:e. September 1-3, 2010 
ln: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). 1APRS. Vol. XXXVI11. Part 3A - Saint-Mandé, France. Septembcr 1-3. 2010 
53 
(e) 
and (e) true-false. 
Hiding to a detection 
1 as an FP (Fig. 5(a)). 
j corresponding to a 
i is marked as an FN 
sponding to a detec 
ted yet: suppose the 
/2 ) •••? br,jk } (k > 1) 
¡stances to the center 
overlapped reference 
es in o r . 7l are sorted 
e ascending order of 
r-.ji • If b d; i 1 and b dti 
te closest overlapped 
e following steps are 
idence between bd,i 
n as TPs (Fig. 5(c)). 
b r .j s (2 < s < к) in 
rlapping entity o r ,j s . 
r distances reference 
mtity than reference 
ее 1 is an FP. 
es b d ,i t (2 < t < l) 
r Od,i t of bd.it con - 
which is obviously 
nd bd.i overlap each 
then bd.i t is marked 
lo not overlap each 
id.%t which becomes 
Otherwise, if Od. M 
: entities (including 
Od.it' 
detection and refer- 
i overlapping entity 
on. the correspond 
ed (as an FP or FN) 
rtice, in most cases 
, the above iterative 
ions. 
tion. Completeness 
Haithcoat, 2005) or 
?ss C r , also known 
?/ have been adopted 
ion rate is the per- 
ities in the detected 
f overlap in the de 
ls the percentage of 
' reference entities, 
rntage of reference 
entities which are overlapped by more than one detected entity 
(see (Awrangjeb et al., 2010) for formal definitions). 
A total of 7 pixel-based evaluation indices are also used, these 
being: completeness C mp . also known as matched overlay (Song 
and Haithcoat, 2005) and detection rate (Lee et al., 2003), cor 
rectness C rp and quality■ Qi v from (Rutzinger et al., 2009); area 
omission error A oc and area commission error A ce from (Song 
and Haithcoat, 2005) and branching factor Bf and miss factor 
Mf from (Lee et al., 2003). 
Root-mean-square-error (RMSE) values (Song and Haithcoat, 2005) 
estimate the geometric positional accuracy. For each one-to-one 
correspondence between detected and reference set, RMSE is 
measured as the average distance between a pair of detected and 
reference entities. Therefore, the RMSE is measured for TPs 
only, but not for FPs, FNs and MDs. 
5 PERFORMANCE STUDY 
5.1 Data Sets 
The test data set employed here was captured over Fairfield, NSW, 
Australia using an Optech laser scanner. Four sub-areas were 
used, the first covering an area of 248?n x 210m (Fig. 2(a)), the 
second covering an area of 155m x 219m. (Fig. 6(a)), the third 
covering an area of 228m x 189?n (Fig. 6(b)) and the fourth cov 
ering an area of 586m x 415??7 (Fig. 6(c)). Last-pulse L1DAR 
data with a point spacing of 0.5m was used. Four RGB colour 
orthophotos with a resolution of 0.15m were available for these 
areas. The fact that the orthoimage did not contain an infrared 
band was circumvented by computing a pseudo-NDVl image us 
ing the assumption that the three image bands R-G-B are in the 
order of IR-Red-Green in order to be used in the standard NDVI 
formula. 
The orthoimagery had been created using a bare-earth DEM, so 
that the roofs and the tree-tops were displaced with respect to the 
LIDAR data. Thus, data alignment was not perfect. Apart from 
this registration problem, there were also problems with shadows 
in the orthophotos, so the pseudo-NDVI image did not provide as 
much information as expected. 
Reference data sets were created by monoscopic image measure 
ment using the Barista software (BaristaSoftware, 2010). All 
rectangular structures, recognizable as buildings and above the 
height threshold Th were digitized. The reference data included 
garden sheds, garages, etc., that were sometimes as small as 10m 2 
in area. Altogether, 70, 62. 60 and 370 buildings from the four 
test scenes formed the reference sets. 
5.2 Results and Discussion 
Table 1 shows the object-based evaluation results and Table 2 
shows the pixel-based evaluation results. The geometric accu 
racy (RMSE) for three scenes was 1.99m, 1.95m, 1.95m and 
2.38m with an average accuracy of 14.5 pixels (2.17?n). 
In object-based evaluation, more than 94% completeness and cor 
rectness resulted in an average 91 % quality with at least 5% build 
ings being detected multiple times. The reference cross-lap rate 
was higher than the detection cross-lap rate, since some nearby 
trees were detected along with the actual buildings. In pixel- 
based evaluation, while 81% of building areas were completely 
detected, resulting in a 19% omission error. 88% of detected ar 
eas were correct, offering a 14% commission error. Since the 
miss factor and omission error were larger than the branching 
Figure 6: Detected buildings on the orthoimages. 
factor and commission error, respectively, the false positive rate 
of the proposed technique is lower than its false negative rate. 
Overall, both in object- and pixel-based evaluations, the proposed 
detection technique performed better on Scene 1 than on Scene 2 
in terms of all indices except multiple detection rate and detection 
overlap rate. There were two reasons for this: a) some buildings 
were detected twice in Scene 1, and b) though in Scene 1 all true 
buildings were detected, in Scene 2 some false buildings (actually 
trees) were detected and some true building parts were missed. 
Scene 3 performed better than Scenes 1 and 2 in pixel-based eval 
uation whereas Scene 3 gave higher cross-lap and detection over 
lap rates in object-based evaluation due to multiple detection of 
complex industrial buildings. Almost the same was observed for 
Scene 4. In the geometric evaluation, in terms of RMSE, there 
was about 0.4m worse positional accuracy for Scene 4 than the 
other three scenes. 
The same Fairfield data set was previously employed by (Rotten- 
steiner et al., 2005), (Rottensteiner et al., 2007) and (Rutzinger 
et al., 2009) to investigate automated building extraction. How 
ever, in those investigations, two different threshold-based evalu 
ation systems were employed and the Dempster-Shafer (DS) de 
tector was evaluated using completeness, correctness and qual 
ity. (Rutzinger et al., 2009) has presented results of pixel-based 
evaluation of the DS detector showing that it can offer higher 
completeness and quality than the proposed detector. However, 
in object-based evaluation the DS detector offered much lower 
completeness and quality than the proposed detector. The supe 
rior performance of the DS detector in pixel-based evaluation was 
largely due to the adopted evaluation systems, (Rottensteiner et 
al., 2005) and (Rutzinger et al., 2009)) which excluded FP and 
FN buildings from evaluation and established many-to-many re 
lationships between the detected and reference sets. Moreover, 
unlike the proposed detector the DS detector was excessively sen 
sitive to small buildings (performance deteriorated with the de 
crease of building size) and buildings smaller than 30m 2 could 
not be detected (Rottensteiner et al., 2007).