Вы находитесь на странице: 1из 33

Singularities and Geometry with

Applications to Imaging
Peter Giblin (University of Liverpool)

Valencia, July 2012, Continued......

THE book on medial


axes and related
matters!

Finally, some applications of


singularities of mappings, that is
where the dimension of the target
is > 1. This brings in ongoing work
with Jim Damon (but mostly hell
talk about that).

()

(())

a fold singularity

Apparent contours of surfaces have been used


extensively in recovery of surface shape and camera
motion from views.

Early work (1985-6) on this was done by Giblin and


Weiss but the major figures in more recent years
include Roberto Cipolla (U of Cambridge).
There is a book Visual Motion of
Curves and Surfaces (Cambridge 2000)
by Cipolla and Giblin explaining the
geometry and applications.
This involves global considerations
whereas I shall stick to local ones here.

Weve seen this


before when
studying the
envelope of
normals to a
parabola!

Whitney showed that the fold and the cusp are the only two
stable mappings from the plane to the plane.

There is a nice way of thinking about apparent contours


which goes back to the description of discriminants in the
first two lectures.

Of course the canonical example of this is the surface of


normals to a parabola and the vertical view of it.

But the point here is that we can write down the


conditions for the projection to be a fold or a cusp using
the versal unfolding conditions for discriminants
(similar to those for bifurcation sets).

Another example: spreading out in the t-direction


the family of circles of fixed radius centred on a
parabola. The envelope of the circles in the plane
is the apparent contour of the surface.

Note that each of these is four equations to be satisfied by


three unknowns t, x, y. So maybe theres no solution?
If we can show G avoids all these bad sets then every
apparent contour of G=0 will be locally smooth, fold or cusp.

There is a technique known as transversality which is used


to show that things which we dont want to happen almost
never do happen.

(Unfortunately this doesnt seem to work in real life.)


Two manifolds in a euclidean space are called transversal at a
point p if their tangent spaces at p span the whole ambient space.

There is then a rather neat argument


given in the book Curves and
Singularities which shows that, when
the surface G=0 is given by a
polynomial equation of degree some
definite d then, for almost all such
polynomials G (all but a set of measure
zero) the three nasty equations above
have no solutions.

The essential ingredient is a transversality theorem of Thom,


which in turn depends on a key result called Sards lemma.
In the present context this says the following.

So almost all surfaces (at least in this algebraic context)


have no singularities other than folds and cusps in their
apparent contours.
This is a typical genericity result and transversality is used
here to prove that equations have no solutions in general.

We can also describe folds and cusps of mappings from 2


variables to 2 variables by means of their critical sets.

A beaks
transition
on
apparent
contours

Of course there are other higher singularities of


such mappings from the plane to the plane. There
are extensive classifications, due to J.Rieger and
D.Mond among others.

Some of these higher singularities will appear later in


Jims lectures, but Ill mention an application where
there are two projection maps, one coming from an
illumination and one from a view projection. This
is joint work with Jim Damon and Gareth Haslinger
and I will only trespass very slightly on Jims
material now. So you may see a few of the next
pictures again shortly!

The setup: One (major) source of illumination, providing the


shade curves (gradual shading off) and cast shadows (sharp)
Illumination is stable,
from a point source far
away

Surfaces are piecewise smooth, that is two or


three surfaces can meet in a crease or corner
light
Cast shadow

light

cast
shadow
of
crease

The next couple of slides relate


directly to calculations I have done
in the previous lectures...

submersion y 0

fold y = 0,
called a shade
curve

The light projection being a cusp map is more interesting...

The mathematics doesnt


distinguish between the two parts
of the shade curve x = 3y2.
The real part casts a shadow on
the opposite wall of the surface

Light producing shade


curve x = 3y2.

Of course this can interact with an apparent contour when we


view the surface from a particular direction:
real shade curve
imaginary
shade curve
(real for
opposite light
direction)
apparent
contour
real cast
shadow

recall we are
only looking
locally

real shade curve


real cast
shadow

Real
Mathematical
shade+shadow
shade+shadow

Now some pictures and examples which


will be elaborated on by Jim in his
lectures.

Were interested in the way contours, markings, creases etc change


as viewpoint changes:

Consider the simplest case of a smooth surface having a


shade curve, no cast shadow and viewed so that there is an
apparent contour.

the shade
curve
behaves,
maybe, like
a boundary

semifold

fold

The only possible snag is that the shade curve is not an


arbitrary smooth curve on M.

A useful geometrical observation


In the tangent plane at p these are conjugate
Light or view direction
M

tangent to
the shade
curve or
contour
generator

shade curve (light) or


contour generator
(view)

Conjugate:
with respect to
the second
fundamental
form

For p = origin, z = ax2 + 2bxy + cy2 + h.o.t., (v1, v2) and (w1,w2) are
conjugate if av1w1 + b(v1w2+v2w1) + cv2w2 = 0. Asymptotic
directions are self-conjugate.

For z = ax2 + 2bxy + cy2 + h.o.t. (v1, v2) and (w1,w2)


are conjugate iff av1w1 + b(v1w2+v2w1) + cv2w2 = 0.
Asymptotic directions are self-conjugate.
At a parabolic point (b=c=0) all directions are
conjugate to the unique asymptotic direction.

So could the shade curve and the contour generator on M ever be


tangent on M?
Say they were tangent to (w1, w2) = (1,0). Then the light and view
directions would both satisfy av1+bv2 = 0.

So assuming (always) that the light and view are distinct then
a = b = 0, so the origin p is parabolic and the common tangent to
shade and contour is the unique asymptotic direction at p.
This means that any cases in the Bruce-Giblin-Goryunov
classification which demand tangency require special geometry,
and may indeed be ruled out altogether.

Semibeaks. Here the


contour generator on the
surface and the shade curve
are tangent at a parabolic
point of the surface (notice
the inflection on the green
apparent contour).

You can see they are tangent by


tilting the surface slightly (and
leaving both curves in the same
place).
This can only occur at a
parabolic point.

The semibeaks is
Codimension 1 since
viewer movement along a
curve in space unfolds the
singularity of projection to
stable ones.

Unfolding of a semibeaks singularity with viewer movement. On the


by this
this transition.
transition.
right are two semifolds, created by

One of the cases which cannot arise with a shade curve is the
double cusp. It can occur on a surface with boundary where the
boundary curve and the apparent contour both project to a cusp in
the image.
The view here is along the common
tangent to the boundary (light blue) and
Surface with
the contour on M (green) at p.
boundary
p

Apparent contour
would be a cusp for
a transparent
surface; here its
half a cusp!

But this common tangent is then


asymptoticself-conjugateon M so if
the light blue curve were a shade curve
either
view and light directions coincide, or
p is a parabolic point, in which case we
do not see a cusp on the apparent
contour (we see an inflexion, an isolated
point or a crossing).

Another singularity (the semigoose) can


only occur at a cusp of Gauss
Surface with an ordinary
parabolic point: intersects its
tangent plane in a cusp
Surface with a (hyperbolic) cusp of
Gauss: intersects its tangent plane
in a tacnode.

z = ax2 + bxy2 + cy4 where b2 > 4ac


(+ some other assorted terms above
these on the Newton diagram)

The End
Thank you for
your attention

Вам также может понравиться