CMBR-based, have been utilized to approximate buildings con-
taining rectangular and non-rectangular outlies.
Figure 9: Approximation result (Binary regions) (in detail view)
Figures 8 and 9 show the approximation results of the extracted
building outlines using classification technique presented in this
paper. For a final 3D reconstruction of the buildings, an average
height is estimated for each building to shape prismatic model.
For each building, the prismatic model is shaped using average
height and polygon nodes which are extracted by approximation
approaches. The polygons are merged to form prismatic models
as illustrated in Figure 10. Additional 3D representation is pro-
vided by superimposing the ortho rectified image of correspond-
ing area to the 3D prismatic models models (Fig. 11).
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
¢
me
æ
Par Meter a
©
€v
Figure 11: Superimposing of ortho rectified image on 3D pris-
matic models (zoomed to a smaller region)
4 EXPERIMENTS
Finally, in this section we compare three-dimensional urban re-
gion reconstruction performances of the proposed two different
approaches. Using two large test area images and DSMs of Tu-
nis city (both in 2000 x 2000 pixel sizes), we have performed
our quantitative performance comparisons. For testing shape de-
tection performances, we applied two different analysis; object
based shape detection performance and pixel based shape detec-
tion performance. The object based shape detection performance
considers a building as correctly detected if the algorithm can de-
tect any shape on it regardless of the detected shape. Based on
object based performance analysis, active shape model based ap-
proach detected 1023 of 1151 buildings correctly (88.87% true
positive percentage) and 20 buildings are detected in non-built
areas (1.00%). However, the prismatic model based approach
detected 953 of the 1151 buildings correctly (82.79% true posi-
tive percentage) and 0 buildings are detected in non-built areas.
The pixel based performances are computed by considering how
many groundtruth building pixels are detected correctly. The ac-
tive shape model based approach detected 82.12 % of building
pixels correctly, and 10.08 % of the pixels in the result shape
are falsely labeled as buildings in non-built areas. However, the
prismatic model based approach detected 96.26 % of building
pixels correctly, and 18.54 % of the pixels in the result shape
are falsely labeled as buildings in non-built areas. Shape detec-
tion results show that, the active shape model based approach Is
more successful to detect building structures, and the prismatic
model can miss building locations if the building shape cannot be
represented with prismatic models. However, if the building lo-
cation is correctly detected, the prismatic approach can estimate
more accurate shape for buildings since it does not contain dis-
continues as the active shape based approach has in the connected
active shape models. For testing height estimation performances,
the |
com
does
chec
fron
pari:
diffe
0.72
ing |
the |
step.
Her
3D (
very
proa
pari
heig
to tl
and
algo
ent :
Tun
final
resu
like
anal
men
Are
ima!
Syst
Are
tail
of I
and
ital
Pho
enc
Can
buil
Inte
Fisc
sen
age
AC]
Frac
Tec
In
tion
Har
dete
Jac
and
