guess. We can distinguish interactive settings and automated
detection by other feature extraction or segmentation tools.
3.5.1 Interactive setting of seed points and
interaction. The most used form for deriving at the initial
state of a snake is the interactive setting of seed points. The
breakpoints of a contour have usually to be defined very
precisely. Grün and Li, 1994 emphasize the importance of a
convenient graphical user interface for that purpose.
Neuenschwander et al., 1995 use Ziplock snakes, given only
start and end point of well defined edge segments. This
method converges rather well, even with quite different
initial states,
Interaction during the optimization is possible by adding
suitable energy to push the snake out of a local minima to
the desired position. This can be very time consuming and
requires the presence of an operator during the whole
extraction process.
3.5.2 Automated initial state for open
contours. Snakes are applied for road extraction (e.g. Quam
and Strat, 1991), where a semi-automatic road tracker applied
to an aerial scene provides the initial road contours. Other
approaches use region growing. In Ruskoné et al., 1994 a
coarse network is provided by automated road detection and
road tracking.
3.5.3 Automated initial state for closed
contours. Closed snakes are used to extract buildings (e.g.
Quam and Strat, 1991). Approximate values can be provided
by region growing or scene partitioning methods with
global thresholds (Fua and Leclerc, 1988). In general seed
points or some reduced user interaction is required. This is
due to the fact, that constraints on the boundaries are not
easy to introduce for generic cases. In Gülch, 1990 the initial
contour is given by the boundaries of the user selected
largest regions which is optimized through a pyramidal
approach until the final result is reached.
The extraction of natural boundaries like vegetation
boundaries is possible (Quam and Strat, 1991), the modeling
is however more difficult and the snakes might require more
interactive seed points and attention. If the accuracy
requirements are lower a more relaxed modeling can be
applied. The snakes described in (Gülch, 1990, 1993) have
been used in satellite imagery to extract the boundaries of
logged forest stands. Rule based multi-spectral classification
and analysis identifies old map polygons with newly logged
forest (Johnsson, 1994). Those map polygons give the
initial state for the snake. The contour has been refined in 60
iterations in a hierarchical approach (fig. 4). The contour has
attracted to the major body (black) of the logged forest stand,
cutting out small details. The major remaining problems are
the rather coarse approximate values from the old forest
maps and the difficult topological relations that can occur if
the logged forest stands cover several old polygons in the
advantageous.
5%
Aa d el.
de
eod er.
N
3
NS
N
^N
GA ees:
P
i
ur,
de
ee
1,
v
e
a
"V.
e
I
"e^ d PP d.
s" s" ds
^
Fig. 4: Logged forest stand in satellite scene.
The initial state for the contour from an automatically identified
old map polygon, is shown together with the extracted boundary
of the logged forest stand (60 iterations).
3.6 Objects of known topology.
In many applications the topology of objects or even size
and shape are known. In those cases the snakes usually
perform better than in more generic cases.
In Ruskoné et al., 1994, the accuracy of an automatically
detected road network is improved by a network of snakes.
The topology is given and kept fixed in this elastic network
which is globally adjusted.
In Gülch, 1995 snakes refine the contour of a signalized
control point, extracted by an automatic rule based region
segmentation. It is assumed that the signalized point can be
described by a closed boundary and that the signal appears as
a homogeneous region in the image. Pixel size, the image
scale and the image patch are known. Also the true type,
shape, size, colour of the signals are supposed to be
available. Satisfying results can be obtained with fixed
parameter settings in rotated and distorted images (fig. 5).
In ultrasonic heart images an initial contour of a heart
chamber has been derived by a blackboard system, based on a
heart chamber model and a set of internal relations in a
relaxation labelling scheme. The quite coarse approximation
is sufficient for further refinement by an active contour
model, as can be seen in fig. 6.
Neuenschwander et al., 1995 assume a spherical topology for
3D deformable surfaces which require a small number of seed
points. The surface is initialized/optimized by automatic
selection of boundary conditions and front propagation.
ANNIS SN
EIN,
DIES SEIN
0 ET Sae
NS SS
e:
etre tete etant date te.
"te
SEEN
SEN
VE es
Grae
edd dud de
edd due,
T
rere
ose
oe
2
ON certe tete
A
SUNNY
NRA ANN,
NARA S
MS: MN S ee
= RR
S NN SS
SE NASE
%
Sr,
e
he
oe
Sr
EE
Ce
su
T
RRR,
s
SEO I
Cr
m
2m
A
T
73
oe
BA
dette
se
ecrit. os
RRR
a Nee
Ce
CR
Dm
és
EEE
deo th dad,
o
i ce
a signalized point (a-e — l-r) a) Image, b) region segmentation
with signal candidate (white region), c) initial, shape from approximated boundary, d) extracted contour of the signal, (40
iterations), e) the coarse position of the signal (circle) is derived as a weighted centre of gravity of the contour points.
282
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
map. In those case dividable snakes (cf. 3.7) would be
Fig. 6
Initi
aîte
3:7
If nott
objects
approa
techniq
Mallad
protrus
The co
indepe
front a
not lin
a stabl
objects
Deling
points
enable
require
which
Tonne:
arbitra
surfac«
elastic
al. |
triangi
global
than s
compu
alway
furthe
3.8
Any a
shoulc
Szelis
measu
descri
nodal
appro
distril
contol
energ
Gülch
mask
