curvature we have the hyperbolic space of J. Bolyai and N. I. Lobachevski.

Hyperbolic geometry. Study of those properties of subsets of Rn which are invari-

ant under the application of any hyperbolic transform; i.e., a transform based on
a hyperbolic metric, such as the Minkowski distance, defined within a hyperbolic
space (i.e., a space with negative intrinsic curvature).
It has been proposed in the literature that the correct geometry of binocular
vision or stereopsis is the hyperbolic one (Luneburg 1947, 1950). The latter gives
an explanation for some well-known perceptual illusions, such as, e.g., a percep-
tually straight line being physically (i.e., in Euclidean – environmental – space)
concave toward the observer (Blank 1978).

Euclidean geometry. Study of those properties of subsets of Rn which are invari-

ant under the application of any Euclidean transform; i.e., a “distance–preserving”
transform based on a Euclidean metric within a Euclidean space (i.e., a space
with vanishing intrinsic curvature). Thus, concepts such as angles and lengths are
meaningful.
Consider the example of a “rigid motion” in the plane R2 , an instance of a
Euclidean transform. Then, the outline of the transformed object is preserved. In
this sense it is a “very rigid” geometry, since any slight deformation of a form tells
us that we are dealing with a different object.
In particular, a circle is different from an ellipse in Euclidean geometry. So is a
square with respect to a parallelogram.
Norman and Todd emphasise that (Norman & Todd 1992: 95):
Throughout the literature on human perception, classical Euclidean geome-
try is by far the most common framework for describing the structure of the
environment.

From Euclidean geometry, we may “descend” in the hierarchy to reach affine ge-
ometry which can be portrayed as Euclidean geometry without a metric or without
a “machinery to measure distances and angles” (Koenderink 1990: 58).

.. Non-metrical geometries

Affine Geometry. Study of those properties of subsets of Rn which are invariant
under the application of any 1–to–1 affine transform (e.g., a transform which
preserves parallelism and ratios of length along any line).
Consider the example of a “shear motion” in the plane R2 . The outline is not
strictly preserved anymore. A square is transformed into a parallelogram. A cir-
cle into an ellipse. As such, squares and circles are not generic objects in affine
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.18 (1260-1323)

 Frederic Fol Leymarie

geometry: they do not define a class of object being invariant under an affine
transformation. In fact the concepts of squares and circles make no sense in affine
geometry; but parallelograms and ellipses do.
Paraphrasing Norman and Todd (Norman & Todd 1992: 96):

In affine geometry the distance metric is allowed to vary in different directions –

i.e., it is anisotropic. Arbitrary stretching transformations, which do not leave
distance invariant, but do preserve other properties such as the sign of Gaussian
curvature of a surface or the parallelism of pair of line segments. are permitted
in this geometry.

In stereopsis, the mapping between the two views has been modelled as an affine
transform, to recover the gradient of horizontal disparity (Jones & Malik 1992).

Projective geometry. Study of those properties of subsets of Rn which are invariant

under the application of any projective transform such as a perspective.
E.g., under a general projective transform, a conic section remains a conic
section, and a collinear set of points remains collinear.
In comparison to affine geometry where all parallelograms are equivalent, in
2D projective geometry, all quadrilaterals are equivalent, i.e., parallelism is not
necessarily preserved.

Ordinal geometry. Study of those properties of subsets of Rn which are invari-

ant under the application of a non–metric transform such that only simple order
relations, such as “greater than” or “lesser than,” are preserved.
James Gibson has argued that some of our perceptual awareness of the en-
vironment is based on such a geometry (Gibson 1950). Todd and Reichel have
suggested that an observer’s knowledge of smoothly curved surfaces can often in-
volve a form of ordinal representation such that neighbouring surface patches are
labelled in terms of which region is qualitatively “closer” to the point of observa-
tion (Todd & Reichel 1989). Note that, occlusion contours are an important source
of information in such a geometry.
A general deformation of a continuum where ordinal structure is preserved
can provide a vast class of transform for such a geometry (see the paragraph on
“General deformation,” §3.1.6, and the paragraph below on “rubber” geometry).

.. Pseudo-metrical geometries

These are geometries which can accommodate a metric but do not absolutely
require one.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.19 (1323-1388)

Thoughts on shape 

“Rubber” geometry (Kendall’s morphometrics). Study of those properties of sub-

sets of Rn which are invariant under the application of any general deformation,
i.e., a continuous and differentiable 1–to–1 mapping.
A famous application of this type of geometry into the field of morphometrics
was made in the early 20th century by D’Arcy Wentworth Thompson in his study
of the evolutionary changes of shapes of certain animals and plants (Thompson
1992).22
In stereopsis, the mapping between the two views has been modelled as a
deformation, to recover the horizontal disparity (Koenderink & van Doorn 1976).
If one gets rid of the requirement of differentiability and only imposes conti-
nuity, one gets a “topological geometry.”

Topological geometry. Study of those properties of subsets of Rn which are invari-

ant under the application of any “topological transform,” such as a homotopy. The
latter may be loosely described as a family of continuous maps which permits to
“continuously deform” a class of sets covering two objects so that they coincide.
For example, imagine having two sets of curves covering two surfaces (our two
objects’ outline here): a homotopy will establish a (continuous) correspondence
between these two covers.23 Therefore, two objects are considered topologically
equivalent, and thus having the same shape, if a deformation exist which can map
one to the other and vice versa. E.g., a sphere and a cube are topologically the same
object, but not a sphere/cube and a torus.
Note that it is has been observed that “topologically equivalent” shapes “pro-
duce stronger perceptions of apparent motion” (Norman & Todd 1992: 105).
A special class of topological geometry concerns sets which are self-similar,
such as the famous “monster curves” and “fractals.”

Fractal geometry. Study of those properties of subsets of Rn which are invariant

under the application of any topological transform and which present the char-
acteristic of self-similarity. Three main forms of self-similarity are considered: (i)
iterated transforms, which have a fixed geometric replacement rule, e.g., the Koch
snowflake which produces an infinite perimeter enclosing a finite area; (ii) re-
currence transforms, which define an iterative sequence at each point in space,
e.g., the Mandelbrot set (Mandelbrot 1982); (iii) random transforms, which use
a stochastic sequence rather than a deterministic one, e.g., the fractal landscapes
of Ken Musgrave (Ebert et al. 2002) or the Lévy flights, a type of random walks
(Mandelbrot 1982).
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.20 (1388-1433)

 Frederic Fol Leymarie

. Horizontal and vertical transform sequences

Having looked at different transforms for shape in isolation, including the “clas-
sical” geometry stratification based on invariance as proposed by Klein, I now
propose a more general bi-partite classification for sequences of transforms which
are applied to the outline and various shape representations of an object. This clas-
sification which does away with the invariance constraint, is inspired in particular
by the early filters developed in Mathematical Morphology, based on idempotence,
by the works of Zucker et al. on the application of differential transforms for the
recovery of shape, and by Michael Leyton’s work on the association of a history for
shape, based on those properties which vary under deformations.
A transform sequence involves at least a pair of transforms, where, typically,
the first one acts directly on the outline of an object, resulting in a first representa-
tion, while the second one then acts on this representation to make more explicit
special selected types of features.
I will distinguish such concatenations of transforms as to whether they operate
(i) along the outline of a shape or tangentially, or (ii) away from it or perpendic-
ularly. I call the two classes “horizontal” and “vertical” transform sequences. This
represents a fundamental dichotomy in the analysis of shape transforms. While
the horizontal class is inspired by a more classical approach to geometry based on
calculus and the study of those properties which remain intrinsic to the object, i.e.,
such that the analysis can be carried on (or along) the outline itself, independently
of the surrounding (or embedding) space, the vertical class is inspired by more
modern thinking in mathematics, such as singularity theory, and is based on an
analysis of the outline under transformation together with its surrounding space.
In the following I provide a number of examples from the field of computer
vision which illustrate how the two classes are used to solve similar problems.

. Horizontal transform sequences

Definition 4.1 (Horizontal transform sequence) It is a shape transform sequence,

such that the first transform operates in the tangent space (or the tangent bundle) of
the outline of the object; in other words, subsequent transforms are constrained to act
along the outline of the object.

Example 4.1 (1-D manifold on the plane: Co–circularity mapping and curvature)
Let a tangent field on a curve segment, α : I ⊂ R → R2 be defined; where this curve
segment represents, e.g., part of the outline (boundary) of a planar object (or of its
projection in the image plane).
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.21 (1433-1495)

Thoughts on shape 

Tangents are co-circular if they lie on a common circular arc of radius r. So one
can take a transform from the tangent space to the co–circularity space for pairs of
tangents.24
Furthermore, co-circularity is directly related to the concept of curvature as one
takes tangents being closer and closer to each other. In the limit, one can define a
real map on the interval, I, used to parameterised the curve to define the concept of
curvature, k, a scalar field:
∆θ 1
k : I → R, k = lim = ,
∆s→0 ∆s r
where s ∈ I is the parameter and θ is the angle between the tangents (or their corre-
sponding normals). Thus, curvature measures deviation from flatness, or collinear-
ity, for successive tangents.
The corresponding (limiting) circle is called the osculating circle (Pogorelov
1958: 57). It has second order contact with the curve α, i.e., both the curve and the cir-
cle agree (locally) in position, tangency and curvature (or up to second derivatives). If
s is the arc–length parameter, then k is equal to the norm (length) of the acceleration
vector α (Pogorelov 1958: 50):
k(s) = α (s) .
In conclusion, if a circle exists such that a pair of tangents (to α) are also tangent
to this circle, then they are said to be co–circular. The (osculating) circle is thus used to
“connect” or relate the pair of tangents in a “horizontal manner,” i.e., along the curve
through a simple interpolating shape, a circle, or through a derived “co–circularity
support,” i.e., a “directed” neighbourhood or thick trace (Parent & Zucker 1989).

Note that, co-circularity is related to Blum’s notion of symmetric chord coordinates

(Blum 1973) and Brady and Asada’s local symmetries (Brady & Asada 1984). In-
stead of relating the pair of tangents through an arc of circle, link them by a straight
line segment. If they are co–circular then this line segment is a chord (w.r. to the arc
of circle) and the angles each tangent makes with respect to this chord are equal
(e.g., see Parent & Zucker 1989: Fig. 5(b)). The extension to surfaces involves the
classical notion of “parallel transport” from differential geometry.

Example 4.2 (2-D manifold: Parallel transport) Let V be a tangent vector field over
a region B of the surface M. How can we relate (or move) a given tangent vector v at
point p ∈ M with another tangent vector w at point q ∈ M? One possibility is through
the notion of (Levi–Civita) parallel transport where a tangent vector, e.g., v, is moved
along a curve α ⊂ M such that it stays parallel with respect to the tangent space TM .
Thus, v may rotate, but only with TM (this depending on the intrinsic or Gaussian
curvature of M).
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.22 (1495-1543)

 Frederic Fol Leymarie

In computer vision, such a notion of parallel transport has been used by Sander
and Zucker in the context of the recovery of the trace of surface points (Sander &
Zucker 1990).25 But, note that a surface can curve in more than one way at any given
point. In particular, for smooth surfaces one can show that at any point of the man-
ifold (normal) curvatures range from a minimum, kmin , to a maximum, kmax , called
principal curvatures. Thus, the object used for relating tangent vectors on M cannot
simply be an osculating sphere (the extension to 3D of the 2D osculating circle above),
but rather needs to be a more flexible object. In the context of differential geometry, the
“canonical” object used is the osculating paraboloid, S, a parabolic quadric surface
having second order contact with the manifold M (Pogorelov 1958: 87).

Therefore, we can compare/relate tangent vectors on M through parallel transport

via an interpolating paraboloid. We need the latter because we do not know a-
priori the exact “form” (trace) of M. One usually has only partial/noisy knowledge
of the trace of M. This is an instance of a “horizontal transform sequence.”

Remark 4.1 (Inadequacy of parallel transport) Sander notes that parallel trans-
port is not the “ideal” way of relating tangent vectors of a surface for the following
reasons (Sander 1988: 38–39):
– Which curve/path to choose on the paraboloid? There is potentially an infinity
of them and, although parallel transport is well-defined for a given curve (i.e.,
it is unique), it gives, in general (when K = 0), different results for different
curves. I.e., the holonomy angle varies with the path, and this is a function of the
Gaussian (intrinsic) curvature of the paraboloid.
– How to transport “directions” (associated with a given initial tangent vector, v,
and its corresponding normal curvature) so as to preserve them? Clearly, arbitrary
paths will not preserve such directions. An interesting choice might be lines of
(principal) curvature, but, then there is the problem of “navigating” through sin-
gularities of the tangent field, i.e., through umbilics (points where all curvatures
are equal: kmin = kmax ).

Sander proposes the notion of “compatibilities via frame bundles” to alleviate

these problems (Sander 1988).26

Example 4.3 (2-D manifold: Compatibilities via frame bundles) The local infor-
mation available in a cross section, σ, of the frame bundle via the first and second
fundamental forms, IS and II S ,27 of the interpolating paraboloid, S, enables the direct
generation of the frame ξα (q) corresponding to moving ξα from p to q.28 This moving
frame will tell “what the local information at” q “looks like according to the informa-
tion at” p (Sander 1988: 41). Thus, it gives “support” to the information at q, “under
the assumption of an underlying smooth surface” (i.e., S).
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.23 (1543-1596)

Thoughts on shape 

This is again an horizontal transform sequence where frames are “moved” accord-
ing to an interpolating surface which best approximates (through some optimisa-
tion measure) the actual surface M we wish to recover (the trace of).
Example 4.4 (2-D manifold: “Classical” Shape from Shading (SfS)) Shape from
shading consists in “inferring geometric structure of a scene from measurements of
its photometric structure” (Langer & Zucker 1994).
The “classical” model studied in the context of this problem, is based on the “Im-
age Irradiance Equation” (IIE) which states that luminance, or image intensity I, is
proportional to some reflectance function or map, R : I(x) = R(N(x)), where R is
assumed dependent on the surface normal, N, i.e., dependent on the surface orienta-
tion. Here, the surface trace is the geometric structure one wishes to recover from the
image data, i.e., it is an “inverse optics problem.” A special case (the “classical” one) is
when the surface is assumed Lambertian,29 i.e., the light source is assumed punctual
and remote (far), and the albedo (the amount of light reflected, a material property),
ρ, is assumed constant over the surface. Then, the IIE can be expressed as (Langer &
Zucker 1994):
I(x) = ρ Is L · N(x) + ρ Ia ,
where L and Is are the direction and intensity of the distant point source, respectively,
and Ia is the intensity of the ambient illumination. The (constant) term ρIa is usually
ignored in the calculations. The term ρIs L is assumed known (or is approximated
empirically). Then, assuming one faces a Monge patch, i.e., the surface topography
can be represented as the graph of a function, z = f (x, y), which is always possible
locally for smooth surfaces, the (unit) normal vector N can be expressed as a function
of the gradient of the surface, i.e., in terms of the slopes, p = ∂z/∂x and q = ∂z/∂y, of
the surface in the x and y directions, respectively:
N · N = vx ∧ vy = (1, 0, p) ∧ (0, 1, q) = (–p, –q, 1) ,
where the vectors vx and vy are tangent vectors to the surface in the x and y direc-
tions. Similarly, for a distant source, we can represent its direction L via a unit vector
(–ps , –qs, 1) expressed in terms of the slopes ps , qs , of a (wave) surface perpendicu-
lar to the light ray. Thus the IIE, in the Lambertian–single–distant–light–source case,
reduces to (Horn 1986: 219):
1 + ps · p + qs · q
I(x, y) = R(p, q) =  ,
1 + p2 + q2
which has been normalised by a factor ρI0 . This equation can be seen as a nonlinear
first order PDE.
Different ways of solving numerically the reduced IIE equation above have been
used in the computer vision literature. All of these are instances of horizontal trans-
form sequences where one, starting from an initial contour, usually taken as an
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.24 (1596-1648)

 Frederic Fol Leymarie

extrema of I, grows solutions along the surface by (numerically) integrating the IIE.
Essentially, one takes small steps from a known height/position30 of the surface
and recovers surface height at the new position.31 To “validate” the calculations
one can integrate solutions on the calculated surface, this along closed paths from
and to the initial (known) starting position; thus, the difference in height should
vanish.32
Example 4.5 (Motion field & optical flow) One may consider a motion (vector)
field as the “object” of interest. The layout of the field is transformed, e.g., under as-
sumptions of smoothness, thus retrieving a smooth vector field possibly corresponding
to some translation and/or rotation of a real physical object in the scene.33
A common hypothesis made in computer vision is that the motion field, which as-
signs a velocity vector to each point in an image, corresponds to the optical flow which
is the apparent motion of the brightness pattern.34 Note that this correspondence is an
idealisation; in many (practical) situations, the motion field and the optical flow are
not equivalent. Nevertheless, optical flow is believed to be a potentially important
source of information towards the retrieval of the true motion field or of some of its
features.
Under the (very restrictive) assumption that brightness varies smoothly one can
derive the Optical Flow Constraint Equation (OFCE) (Horn 1986: 282), which is
nothing but the 1–jet for I:35
Ix u + I y v + It = 0 ,
where u = dx/dt and v = dy/dt are the optical flow components, and the (partial)
derivatives Ix , Iy and It are estimated from the image. We may re-write this equation
as follows:
∂I
+ ∇I · (u, v) = 0 .
∂t
This can be shown to correspond to a conservation law for an incompressible
medium, where the image intensity, I, corresponds to the density of the medium (Fitz-
patrick 1988). In this context, the assumption of incompressibility for a density image
is analogous to the assumption that image brightness of a surface point is independent
of the motion of that point (a very restrictive assumption indeed).
Note that the OFCE is a local constraint on brightness velocities. In order to re-
cover a “consistent” (global) optical flow, one has to filter or transform somehow
the initial flow field given by the values (u, v) at each image point. E.g., one may
convolve the field with an averaging kernel. Or one may update information on the
basis of the available local geometric structure (i.e., tangent direction, curvature)
within a relaxation network. The latter is reminiscent of ex. 4.3, where compatibil-
ities are (here) in terms of tangent bundles, i.e., the problem can be cast as one of
selecting consistent sections σ through the tangent bundle of the “motion surface”
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.25 (1648-1706)

Thoughts on shape 

(i.e., motion field). Thus the flow field is integrated along the underlying surface
(or “base”), and we can picture it as an horizontal transform sequence.36
I covered in the above a number of illustrative examples of horizontal trans-
form sequences from the field of computer vision, from contour tracing to shape
from shading and shape from motion. In the next section, I consider the “comple-
mentary” vertical ones, where the transforms operate in the space away from the
outline of an object, rather than along it.

. Vertical transform sequences

Definition 4.2 (Vertical transform sequence) It is a shape transform sequence, such

that the first transform operates in the normal space (or the normal bundle) of the
outline of the object; in other words, subsequent transforms are constrained to act
above (or below – but not along, i.e., acting away from) the outline of the object.
Example 4.6 (Medial symmetry graphs (MSG) in 2D) There exists a few different
but equivalent or related definitions of the concept of a medial symmetry set of a
2D shape, taking the form of a skeletal graph.37 One interest in having a variety of
definitions is that each one leads to different operational schemes for implementing
the transform sequence. Below we give the main different definitions to be found in
the literature.
Contact circles: Consider the closure of the loci of centres of disks tangent to the out-
line at two or more points; such bitangent disks are called “contact circles.” The
centres of these circles trace a symmetry set (SS ) which in 2D takes the form of
a planar graph. The most commonly used subset of the SS is the medial axis
(MA) of Harry Blum (Blum 1973) for which all contact circles are constrained
to be maximal, i.e., such that they do not contain any points of the outline other
than the contact (tangency) points.38
Quenching waves: Imagine a wave propagating from the boundary with a constant
speed. Shocks occur where wavefronts merge, intersect or collapse. The locus of
each shock correspond to the centre of a contact circle part of the SS , since such a
point is at equal distance from the two contact points on the outline, generators of
the pair of quenching waves. Or imagine a fire propagating over a grassfield and
initiated by the outline of the object. The loci of quenching points of the grassfire
correspond to Blum’s MA. That is, if we let waves intersect and keep propagating
we get the SS , and if we terminate propagation as soon as waves create a first
shock we get the MA.
Voronoi diagrams and cut locus: Consider the set of points being closest to a given
outline sample. This is the natural framework to work with the Voronoi Diagram
(VD) (Okabe et al. 2000), a complete partition of space made of cells whose
boundary, the Voronoi Graph, is equivalent to the MA. Each cell (or Voronoi
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.26 (1706-1760)

 Frederic Fol Leymarie

region, VR) is made from the loci of the R2 plane closest to one single outline
point. A related concept is that of the Cut Locus, the closure of the set containing
all points that have at least two shortest segments to the outline (Wolter 1992).
Ridges of a distance surface: Consider the surface generated by applying a distance
transform to the object’s form. Map distance as a height function. Then consider
the loci of points on ridges of this topographical map (Kotelly 1963; Blum 1973;
Leymarie & Levine 1992b).
Mid-chord symmetry and process inferring symmetry axis (PISA): An alternative
to the centre of contact circles, as is used to trace the MA, is to consider instead (i)
the mid-chord point defined between the two contact points on the outline (Blum
1973; Brady & Asada 1984), or (ii) the mid-point of the shortest subtented arc of
the circle (Leyton 1987).
Shock graph: Consider the radius function associated to the contact circles. It defines
a (1D) vector field along the trace of the MA. The singularities of this function
when linked by the intermediary MA segments define a more compact graph
representation of the MA. These singularities, the sources, relays and sinks of
the radius function, correspond to endpoints, junctions and saddle points of the
MA (Kimia 2003).

Now let us consider the boundary of the object as either a piecewise 1D manifold
or a discrete set of samples. Then the MSG is a transform which maps this bound-
ary onto the medial symmetry set in one of the forms given above. This transform
acts along normal rays to the boundary of the shape. That is, it acts along the nor-
mal bundle. Points where normal rays meet correspond to symmetry points. This
is an instance of a vertical transform sequence.

Concept 4.1 (Wolter’s topological shape theorem of the MA) Franz-Erich Wol-
ter has shown, in his early works, how the MA of a solid O (i.e., such that the outline
defines an inside and an outside for the object under scrutiny) “contains the essence
of the topological type” of O (Wolter 1992; Wolter et al. 2004). In summary, Wolter
has proven for the cases of smooth outlines and, in 2D, for piecewise-smooth outlines,
that the MA is the deformation retract of the solid O; and, therefore, it preserves
the homotopy type of O, which can be retrieved, for example, from the (topological)
graph structure of the MA.

Remark 4.2 (Relation between the MSG and co-circularity mapping) In the case
of the MSG, two points are put in relation with each other by searching where their
normal rays meet. This locus defines the contact circle.
In the case of the co–circularity mapping, the osculating circle is used to relate
pairs of tangents of boundary points by using the circular shape as an interpolant (ex.
4.1, §4.1).
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.27 (1760-1818)

Thoughts on shape 

Thus, the same object, a circle in close contact with the outline, is used in two
different manners, for relating pair of points of the boundary which share a “local
symmetry,” albeit of different nature.

Remark 4.3 (Planar and space curves) In the above we considered only outlines
tracing planar curves in R2 . The discussion also applies to space curves in R3 , al-
though the normal bundle is then more complex: it involves both the normals and
bi–normals (Bruce & Giblin 1992).

Example 4.7 (MSG on curved surfaces) Wolter et al. have extended the concept of
a medial symmetry graph from the R2 plane to curved surfaces embedded in R3
(Rausch et al. 1996; Kunze et al. 1997; Wolter & Friese 2000). They use the no-
tion of the cut locus for this purpose, where shortest paths are now defined along
shortest geodesics from an outline defined on the curved surface.39 In essence, this is
an extension from the Euclidean space (flat 2D manifolds, i.e., with zero curvature)
to the Riemannian space (curved 2D manifolds).

The above notion of medial symmetry sets for 1D outlines is naturally extended to
2D outlines, i.e., to the traces of surfaces of 3D objects.

Example 4.8 (Medial symmetry sets in 3D) Medial symmetries of 3D objects can
be captured by replacing contact disks by contact spheres.

Again, one may view these medial symmetries in different flavours. For example,
a wave might propagate from the surface boundary along the normal rays to the
boundary. Singularities of this normal mapping correspond to the symmetry set.
Most of the definitions used for the 2D case have been extended to 3D, with the
exceptions of mid-chord symmetries and Leyton’s PISA.

Remark 4.4 (Medial symmetries in 3D and compatibilities via frame bundles) In

both cases we are putting in relation pairs of surface points. In the case of medial sym-
metries we do so by navigating along normal rays to the surface, while in Sander and
Zucker’s horizontal transform sequence one navigates along the surface according to
an interpolating paraboloid.
Note that different objects are used to relate pairs of points: osculating paraboloids
in the horizontal transform sequence of Sander and Zucker and contact spheres for
the “vertical” medial symmetries.
Note also that, the “compatibilities via frame bundles” are used in practice to
relate surface points within a relatively small neighbourhood along the surface – a
kind of “sausage” neighbourhood also called “thick–trace” – while in the vertical case,
much larger (3D) neighbourhoods may be used to compute symmetries. In principle,
all the available region of space under scrutiny is used.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.28 (1818-1877)

 Frederic Fol Leymarie

The main difference between medial symmetry sets in 2D and 3D, is that, in 3D,
for outlines sampling the surfaces of volumetric object, the resulting set is not a
graph in general, but rather a hypergraph made of intersecting medial surfaces; in
other words the medial set has one more dimension. To benefit from the use of a
graph representation, e.g., in recognition tasks, a number of additional transforms
are available.
Example 4.9 (Medial symmetry graphs in 3D) Medial symmetries of 3D objects
constrained to take the form of a 3D (non-necessarily planar) graph. Below we give
the main definitions to be found in the literature.
Generalised cylinders and 3D curve skeletons: A generalised cylinder is defined by
(i) an axis, (ii) a radius functional, often called “sweep function,” and (iii) a set
of rules to define branching points and behaviours. The axis is typically a 3D space
curve. The radius functional specifies the trace of the outline perpendicularly to
the axis, and may vary as it “sweeps” along the axis.40 The set of rules defines what
happens at the end points of an axis: in which directions other axes with their own
radius functional may branch-out. While generalised cylinders are useful in gen-
erating pre-specified objects, 3D curve skeletons are an attempt to define from the
outline an equivalent representation. These come under different names, repre-
senting slightly different definitions which we organise in two main categories: (i)
curve skeletons (Borgefors et al. 1999), skeleton graphs (Sundar et al. 2003) or
ik-skeletons (Wade & Parent 2002) are obtained by thinning a discrete distance
field, such as resulting from an object voxelisation, and (ii) Reeb graphs (Hilaga
et al. 2001) also called Level Set Diagrams (Verroust & Lazarus 2000) are ob-
tained by a simulated flooding of the object as a solid, tracing midpoints of level
sets as the object is filled-in.
Medial and shock scaffolds: The 3D MA of Blum has been generalised by Ley-
marie, Kimia and Giblin by defining a unique 3D graph structure from which
the MA itself can be recovered (Leymarie & Kimia 2001; Leymarie 2003; Gib-
lin & Kimia 2004; Leymarie & Kimia 2007). This is an extension of the concept
of a shock graph in 2D and it comes into two main types: (i) the Medial Scaf-
fold (MS ) which links the singularities of the radius function associated to the
MA for two types of medial curves: axial shock curves where three or more MA
surfaces intersect, and rib curves which are associated with surface ridges of the
object’s outline; (ii) the Shock Scaffold (SC ) which extends the MS by defining
additional curves linking the singularities of the radius function along the MA
surfaces.
Critical nets and molecular graphs: Consider each 3D Voronoi region, VR, of the
Voronoi diagram, VD, and its decomposition in a set of k-faces which are home-
omorphic to a closed ball, i.e., such that each VR can be continuously deformed
to a closed ball of same dimensions (Bredon 1993). The collection gives a finite
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.29 (1877-1926)

Thoughts on shape 

closed covering of Rn implying that the VD is a CW-complex representation of

space.41 Another classic result, from Morse theory, a branch of differential topol-
ogy (Milnor 1963), associates to each of the Voronoi k-face a critical point of
index n – k for a given Morse function, which we can in practice take as a func-
tion of the distance to the outline.42 The graph obtained by connecting in this
“n+1-level hierarchy” the critical points, seen as nodes, via links which are unique
topological “separatrices” is called a critical net in Crystallography (Johnson et
al. 1996; Johnson 1999), and a molecular graph in Computational Chemistry
(Bader 1990: Ch. 2 & 3); such graphs are used to described atomic structures, e.g.,
crystals or proteins, and the topology of their charge density.43

Generalised cylinders and 3D curve skeletons are well-adapted to capture the

shape of tubular-like objects. When describing more general objects in 3D, the
scaffolds, MS or SC , are to be preferred. While, in theory, the Critical nets and
Molecular graphs can be seen as an extension of the SC , they have only been
studied for the restricted case where the outline is sampling an outward surface
bounding a density cloud centred on atomic centres of molecules, and for non-
degenerate underlying Morse functions, constraints not imposed on the MA and
its representation as a graph via the MS or the SC .
Similarly to the 2D MSG case, all of the above transform sequences in 3D
operate away from the outline, along the normal rays to the object’s underlying
surface: they are all instances of vertical transform sequences.

Example 4.10 (Shape from shading: Case of diffuse light source conditions)
Paraphrasing from the work of Langer and Zucker (Langer & Zucker 1994):
Under diffuse lighting conditions, surface luminance depends primarily on the
amount of sky visible from each surface patch, while surface normal and mutual
illumination are of secondary importance. This dependence – on the visible sky –
is modelled in terms of a surface aperture function: the percentage of incident di-
rections visible in a local hemisphere above “ground” or above a surface patch. An
approximately linear constraint between image intensity and surface aperture is
found.
While in the case of “classical SfS” (§4.1), where one assumes a point light source
(at infinity), the local geometry constraints lie along the surface, in the case of the
diffuse lighting problem, the local geometric constraints are found in the ambient
light above the surface.

In this “SfS on a cloudy day,” once surface apertures are estimated from an initial
image, a recursive marching algorithm is used to refine depths for a viewer as-
sumed seeing the scene from above. The entire space is discretized via an Eulerian
(fixed) grid and depth is propagated one node at a time; the method resembles
a blanket slowly falling on the scene which starts folding locally according to the
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.30 (1926-1985)

 Frederic Fol Leymarie

surface aperture values. All computations take place in the ambient space above
the object’s surface, yet another instance of a vertical transform sequence.

Example 4.11 (Biological motion field: Compact medial node representation)

Ilona Kovács et al. have proposed that special medial nodes along a thick trace of the
MA represent a plausible natural and effective mechanisms for the computation of
motion fields in animal visual systems. A pseudo-distance function, Dε , is defined
which captures medial symmetries within a pair of disks defining an annulus region,
where ε denotes the thickness of the annulus. Under this metric, the special nodes of
the modified MA are those which locally maximise the amount of outline trace they
capture: they are “the most informative points along the skeleton” (MA) of an object
in motion (Kovács et al. 1998). Such a compact medial representation is inspired by
the chronophotographs of Etienne-Jules Marey (circa 1880) and G. Johansson’s work
(circa 1970) on the perception of biological motion for point-light walker displays.

The Dε distance function of Kovacs et al. is computed for each frame of a video
sequence, and can be applied in each frame directly to the available outlines. It is
comparable to other “annular” medial symmetry schemes (Kelly & Levine 1995;
van Tonder & Ejima 2000, 2003), and this group also falls in the class of vertical
transform sequences. Note that as ε goes to zero, the method may converge to the
shock graph representation, where the special medial nodes along the MA are
those for which the associated distance (or radius) function goes through a critical
point (i.e., a source, relay or sink).

. Mixed transform sequences

It is possible to define sequences of transforms which mix horizontality and ver-

ticality. Typically, this proves useful when one considers instantiation of general
deformations. An important example of such mixed transform sequences, which
we briefly consider below, is the so called reaction-diffusion case.

.. Reaction-diffusion transform sequence

Definition 4.3 (Reaction-diffusion transform sequence (RDTS)) Transform sequence
which operates in both the normal and tangent space of the object’s outline on the basis
of a differential equation. The action of the RDTS is equivalent to a combined normal
flow – the “reaction” part – and a tangential flow – the “diffusion” part.

Example 4.12 (Curve evolution) Let the evolution of a piecewise continuous 2D

outline be modelled by a differential equation of the form:
∂α
= Υ(·) N ,
∂t
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.31 (1985-2035)

Thoughts on shape 

where α(s, t) = (x(s, t), y(s, t)) is the position vector of the outline, s is the length
parameter (not necessarily arclength), Υ(·) is the deformation functional, and N is the
unit normal vector (typically oriented “inward” with respect to the object’s outline).
The following deformation functional: Υ ≡ β0 + β1 κ, where κ stands for the
curvature of a curve segment α, is often considered in differential geometry (curve
shortening and smoothing), physics of interfaces, and computer vision (Catté et al.
1992; Kimia et al. 1995; Sethian 1999).

When β0 = 0 the evolution reduces to a pure diffusion (horizontal) transform

sequence: a curvature driven evolution which shrinks the curve (to a circular dot
for a closed outline). When β1 = 0 the evolution reduces to a pure reaction (ver-
tical) transform sequence: an eikonal wave propagation for the outline developing
shocks (MA symmetries). When both effects are active (i.e., β0 = 0 and β1 = 0)
shape is visualised as a scale-space of deformations (Kimia et al. 1995). How to ex-
plore this scale space is neither systematic, nor obvious. Furthermore, it depends
on two “parameters” (β0 and β1 ) which may be adapted to different applications.

Example 4.13 (Alan Turing’s morphogens) The most famous RDTS was proposed
by Alan Turing (circa 1950) to model the changing spatial concentrations of chemical
substances he called “morphogens” (Turing 1952). By varying the relative influence
of reaction versus diffusion, different regular patterns observable in nature have been
simulated, such as the stripes of zebras and the spots of leopards (Murray 2004).

Having introduced in the above a classification of transform sequences based on

the dichotomy of their action along the outline or away from it, I now consider in
the next section systemic definitions of transform sequences in the form of “gram-
mars:” i.e., sets of rules permitting to group, assemble, separate, re-organise, object
elements with the goal of capturing complexity in forms.

. Grammars for shape

Definition 5.1 (Shape grammar) (i) It consists in a set of rules which permits to
replace one outline by another, by giving a “shape–interpolation,” i.e., a plausible de-
velopmental history. (ii) Alternatively, it is a set of rules which permits to add/delete
(or cut and paste) parts/components to an outline as a function of its derived shape.

An example of the former is Leyton’s process-grammar (§5.1) and an example of

the latter is Richards and Hoffman’s codon–grammar (§5.2), both of which will be
described below.
Thus objects of various complexities may be generated by an iteration of a
finite number of simple outline transformation instructions. As such, a shape
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.32 (2035-2084)

 Frederic Fol Leymarie

grammar provides a discrete form of a language which can be based on shape

transforms.
Various shape grammars have been developed in the literature. E.g., the gener-
ation of self–similar fractal objects is possible with very simple grammars (Lord &
Wilson 1984: §8.1). Selection rules, which forbid the addition of a subunit under
certain conditions, have been used by Ulam to generate less regular patterns (Lord
& Wilson 1984: §8.2). Trees and river systems, crystals, tessellations and space fill-
ing organisations are other examples of domain of applications of shape grammars
as object generators (Lord & Wilson 1984). An important example of early work is
to be found in “L-systems,” also called “Lindenmayer systems” or “parallel string-
rewrite systems,” which are made from productions rules used to define a tracing
of piecewise linear segments with joints parameterised by rotation angles (Linden-
mayer 1968; Prusinkiewicz & Lindenrnayer 1999). These rules also are a compact
way to iteratively repeat constructive sequences in the description of fractals, often
used to model groups of plants, flowers, leaves, and so on (Ferraro et al. 2005).

Remark 5.1 (The inverse problem of vision) Note that, for the representation and
description of an object, toward its recognition, a shape grammar will prove particu-
larly useful if it is possible to “invert” the rules, or at least identify them, when given
a particular outline of an object. This would then permit to identify the process which
gave the “final” shape, or alternatively would give cues in identifying parts or object
components.

Example 5.1 (Biederman’s recognition-by-components theory) The RBC theory

of object recognition by Biederman et al. is an important example of a shape grammar
which aims at addressing the “inverse problem” of vision (Biederman 1987; Hummel
& Biederman 1992; Biederman 1995). The RBC theory starts from the hypothesis
that a finite compact family of (36) object parts (or components) called geons is suffi-
cient to represent most commonly encountered objects. Geons include primitive parts
(or geometric units) such as blocks, cylinders, spheres, which are parameterised by a
number of (“non-accidental”) features encoding size, tapering, symmetry axes, bends,
and so on. Geons can be extracted or re-constructed from outline traces in images.
Their sequencing in representing a complex object is specified by relations (geometrical
and topological) amongst parts.

Since its proposal in the late 1980’s the RBC theory has been implemented in a
few systems aiming at performing object recognition. Bergevin and Levine im-
plemented the first comprehensive system based on RBC called PARVO (Primal
Access Recognition of Visual Objects) (Bergevin & Levine 1993). Wu and Levine
restricted Biederman’s set of 36 geons to only 7 defined as parameterised subsets
of superellipsoids for potentially faster processing (Wu & Levine 1997). Of related
scope is the earlier work by Pentland et al. using parameterised object primitives to
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.33 (2084-2143)

Thoughts on shape 

model parts (Pentland & Scarloff 1991). An account of the various approaches and
issues in using geons until the mid-1990’s can be found in (Dickinson et al. 1997).
More recent work can be found in (Siddiqi & Kimia 1995; Siddiqi et al. 2001; Sing
& Hoffman 2001).
Remark 5.2 (Dynamic shape and genetic grammars) One can generalise shape
grammars within the context of cellular automata where some randomisation is in-
troduced in the manifestation of the “rules” leading to “dynamic” shapes; for example
see the works of Wolfram et al. (Wolfram 1994) and more recent studies in biological
pattern genesis (Deutsch & Dormann 2005).
Another possible generalisation is in the context of genetic programming where
mutations and the natural mixing of a pool of “genes” (possibly representing shape
components or features) is used to obtain evolving “natural” shapes; for example see
the early works of Dawkins on biomorphs (Dawkins 1986), and of Latham and Todd
on genetic art (Todd & Latham 1992), and more recent works in arts and design
(Bentley 1999) and biological sciences (Kumar & Bentley 2003).
In the following I consider and compare two important grammars which emerged
in the field of visual perception analysis in the 1980’s, this to illustrate some of the
issues one has to face in designing a “shape language.”

. Leyton’s process–grammar

Leyton’s starts from the assumption that “shape [can be] understood as the out-
come of [physical] processes that formed it” (Leyton 1988). Thus, a shape is “de-
fined as the outcome of a history which is psychologically decomposed into phases
where qualitatively different processes were acting;” and, “categories of objects
are psychologically stratified into levels corresponding to phases in a formation
history.”
Leyton has looked at two types of process–history: when considering the ob-
ject as a whole, and when considering it only through a few salient features – i.e.,
the curvature extrema of the outline. In the following I only consider the latter
approach.
For smooth, planar, simply connected outlines, Leyton shows that a simple
grammar of six operations is sufficient to represent the relationships between any
two shapes.
In order to retrieve the processes which are (potentially) acting on a shape,
Leyton suggests that it is necessary to recover the curvature extrema of the bound-
ing contour of an object and its associated symmetry axes. The inference of pro-
cesses from a single outline requires two stages
1. From curvature extrema to symmetry axes.
2. From symmetry axes to processes.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.34 (2143-2205)

 Frederic Fol Leymarie

Curvature extrema and symmetry axes are put into correspondence through the
following principle.44

Theorem 5.1 (2D symmetry–curvature duality (Leyton 1987)) Any segment of a

smooth planar curve bounded by two consecutive curvature extrema of the same type
(i.e., max. or min.) has a unique differential symmetry axis which terminates at a
curvature extremum of the opposite type (i.e., min. or max., respectively).45

Thus, this “duality” between curvature extrema and symmetry axes gives a rule for
relating them. Furthermore, symmetry axes are postulated as the “records of de-
formational processes.” E.g., the longer an axis is, the more significant the attached
process shall be. A process in this context may be thought of as an internal or ex-
ternal force acting on the shape. Significance may be taken in terms of magnitude
or duration in time.

Proposition 5.1 (Interaction principle (Leyton 1984)) The symmetry axes of a

perceptual organisation are interpreted as the principal directions along which pro-
cesses are most likely to act or have acted.

This proposition was justified as follows by Leyton (Leyton 1985, 1988):

A transformation, acting on an organisation, is one in which symmetry axes

become invariant lines or eigenspaces under the transformation. Thus the trans-
formation will tend to preserve the symmetries – i.e., be structure–preserving on
the organisation.

By combining the above two principles one gets:

Corollary 5.1 ((Leyton 1988)) Each curvature extrema implies a process whose
trace is the unique symmetry axis associated with, and terminating at, that ex-
tremum.
Thus, a process explains the extrema.

Corollary 5.2 (Asymmetry rule (Leyton 1988)) Processes are understood as creat-
ing greater curvature variation, i.e., greater information in the sense of mathematical
information theory.46

Remark 5.3 (Simplest shape is the circle) If one uses the “asymmetry rule” above to
extrapolate backward in time, the “ultimate starting point for the shape must have
been the circle.” The latter possess the least amount of curvature variation, hence
information.

Definition 5.2 (Types of curvature extrema) Four types of curvature extrema are
defined: local minimum and maximum, denoted by m and M, respectively, each being
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.35 (2205-2289)

Thoughts on shape 

possibly of negative (–) or positive (+) curvature. The semantic interpretation of these
extrema is as follows:
M+ : “Protrusion” m– : “Indentation”
m+ : “Squashing” M– : “Internal resistance”

Leyton’s Process–grammar
The process–grammar consists of two types of extrapolation rules: Continuation
and Bifurcation rules, denoted by C and B, respectively. Continuation implies
that no new extremum is created during the (deformation) process (but an ex-
tremum may change in nature), while a bifurcation implies that a new, interven-
ing, extremum must be introduced. The (rewrite) rules of the grammar can be
summarised as follows:
CM + : M+ → 0M + 0 “Protrusion continues”
Cm– : m– → 0m– 0 “Indentation continues”
Cm+ : m+ → 0m– 0 “Squashing continues until it indents”
CM – : M – → 0M + 0 “Internal resistance continues until it protrudes”

BM + : M + → M + m+ M + “Protrusion bifurcates”
Bm– : m– → m– M – m– “Indentation bifurcates”
Bm+ : m+ → m+ M + m+ “Protrusion introduced”
BM – : M – → M – m– M – “Indentation introduced”
NB: Continuation at protrusion, M + , and indentation, m– , are structurally trivial:
i.e., no change in the extremum’s nature occurs. As such they do not need be con-
sidered explicitly in the grammar. Thus, only the last six rules are to be used in
practice.

By applying the rules of the above process–grammar to simple objects, such

as an ellipse, one can generate a succession of more complex outlines. Note that,
all outlines can in principle be derived from the simplest one in the hierarchy. It
turns out that the shape–space so generated is highly structured (e.g., see Figures
16 and 17 in Leyton (1988)). However, there exists more than one “path,” in gen-
eral, through this shape–space, when travelling from a simple outline to a more
complex one. Leyton suggested the following simple heuristic for constraining the
search in this space:

Heuristic 5.1 (Size–is–time (Leyton 1988, 1989)) In the absence of information to

the contrary, size corresponds to time. I.e., later processes have a shorter time to de-
velop than earlier ones and thus in general they tend to result in smaller deformations
of an outline.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.36 (2289-2356)

 Frederic Fol Leymarie

Leyton also suggested that deblurring of the boundary (coordinates) could provide
a way of finding a single interpretation or path through the shape–space (Ley-
ton 1989). Leyton claims that deblurring is the appropriate shape transform to
use because it “embodies the size–is–time heuristic;” and in particular: “deblur-
ring incrementally moves the boundary in the direction of increased curvature
variation.”
Remark 5.4 (Figure–ground reversal) Figure–ground reversal is simply dealt with
by introducing a duality operation, D, which interchanges extrema labels as follows:
M+ ↔ m– : protrusion ↔ indentation
M– ↔ m+ : squashing ↔ internal resistance

. Richards and Hoffman’s codon–grammar

This is an extrema–grammar rather than a process–grammar, i.e., it is based only

on the curvature extrema of the object’s outline (Richards & Hoffman 1985).
Definition 5.3 (Codon) A boundary segment made of a triple of extrema, such that
two are minima, m, and are endpoints for the segment.
There are five types of non–trivial codons for planar, smooth, and simply–
connected curve segments: {m+ M + m+ , m– M – m– , m+ M + m– , m– M + m+ , m– M + m– }.
These are denoted 0+ , 0– , 1+ , 1– , 2, respectively (Richards & Hoffman 1985) – i.e.,
where the numeral represents the number of zeroes of curvature present in the curve
segment.
If one considers there duals, to account for figure–ground reversal, one gets a
family of ten codons (Leyton 1988).

Codon–grammar
The codon–grammar consist of two types of rules:
Level–preserving: Replace a codon by a codon.
Level–increasing: Replace an extremum with a codon or replace a codon with a
pair of consecutive codons.

The codon–grammar consists of eighteen rewrite rules (see Leyton 1988: Tbl. 2 for
a list of these).

Remark 5.5 (Comparison of the process and codon grammars)

1. The Process–grammar is more economical: only six rules are necessary in com-
parison of the eighteen rules of the Codon–grammar.
2. Every codon rule can be represented by a sequence of Process–grammar rules.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.37 (2356-2428)

Thoughts on shape 

3. The codon substitutions are cutting and pasting operations, while the Process-
grammar rewrite rules are developmental, i.e., concerned with growth.

Note that both the Process and Codon grammars can be implemented on the basis
of horizontal or vertical transform sequences, to identify curvature extrema and
their relationships. An example of an horizontal approach is the curvature mor-
phology system of Leymarie and Levine which maps an outline to a set of valleys
and peaks of curvature where the scale of an extremum is characterised by the size
of a valley (mapping to a concavity for the outline) or a peak (a convexity for the
outline) (Leymarie & Levine 1989). An example of a vertical approach is the work
of Tek and Kimia based on the 2D MA where the collection of axes can be used
to rank order a set of curvature extrema (sitting at one end of each axis) by a no-
tion of significance or scale: how important each extremum is in relation to other
nearby ones and the overall shape (Tek & Kimia 2001). Significance is directly as-
sociated to symmetry axes in how long they are or how large an object’s sub-part
they capture; as such this system provides an implementation of some of Leyton’s
ideas, and in particular his “size is time” heuristic (5.1).
An important open question when designing a practical shape language, is
whether a discrete form of a family of transforms for shape is more adequate than
a continuous one, such as those provided by the Reaction-Diffusion Transform Se-
quences (RDTS; §4.3.1). Note that a discrete, rule-based approach provided by
a shape grammar such as Leyton’s or the Codon system, has the appeal of being
more easily implementable in a reasoning system; e.g., via an expert-system. Do-
ing reasoning in a continuous shape–space derived from an approach such as the
RDTS does not appear, at first glance, as easy, since the space is not explicitly as
well “compartmentalised.” A similar issue occurs in neuroscience where the lead-
ing paradigm is one where perception relies on discrete time events and processing
(VanRullen & Koch 2003).

. Conclusion

In the above “thoughts” I have tried to characterise the meaning of what “shape”
is and how it can be related to fundamental mathematical concepts, via the notion
of transforms: mathematical constructs permitting to characterise the probing of
the structural field of objects, thus revealing their shape. I have also given many ex-
amples based on my understanding of visual perception and computer vision, to
illustrate and better focus the effort. A particular emphasis was the definition of a
bi-partite classification of sequences of transforms, whether they first act along the
outline or away from it: what I call horizontal and vertical transform sequences.
This dichotomy is useful as a mean to compare and distinguish popular shape
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.38 (2428-2457)

 Frederic Fol Leymarie

transforms, as was illustrated with examples from the field of computer vision,
and it could be extended to other “visual” fields. For example, in CAD-CAM (Pa-
trikalakis & Maekawa 2002), one can distinguish methods based on the differential
geometry of curves and surfaces as well as of intersections – essentially sets of hor-
izontal transform sequences – from the methods based on offsets for curves and
surfaces – i.e., vertical transform sequences. How the horizontal versus vertical
approaches precisely compare, when applied to similar or identical problems –
e.g., when identifying corners or curvature extrema – is a study which remains
to be conducted. Furthermore, such a dichotomy may also be useful to study the
human visual system itself. The hypercolumn system of receptive fields first iden-
tified by Hubel and Wiesel (Heeger et al. 1996) can be understood as the basis of a
(differential) geometric engine used to generate tangent bundles, curvature maps,
and so on, from outline samples (Zucker 1987); in other words, we can portrait (at
least part of) the hypercolumn system of the visual cortex as implementing certain
horizontal transform sequences. Kovács et al. have observed in human subjects
differential contrast sensitivity maps for 2D shapes which are consistent with a
medial function representing the percentage of outline samples equidistant from
the observation point within a tolerance level (Kovács et al. 1998). This has lead
them to hypothesize a medial MA-based shape processing method for human
vision, an instance of vertical transform sequences.
This “essay” is also the beginning of a survey in the context of visual analysis
with the goal of addressing the very nature of shape understanding. Most topics
surveyed in this manuscript would benefit from a more detailed description, with
illustrative examples. Also, some topics have only been touched upon: in particu-
lar the last topic of shape grammars, and the genesis of shape, which would benefit
from a comparative study of the different approaches now available in computing,
such as “shock grammars” (Siddiqi & Kimia 1996; Sebastian et al. 2004). The lat-
ter can be seen as recent developments in computing inspired from the original
approach to morphogenesis developed by the mathematician René Thom in the
1970’s, the founder of catastrophe theory (Thom 1989), which lead to the emer-
gence of singularity theory and its applications (Arnold 1989; Bruce & Giblin 1992;
Poston & Stewart 1996).

Notes

. One can also talk of the shape of a region of space, even if empty (φ = 0), but defined by
some outline: e.g., the region of space in between objects such as the voids between galaxies in
the universe (Weygaert 2003; Jones et al. 2005).
. Leyton talks about “transfer structure” rather than “structural field,” notions I consider
equivalent for the purpose of my somewhat imprecise study.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.39 (2457-2533)

Thoughts on shape 

. Refer to Leyton for an explanation of how memory and shape can be tied up together (Leyton
2001).
. The Hough transform maps edgels of a binary image to a parameter space. In the case of a
square one would use the slope and intercept of the equation of a straight line as the parameters.
Edgels which fall on nearby line segments will cluster in the parameter space.
. CAD-CAM: Computer Aided Design and Manufacturing.
. This is similar to the Hough Transform where lines in a binary image are mapped to points in
the parameter space.
. A “bias” in favour of invertible transforms for shape may have emerged in the signal process-
ing field, whose original concerns about (lossless) data transmission are very relevant, e.g., for
communication systems.
. The structural element is typically defined as a set of pixels or voxels, similarly to the input
dataset (an image array). Set operations, in particular Minkowski’s addition and subtraction,
are then used to construct higher order operators.
. E.g., this rules out the well-known Marr–Hildreth detector of zero–crossings of the Laplacian
of an image I, which creates “shadow edges,” i.e., edges of edges, if applied twice.
. Unless we change the constraint set or perturb the data set, of course – which is what one
may try to do to reach a global minimum.
. Adapted boundaries: hyper–surfaces on which one of the coordinate is constant.
. This is a restriction of a conformal map (i.e., which preserves local angles) which in the
literature is applied as a local constraint to more general functions, such as in complex analysis.
. A real square matrix A is positive definite if vT Av > 0. This is a useful property to solve
a linear system of equations, Ax = y by a (Cholesky) decomposition in a (upper) triangular
matrix.
. For more details refer to Arnold’s (Arnold 1989: Ch. 8) and Guillemin’s (Guillemin & Stern-
berg 1984) manuscripts.
. For a good introduction on this topic, refer to Gurtin’s manuscript (Gurtin 1982).
. The original statement gives a pair of solutions in general, only one of which is of practical
interest: the envelope at an increasing distance from the original sources of propagation.
. In differential geometry, ν is called the gauge figure or indicatrix: i.e., the set of (possible)
velocity vectors in every direction at q ∈ E3 (Koenderink 1990). In general, ν can vary in space,
which proves necessary to model propagation at interfaces or in non-homogeneous medium.
In the domain of visual perception, a varying index could be used to model contextual prop-
agation to favour certain orientations versus others (like horizontal versus vertical proximity
relationships between boundary elements).
. See also the work of Kotelly on Blum’s transform (Kotelly 1963). For a more recent applica-
tion of the distance surface τ = Φ(x, y) to the 2D skeletonization problem, when simulating a
grassfire on a square lattice, see the work of Leymarie and Levine (Leymarie & Levine 1992b).
. Another way of thinking of this concept, is in term of the system of normals to smooth
surfaces in Euclidean space. In neighbourhoods of such smooth surfaces, their normals (rays)
form a smooth fibration – they constitute tightly packed fibres (beams) filling-in space, hence
creating a regular distance hypersurface – but, at some distance from the surface, the various
normals begin to intersect one another (Arnold 1989: 480).
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.40 (2533-2604)

 Frederic Fol Leymarie

   
. A real function f : I → R such that ∃ a constant K ≥ 0, such that f (x) – f (y) ≤ K x – y
for all x, y in the interval I, is called a Lipschitz continuous function: i.e., it is limited in how fast
it can change.
. Historical note: This generalisation of the concept of Euclidean geometry was proposed by
Bernhard Riemann in 1854 in his lecture to qualify for the post of lecturer in Gauss’s department
at the University of Göttingen: “Über die Hypothesen welche der Geometrie zu Grunde liegen”
(Lord & Wilson 1984).
. For a more recent account on the field of morphometrics, refer to the works of Fred Book-
stein et al. (1992, 1997).
. Two (mathematical) objects are said to be homotopic if one can be continuously deformed
into the other (after Henri Poincaré, circa 1900 (Collins 2004)).
. For the “classic” application to the computer vision problem of the recovery of the trace of
curve points in an image, refer to the work of Parent & Zucker (1989).
. This work can be seen as an extension to surfaces of the ideas of Parent and Zucker developed
for curves (Parent & Zucker 1989).
. A “frame bundle” on a manifold M is a set of orthonormal (coordinate) frames defined at
every point p ∈ M: it determines the possible choices for an orthonormal basis of the tangent
space of M, TM ; it also reflects the ambiguity of choosing coordinates on M. The choice of a
(local) cross section of the frame bundle gives a “moving frame” used to calculate “tensors”
such as curvatures.
. The first fundamental form defines the inner product on the tangent space (of pairs of tan-
gent vectors): IS (v, u) = v · u. The second fundamental form defines a symmetric billinear form
on the tangent space: II S (v, u) = (v) · u, where  is the “shape operator” giving the (nega-
tive) derivative of the unit normal vector field of a surface (a kind of curvature). The ratio II S /IS
computes normal curvature to the surface.
. A frame on a 2-D manifold M is an orthonormal coordinate system, such that two of its
axes span the tangent space TM at any given point p ∈ M. Note that at each point of M, ∃ an
infinite number of such frames (an equivalence class) through rotation around the normal third
axis. Hence the need for the notion of bundle of frames.
. “An ideal Lambertian surface is one that appears equally bright from all viewing directions
and reflects all incident light, absorbing none” (Horn 1986: 212). The luminance of a Lambertian
surface depends only on the angle between the surface normal and the incident light ray.
. For I(x, y) = Imax one can set the height z = zmax (i.e., either the brightest or darkest spot in
a image) to some arbitrary constant.
. Refer to Berthold Horn’s manuscript for more details on the methods of “characteristic
strips” and “variational relaxation” for solving this classical shape from shading problem (Horn
1986: Ch. 11).
. This is an example of the Jacobi identity for Lie transforms (§3.1.5).
. For work along these lines, where a geometrical image transformation is put in corre-
spondence with the physical motion of the imaged objects, see e.g., the work of Fitzpatrick
(1988).
. Historical note: Gibson et al. coined the term “optical flow;” e.g., refer to (Gibson et al.
1959) for some of their early ideas on the subject.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.41 (2604-2692)

Thoughts on shape 

. A n–jet is a Taylor series expansion of which one keeps only the terms up–to (and including)
nth –order derivatives.
. This is an extension to the optical flow problem of the ideas on “shading flows and scenel
bundles” of Zucker et al. (Breton et al. 1992).
. I.e., it resembles a stick figure or skeleton for an anthropomorphic form.
. Using the “closure” requirement in the definition insures to capture limit points as ends of
SS branches. For example, in computing the MA of an ellipse, this ensures the two focal points
are included as end points.
. The result can as well be “stated by using maximal geodesic circles, [making it a closer
cousin] to the Euclidean maximal circle concept” (private communication from F.-E. Wolter).
. The origins of generalised cylinders can be traced back at least to the 18th century geome-
ter Gaspard Monge who defines “pipe” surfaces as the envelope of the set of spheres of radius
r centred on a space curve defining the axis or “spine” of the pipe (Monge 1850). Pipe sur-
faces are used in CAD-CAM for shape reconstruction, blending of surfaces, surface transitions,
numerically-controlled (NC) verification (Patrikalakis & Maekawa 2002: 353).
. A CW-complex, i.e., a “closure finite weak topology cell complex,” is a finite collection of
non-empty regions called “cells” homeomorphic to a closed ball, such that (i) the cells have
pairwise disjoint interiors, and (ii) the boundary of each cell is the union of other cells (of
lower dimensions) (Edelsbrunner & Shah 1997). CW-complices, a key tool in algebraic topology
(Bredon 1993), are a “nice” class of topological spaces which are general in scope and simple in
construction.
. Critical points of a function f on a manifold M are found where the gradient vanishes, i.e.,
∇f = 0. Such a function is said to be “Morse” if its critical point are non-degenerate, i.e., they
are isolated.
. Such “molecular” shape descriptions have recently started being used in Computer Graph-
ics, for modelling objects via meta-balls, that is, networks of spheres of varying radii, connected
into a 3D graph representation (Hart 1999).
. This was first proved only in the case of regular, simply connected, planar curve. It was then
extended to deal with the more interesting case of first order discontinuities, such as corners,
creases and cusps, in Hayes & Leyton (1989).
. The 3D version of this theorem states that “At each curvature extremum on a principle line of
curvature, there is a unique symmetry sheet that terminates at that extremum” (Yuille & Leyton
1990). Technically speaking, this argument is always valid (for 2D and 3D problems) when con-
sidering the symmetry set (SS) (Bruce & Giblin 1992; Giblin & Kimia 2004). When considering
only the MA (subset of the SS) some axes may be missing in correspondence to surface ridges,
i.e., lines of curvature extrema of the boundary. Recall that the MA sheets terminate at centres
of curvature associated to ridges, rather than on the bounding surface itself.
. Here information is identified with variety. Refer to Howard Resnikoff who has shown that
the measurement of curvature variation has the same form as the measurement of information in
information theory (Resnikoff 1989).
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.42 (2692-2819)

 Frederic Fol Leymarie

References

Arnheim, R. (1974). Art and Visual Perception: A Psychology of the Creative Eye. Berkeley, CA:
University of California Press. New version; expanded and revised edition of the 1954 origi-
nal.
Arnold, V. I. (1989). Mathematical Methods of Classical Mechanics, volume 60 of Graduate Texts
in Mathematics. Springer-Verlag, 2nd edition.
Bader, R. F. W. (1990). Atoms in Molecules – A Quantum Theory, volume 22 of International
Series of Monographs on Chemistry. Oxford, UK: Clarendon Press.
Bamberg, P., & Sternberg, S. (1988). A Course in Mathematics for Students of Physics, volume 1.
Cambridge University Press.
Bentley, P. J. (Ed). (1999). Evolutionary Design by Computers. Morgan Kaufmann.
Bergevin, R., & Levine, M. D. (1993). Generic object recognition: Building and matching coarse
descriptions from line drawings. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence, 15, (1), 19–36.
Biederman, I. (1987). Recognition-by-components: A theory of human image understanding.
Psychological Review, 94, 115–147.
Biederman, I. (1995). Visual object recognition. In S. F. Kosslyn & D. N. Osherson (Eds.), An
Invitation to Cognitive Science, volume 2 of Visual Cognition, Chapter 4 (pp. 121–165). MIT
Press.
Blank, A. A. (1978). Metric geometry in human binocular perception: Theory and fact. In E. L.
J. Leeuwenberg & H. F. J. M. Buffart (Eds.), Formal Theories of Visual Perception, Chapter 4
(pp. 83–102). John Wiley and Sons.
Blum, H. (1967). A transformation for extracting new descriptors of shape. In W. Wathen-Dunn
(Ed.), Models for the Perception of Speech and Visual Form (pp. 362–380). Cambridge, MA:
MIT Press. Proceedings of a Symposium held in Boston, MA, November 1964.
Blum, H. (1973). Biological shape and visual science. Journal of Theoretical Biology, 38, 205–287.
Bookstein, F. L. (1992). Morphometric Tools for Landmark Data: Geometry and Biology. Cam-
bridge University Press.
Bookstein, F. L. (1997). Landmark methods for forms without landmarks: Morphometrics of
group differences in outline shape. Medical Image Analysis, 1(3), 225–243.
Borgefors, G. et al. (1999). Computing skeletons in 3D. Pattern Recognition, 32(7), 1225–1236.
Brady, M., & Asada, H. (1984). Smoothed local symmetries and their implementation. Interna-
tional Journal of Robotic Research, 3(3), 36–61.
Bredon, G. E. (1993). Topology and Geometry, Volume 139 of Graduate Texts in Mathematics.
Springer-Verlag.
Breton, P, Iverson, L., Langer, M., & Zucker, S. W. (1992). Shading flows and scenel bundles: A
new approach to shape from shading. In G. Sandini (Ed.), Proceedings of the 2nd European
Conference on Computer Vision (ECCV), Volume 588 of Lecture Notes in Computer Science
(pp. 135–150). Santa Margherita, Ligure, Italy. Springer-Verlag.
Bruce, J. W., & Giblin, P. J. (1992). Curves and Singularities: A Geometrical Introduction to
Singularity Theory, 2nd edition. Cambridge University Press, UK.
Burt, P. J. (1981). Fast filter transforrns for image processing. Computer Graphics and Image
Processing (CGIP), 16, 20–51.
Burt, P. J. (1983). Fast algorithms for estimating local image properties. Computer Vision, Graph-
ics and Image Processing (CVGIP), 21, 368–382.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.43 (2819-2958)

Thoughts on shape 

Catté, F., Lions, P.-L., Morel, J.-M., & Coll, T. (1992). Image selective smoothing and edge
detection by nonlinear diffusion. SIAM Journal on Numerical Analysis, 29(1), 182–193.
Collins, G. P. (2004). The shapes of space. Scientific American, 291, 94–103.
Danielsson, P.-E. (1980). Euclidean distance mapping. Computer Graphics and Image Processing
(CGIP), 14, 227–248.
Dawkins, R. (1986). The Blind Watchmaker. W. W. Norton and Company. Reissued in 1990 (UK)
and 1996 (USA).
Deutsch, A., & Dormann, S. (2005). Cellular Automaton Modeling of Biological Pattern Forma-
tion. Modeling and Simulation in Science, Engineering and Technology. Birkhäuser.
Dickinson, S. J., Bergevin, R., Biederman, L., Eklundh, J.-O., Munck-Fairwood, R., Jain, A. K., &
Pentland, A. (1997). Panel report: The potential of geons for generic 3-D object recognition.
Image and Vision Computing, 15(4), 277–292.
Ebert, D. S., Musgrave, F. K., Peachey, D., Perlin, K., & Worley, S. (2002). Texturing and Model-
ing – A Procedural Approach. Morgan Kaufmann, 3rd edition.
Edelsbrunner, H., & Shah, N. R. (1997). Triangulating topological spaces. International Journal
of Computational Geometry and Applications, 7(4), 365–378.
Ferraro, M., & Caelli, T. M. (1988). Relationships between integral transform invariances and
Lie group theory. Journal of the Optical Society of America – A, 5(5), 738–742.
Ferraro, P., Godin, C., & Prusinkiewicz, P. (2005). Toward a quantification of self-similarity in
plants. Fractals, 13(2), 91–109.
Fitzpatrick, J. M. (1988). The existence of geometrical density image transformations corre-
sponding to object motion. Computer Vision, Graphics and Image Processing (CVGIP), 44(2),
155–174.
Giblin, P. J., & Kimia, B. B. (2004). A formal classification of 3D medial axis points and their
local geometry. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 26(2),
238–251.
Gibson, E. J., Gibson, J. J., Smith, O. W., & Flock, H. (1959). Motion parallax as a determinant
of perceived depth. Journal of Experimental Psychology, 8(1), 40–51.
Gibson, J. J. (1950). The Perception of the Visual World. Boston, MA: Houghton Mifflin Cie.
Reprinted in 1974 by Greenwood Press, Westport, CT.
Goodman, A. W. (1964). A partial differential equation and parallel plane curves. American
Mathematical Monthly, 71, 257–264.
Guillemin, V., & Sternberg, S. (1984). Symplectic Techniques in Physics. Cambridge University
Press, UK.
Gurtin, M. E. (1982). An Introduction to Continuum Mechanics, volume 158 of Mathematics in
Science and Engineering Series. Harcourt Publishers.
Hansen, O. (1992). On the Use of Local Symmetries in Image Analysis and Computer Vision.
PhD thesis, Laboratory of Image Analysis, Institute of Electronic Systems, Aalborg Univer-
sity, Denmark.
Hart, J. C. (1999). Using the CW-complex to represent the topological structure of im-
plicit surfaces and solids. In Proc. Implicit Surfaces ’99 (pp. 107–112). Eurograph-
ics/SIGGRAPH, ACM.
Hayes, P J., & Leyton, M. (1989). Processes at discontinuites. In Proceedings of the 11th Joint
Conference on Artificial Intelligence, volume 2 (pp. 1267–1272). Detroit, MI: American Asso-
ciation for A.I.
Heeger, D. J., Simoncelli, E. P., & Movshon, J. A. (1996). Computational models of cortical vi-
sual pracessing. In Proceedings of the National Academy of Science, volume 93 (pp. 623–627).
Washington, DC.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.44 (2958-3082)

 Frederic Fol Leymarie

Hilaga, M., Shinagawa, Y , Kohmura, T., & Kunii, T. L. (2001). Topology matching for fully
automatic similarity estimation of 3D shapes. In Proc. SIGGRAPH 2001 (pp. 203–212). Los
Angeles, CA. ACM.
Hoffman, W. C. (1966). The Lie algebra of visual perception. Journal of Mathematical Psychology,
3, 65–98. Errata in vol. 4, pp. 348–349, 1967.
Hoffman, W. C. (1978). The Lie transformation group approach to visual neuropsychology. In
E. L. J. Leeuwenberg & H. F. J. Buffart M. (Eds.), Formal Theories of Visual Perception, Chapter
2 (pp. 27–66). John Wiley and Sons.
Hoffman, W. C. (1984). Figural synthesis by vectorfields: Geometric neuropsychology. In P. C.
Dodwell & T. Caelli (Eds.), Figural Synthesis, Chapter 8 (pp. 249–282). Lawrence Erlbaum
Associates.
Horn, B. K. P. (1986). Robot Vision. Electrical Engineering and Computer Science Series. Cam-
bridge, MA: MIT Press.
Hummel, J. E., & Biederman, I. (1992). Dynamic binding in a neural network for shape recog-
nition. Psychological Review, 99, 480–517.
Johnson, C. K. (1999). Crystallographic topology 2: Overview and work in progress. In V. Alex-
iades& G. Siopsis (Eds.), Trends in Mathematical Physics, Studies in Advanced Mathematics.
AMS, Providence, RI.
Johnson, C. K., Burnett, M. N., & Dunbar, W. D. (1996). Crystallographic topology and
its applications. Oak Ridge National Laboratory. Preprint available: http://www.ornl.gov/
sci/ortep/topology/preprint.html
Jones, B. J. T., Martínez, V. J., Saar, E., & Trimble, V. (2005). Scaling laws in the distribution of
galaxies. Reviews of Modern Physics, 76, 1211–1267.
Jones, D. G., & Malik, J. (1992). Determining three dimensional shape from orientation and
spatial frequency disparities. In G. Sandini (Ed.), Proceedings of the 2nd European Conference
on Computer Vision (ECCV), volume 588 of Lecture Notes in Computer Science (pp. 661–669).
Santa Margherita, Ligure, Italy. Springer-Verlag.
Kass, M., Witkin, A., & Terzopoulos, D. (1988). Snakes: Active contour models. International
Journal of Computer Vision (IJCV), 1(4), 321–331.
Kelly, M. F., & Levine, M. D. (1995). Annular symmetry operators: A method for locating and de-
scribing objects. In Proceedings of the 5th International Conference on Computer Vision (ICCV)
(pp. 1016–1021).
Kimia, B. B. (1991). Toward a Computational Theory of Shape. PhD thesis, Electrical Engineering,
McGill University, Montreal, Canada.
Kimia, B. B. (2003). On the role of medial geometry in human vision. Journal of Physiology,
97(2–3), 155–190.
Kimia, B. B., Tannenbaum, A. R., & Zucker, S. W. (1995). Shapes, shocks, and deformations I:
The components of two-dimensional shape and the reaction-diffusion space. International
Journal of Computer Vision, 15(3), 189–224.
Klein, F. (1893). Vergleichende Betrachtungen über neuere Geometrische Forschungen (“Erlan-
gen Programm”). Mathematische Annalen, 43, 63–100.
Koenderink, J. J. (1990). Solid Shape. A.I. Series. Cambridge, MA: MIT Press.
Koenderink, J. J., & van Doorn, A. J. (1976). Geometry of binocular vision and a model for
stereopsis. Biological Cybernetics, 21(1), 29–35.
Kotelly, J. C. (1963). A mathematical model of Blum’s theory of pattern recognition. Research
report AFCRL-63-164, U.S. Air Force Cambridge Research Laboratories.
Kovács, L., Feher, A., & Julesz, B. (1998). Medial-point description of shape: A representation
for action coding and its psychological correlates. Vision Research, 38, 2323–2333.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.45 (3082-3203)

Thoughts on shape 

Kumar, S., & Bentley, P. J. (Eds). (2003). On Growth, Form and Computers. Academic Press.
Kunze, R., Wolter, F.-E., & Rausch, T. (1997). Geodesic Voronoi diagrams on parametric sur-
faces. In CGI ’97 (pp. 230–237). IEEE, Computer Society Press.
Lagarde, J. W. (1990). Constraints and their satisfaction in the recovery of local surface structure.
Master’s thesis, Electrical Engineering, McGill University, Montreal, Canada.
Langer, M. S., & Zucker, S. W. (1994). Shape-from-shading on a cloudy day. Journal of the Optical
Society of America – A, 11(2), 467–478.
Leymarie, F. (1990). Tracking and describing deformable objects using active contour mod-
els. Technical report CIM-90-9, McGill University, Dept. of Electrical Engineering, Montreal,
Canada. www.lems.brown.edu/∼leymarie/meng/.
Leymarie, F., & Levine, M. D. (1989). Shape features using curvature morphology. In W. A.
Pearlman (Ed.), Proc. of the SPIE Conf. on Visual Communications and Image Processing IV,
volume 1199 (pp. 390–401).
Leymarie, F., & Levine, M. D. (1992a). Fast raster scan distance propagation on the discrete
rectangular lattice. Computer Vision, Graphics and Image Processing – Image Understanding
(CVGIP-IU), 55(1), 84–94.
Leymarie, F., & Levine, M. D. (1992b). Simulating the grassfire transform using an active con-
tour model. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 14(1),
56–75.
Leymarie, F., & Levine, M. D. (1993). Tracking deformable objects in the plane using an active
contour model. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 15(6),
617–634.
Leymarie, F. F. (2003). Three-Dimensional Shape Representation via Shock Flows. PhD thesis,
Brown University, Providence, RI, USA. www.lems.brown.edu/∼leymarie/phd/.
Leymarie, F. F., & Kimia, B. B. (2000). Discrete 3D wave propagation for computing morpho-
logical operations from surface patches and unorganized points. In J. Goutsias, L. Vincent, &
D. Bloomberg (Eds.), Math. Morphology and its Applications to Image and Signal Processing,
volume 18 of Computational Imaging and Vision Series (pp. 351–360). Kluwer Academic.
Leymarie, F. F., & Kimia, B. B. (2001). The shock scaffold for representing 3D shape. In C. Arcelli,
L. P. Cordella, G. S. d. B. (Eds.), Visual Form, number LNCS 2059 (pp. 216–229). Springer-
Verlag.
Leymarie, F. F., & Kimia, B. B. (2007). The medial scaffold of 3D unorganized point clouds. IEEE
Transactions on Pattern Analysis and Machine Intelligence. Accepted for publication.
Leyton, M. (1984). Perceptual organization as nested control. Biological Cybernetics, 51, 141–
153.
Leyton, M. (1985). Generative systems of analyzers. Computer Vision, Graphics and Image Pro-
cessing (CVGIP), 31(2), 201–241.
Leyton, M. (1987). Symmetry-curvature duality. Computer Vision, Graphics and Image Process-
ing (CVGIP), 38, 327–341.
Leyton, M. (1988). A process grammar for shape. Artificial Intelligence Journal, 34(2), 213–247.
Leyton, M. (1989). Inferring causal history from shape. Cognitive Science, 13, 357–387.
Leyton, M. (2001). A Generative Theory of Shape. Number LNCS 2145. Springer-Verlag.
Lindenmayer, A. (1968). Mathematical models for cellular interactions in development: Parts I
and II. Journal of Theoretical Biology, 18, 280–315.
Lord, E. A., & Wilson, C. B. (1984). The Mathematical Description of Shape and Form. Ellis
Horwood Series in Mathematics and its Applications. Statistics and Operational Research.
Halsted Press.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.46 (3203-3346)

 Frederic Fol Leymarie

Luneburg, R. K. (1947). Mathematical Analysis of Binocular Vision. Princeton, NJ: Princeton

University Press. Published for the Hanover (Dartmouth Eye) Institute, N.H., U.S.A.
Luneburg, R. K. (1950). The metric of binocular visual space. Journal of the Optical Society of
America (JOSA), 40(10), 627–642.
Mandelbrot, B. (1982). The Fractal Geometry of Nature. W. H. Freeman & Co.
Milnor, J. W. (1963). Morse theory. In Annals of Mathematics Studies, volume 51. Princeton, NJ:
Princeton University Press.
Monge, G. (1850). Application de l’Analyse à la Géométrie. Bachelier, Paris, France.
Montanari, U. (1969). Continuous skeletons from digitized images. Journal of the Association for
Computing Machinery, 16(4), 534–549.
Murray, J. D. (2004). Mathematical Biology, 3rd edition, in 2 volumes. Springer.
Norman, J. F., & Todd, J. T. (1992). The visual perception of 3-dimensional form. In G. A. Car-
penter & S. Grossberg (Eds.), Neural Networks for Vision and Image Processing, Chapter 4 (pp.
93–110). Cambridge, MA: MIT Press.
Okabe, A., Boots, B., Sugihara, K., & Chiu, S. N. (2000). Spatial tessellations: Concepts and
applications of Voronoi diagrams. Probability and Statistics series, 2nd edition. Wiley.
Owens, R., Venkatesh, S., & Ross, J. (1989). Edge detection is a projection. Pattern Recognition
Letters, 9(4), 233–244.
Papathomas, T. V., & Julesz, B. (1989). Lie differential operators in animal and machine vision.
In J. C. Simon (Ed.), From Pixels to Features (pp. 115–126). North-Holland. Proceedings of a
Workshop held at Bonas, France, 22–27 August 1988.
Parent, P., & Zucker, S. W. (1989). Trace inference, curvature consistency and curve detection.
IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 11(8), 823–839.
Patrikalakis, N. M., & Maekawa, T. (2002). Shape Interrogation for Computer Aided Design and
Manufacturing. Springer.
Pentland, A., & Scarloff, S. (1991). Closed-form solutions for physically based shape model-
ing and recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI),
13(7), 715–729.
Perlick, V. (2004). Gravitational lensing from a spacetime perspective. Living Reviews in Relativ-
ity, 7(9). http://www.livingreviews.org/lrr-2004-9
Pogorelov, A. V. (1958). Differential Geometry. P. Noordhoff, Groningen, The Netherlands.
Poston, T., & Stewart, I. (1996). Catastrophe Theory and Its Applications, new edition. Dover.
Prusinkiewicz, P., & Lindenmayer, A. (1999). The Algorithmic Beauty of Plants, 2nd edition.
Springer-Verlag.
Rausch, T., Wolter, R.-E., & Sniehotta, O. (1996). Computation of medial curves in surfaces. In
Conference on the Mathematics of Surfaces VII (pp. 43–68). Institute of Mathematics and its
Applications (IMA).
Resnikoff, H. L. (1989). The Illusion of Reality: Topics in Information Science. Springer-Verlag.
Richards, W., & Hoffman, D. D. (1985). Codon constraints on closed 2D shapes. Computer
Vision, Graphics and Image Processing (CVGIP), 31, 265–281.
Ronse, C. (1993). On idempotence and related requirements in edge detection. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence (PAMI), 15(5), 484–491.
Sander, P. T. (1988). On Reliably Inferring Differential Structure from Three-Dimensional Images.
PhD thesis, Electrical Engineering, McGill University, Montreal, Canada.
Sander, P. T., & Zucker, S. W. (1990). Inferring surface trace and differential structure from
3-D images. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 12(9),
833–854.
JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.47 (3346-3499)

Thoughts on shape 

Sebastian, T. B., Klein, P. N., & Kimia, B. B. (2004). Recognition of shapes by editing their shock
graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(5), 550–571.
Serra, J. (1982). Image Analysis and Mathematical Morphology, volume 1. Academic Press.
Sethian, J. A. (1999). Level Set Methods and Fast Marching Methods: Evolving Interfaces in Com-
putational Geometry, Fluid Mechanics, Computer Vision, and Materials Science. Cambridge
Monograph on Applied and Computational Mathematics, 2nd edition. Cambridge Univer-
sity Press, UK.
Siddiqi, K., & Kimia, B. (1995). Parts of visual form: Computational aspects. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 17(3), 239–251.
Siddiqi, K., & Kimia, B. B. (1996). A shock grammar for recognition. In Proceedings of the
Conference on Computer Vision and Pattern Recognition (CVPR’96), (pp. 507–513).
Siddiqi, K., Kimia, B. B., Tannenbaum, A., & Zucker, S. W. (2001). On the psychophysics of the
shape triangle. Vision Research, 41(9), 1153–1178.
Siddiqi, K., Kimia, B. B., Tannenbaum, A., & Zucker, S. W. (1999a). Shocks, shapes, and wiggles.
Image and Vision Computing, 17(5–6), 365–373.
Siddiqi, K., Shokoufandeh, A., Dickinson, S. J., & Zucker, S. W. (1999b). Shock graphs and shape
matching. International Journal of Computer Vision, 35(1), 13–32.
Sing, M., & Hoffman, D. D. (2001). Part-based representations of visual shape and implica-
tions for visual cognition. In T. F. Shipley & P. J. Kellman (Eds.), From Fragments to Objects:
Segmentation and Grouping in Vision, Chapter 9 (pp. 401–459). Elsevier Science.
Soille, P. (2004). Morphological Image Analysis, corrected 2nd edition. Springer.
Stavroudis, O. N. (1972). The Optics of Rays, Wavefronts, and Caustics, volume 38 of Pure and
Applied Physics. Academic Press.
Sundar, H., Silver, D., Gagvani, N., & Dickinson, S. J. (2003). Skeleton based shape matching
and retrieval. In Proc. of IEEE International Conference on Shape Modeling and Applications
(pp. 130–142). Seoul, Korea.
Tek, H., & Kimia, B. B. (2001). Boundary smoothing via symmetry transforrns. Journal of
Mathematical Imaging and Vision, 14(3), 211–223.
Terzopoulos, D., Witkin, A., & Kass, M. (1987). Symmetry-seeking models for 3D object recog-
nition. International Journal of Computer Vision (IJCV), 1(3), 211–221.
Thom, R. (1989). Structural Stability and Morphogenesis. Advanced book classics. Addison Wes-
ley.
Thompson, D. W. (1992). On Growth and Form. Dover Publications, complete revised edition
of the 1917 original edition.
Todd, J. T., & Reichel, F. D. (1989). Ordinal structure in the visual perception and cognition of
smoothly curved surfaces. Psychological Review, 96(4), 643–657.
Todd, S., & Latham, W. (1992). Evolutionary Art and Computers. Academic Press.
Turing, A. (1952). The chemical basis of morphogenesis. Philosophical Transactions of the Royal
Society of London, 237(641), 37–72.
van Tonder, G. J., & Ejima, Y. (2000). The patchwork engine: Image segmentation from image
symmetries. Neural Networks, 13(3), 291–303.
van Tonder, G. J., & Ejima, Y. (2003). Flexible computation of shape symmetries within the
maximal disk paradigm. IEEE Transactions on Systems, Man, and Cybernetics, Part B, 33(3),
535–540.
VanRullen, R., & Koch, C. (2003). Is perception discrete or continuous? Trends in Cognitive
Sciences, 7(5), 207–213.
Verroust, A., & Lazarus, F. (2000). Extracting skeletal curves from 3D scattered data. The Visual
Computer, 16(1), 15–25.
 JB[v.20020404] Prn:23/10/2006; 9:07 F: AICR6714.tex / p.48 (3499-3558)

 Frederic Fol Leymarie

Wade, L., & Parent, R. E. (2002). Automated generation of control skeletons for use in anima-
tion. The Visual Computer, 18(2), 97–110.
Weygaert, R. v. d. (2003). The cosmic foam: Stochastic geometry and spatial clustering across
the universe. In E. Feigelson & G. Babu (Eds.), Statistical Challenges in Modern Astronomy III
(pp. 175–196). Springer-Verlag.
Wolfram, S. (1994). Cellular Automata and Complexity: Collected Papers. Addison-Wesley.
Wolter, F.-E. (1985). Cut Loci in Bordered and Unbordered Riemannian Manifolds. PhD thesis,
Technical University of Berlin, Department of Mathematics, Germany.
Wolter, F.-E. (1992). Cut locus and medial axis in global shape interrogation and representation.
Sea grant report, MIT, Cambridge, MA.
Wolter, F.-E., & Friese, K.-I. (2000). Local and global geometric methods for analysis interroga-
tion, reconstruction, modification and design of shape. In Proceedings of Computer Graphics
International (CGI’00) (pp. 137–151). Geneva, Switzerland. IEEE Computer Society.
Wolter, F.-E., Peinecke, N., & Reuter, M. (2004). Geometric Modeling of Complex Shapes and En-
gineering Artifacts, volume 1 of Encyclopedia of Computational Mechanics, Chapter 16. Wiley.
Wu, K., & Levine, M. D. (1997). 3-D shape approximation using parametric geons. Image Vision
and Computing, 15(2), 143–158.
Yuille, A., & Leyton, M. (1990). 3-D symmetry-curvature duality theorems. Computer Vision,
Graphics and Image Processing (CVGIP), 52, 124–140.
Zucker, S. W. (1987). The emerging paradigm of computational vision. Annual Review of Com-
puter Science, 2, 69–89.

View publication stats