Images 
YIQ color 
space 
Udine detection tech- 
>hown in gray-scale). 
/ building mask. 
h of its point in the 
y a completely black 
if its point in the sec- 
f a completely white 
•imary building mask 
•eturns below X), and 
ate filled areas from 
>ove the same height 
ated masks for a test 
imary building mask 
wed to obtain these 
;k shapes in M p are 
d extended. Finally. 
n edge detector and 
minimum building 
ected on each curve 
t al., 2009). On each 
i corner and an end- 
rs are not available, 
In order to properly 
lding edges, a least- 
lied. With each line 
i-point’ indicates on 
:d. In older to avoid 
detected tree-edges, the mean of sigma of the NDVI value T is 
calculated on both sides of each line segment. If it is above a 
threshold T n dvi — 04 for any side, the line segment is classed as 
a tree-edge and removed. 
In the second step, the line segments are adjusted and extended. 
The adjustment is based on the assumption that the longer lines 
are more likely to be building edges. In an iterative procedure 
starting from the longest line and taking it as a reference, the 
angle between the reference and each line in its neighbourhood is 
estimated. The lowest rotation angle 0 r is recorded for each line 
over all iterations (for all long lines taken as references). After the 
iterative procedure, each line is rotated with respect to its centre 
by 0 r - If a rotation angle is not recorded for a line it is removed 
as a tree-edge. Each adjusted line is then extended iteratively by 
considering the percentage of black pixels (more than 70%) and 
applying the NDVI threshold to the building side. 
Finally, initial buildings are formed among the extended line seg 
ments. In an iterative procedure, an initial building position is 
detected using the first longest line segment, another using the 
second longest line segment and so on. Before detecting a rect 
angle using a line segment in each iteration, the line segment is 
first tested to ascertain w'hether it is already in a detected building. 
In order to detect an initial building on a line segment, an initial 
rectangle (of width 1.5m) is formed on the building side and then 
three of its sides are extended outwards with respect to P, n us 
ing the same technique applied to extend the extracted lines. Fig. 
3(a) shows the initial detected buildings on the test scene. 
3.3 Final Buildings 
The final building positions are obtained from their initial posi 
tions by extending each of the four sides. Image colour informa 
tion and the two masks M p and M s are considered during the 
extension. The colour information is basically used to extend the 
initial positions; M p is used to avoid unexpected extension of an 
initial position over more than one actual buildings, and M s is 
used to avoid unexpected extension of an initial position beyond 
the actual building roof. 
An initial building position may go outside the actual building 
roof due to a misregistration between the orthoimage and the LI- 
DAR data. In order to avoid this, since the initial position will be 
extended outwards while obtaining the final position, its length 
and width are reduced by 15% before extension. For each re 
duced building position ABCD. the dominant colour threshold 
pairs are estimated using colour histograms for intensity Y. hue 
/ and saturation Q, respectively. Each dominant colour threshold 
pair indicates a range denoted by its low l and high h values. 
There may be overlaps between the detected initial positions. It 
is hard to decide which overlap is unexpected and which is natu 
ral. If an initial building is completely within an already extended 
building or building part, it is removed assuming that it is an un 
expected overlap. Otherwise, it is extended assuming that it is a 
natural overlap. 
The initial building positions are sorted in descending order of 
their length or area, since both of these sorted lists were found 
to offer the same performance. Starting from the initial build 
ing having the longest length or largest area, its four sides are 
extended outwards separately. While extending each side in an 
iterative procedure, the percentages of black pixels in both M p 
and M s and of dominant colour components within the estimated 
colour threshold pairs are estimated. The side is extended if per 
centages of black pixels are above 90% and those of dominant 
colour components are above 40%. Fig. 3(b) shows the final 
detected buildings on the test scene. 
(a) (b) 
Figure 3: (a) Initial and (b) Final buildings. 
4 PROPOSED EVALUATION SYSTEM 
The proposed threshold-free evaluation system makes one-to-one 
correspondences using nearest centre distances between delected 
and reference buildings. The reference buildings are obtained us 
ing manual measurement from the orthoimagery. Altogether 15 
indices are used in three categories (object-based, pixel-based and 
geometric) to evaluate the performance. Most of these have been 
adopted from the literature and the rest are proposed for a more 
complete evaluation. 
4.1 Detected and Reference Building Sets 
For evaluation, two sets of data were used, in which each building 
is represented either as a rectangular entity, for T shape building, 
or a set of rectangular entities, for ‘L’, ‘U’ and ‘C’ shapes. The 
first set Bj — where 0 < i < m and m is the number 
of detected rectangular entities, is known as the detected set. It 
is obtained from the proposed detection technique. Each entity 
hd.i is an array of four vertices and the centre of a rectangular 
detected entity. The second set B r = {b r .j}. where 0 < j < n 
and n is the number of reference entities, is termed the reference 
set. It is obtained from manual building measurement within the 
orthoimagery. Each entity b r .j is an array of four vertices and the 
centre of a rectangular reference entity. 
To find the reference set B r , manual image measurement is used. 
Any building-like objects above the height threshold Ty, are in 
cluded in B r . As a result some garages (car-ports) whose heights 
are above T), are also included, but some building parts (veran 
das) whose heights are below Th are excluded. Different building 
parts are referred to separate rectangular entities. Consequently, 
there is one entity for T shape, two entities for ‘L’ shape, three 
entities for ‘U’ shape, four entities for ‘C’ shape and so on. 
4.2 Overlapping Sets 
It is natural that different rectangular entities of the same building 
overlap each other. In B r , two overlapping entities must always 
belong to the same building and represent two connected building 
parts (Fig. 4(a)). Such an overlap is defined as a natural overlap 
and for identification purposes a building identification number 
bid is assigned to each reference entity, this being stored in b r .j, 
in addition to the four vertices. Entities of the same building are 
assigned the same bid, but those of the different buildings are 
assigned different bid values. 
In Bd, the situation is different. Here two overlapping entities 
may belong to the same building and represent two connected 
building parts. In such a case, this overlap is a natural overlap 
(Fig. 4(a)) and it is not counted as an error in the proposed eval 
uation. In all other cases, the overlap is counted as an error in 
the evaluation system. For example, the overlapping entities may 
represent the same building (multiple detection, Fig. 4(b)) or con 
stitute combinations of true and false detections (Figs. 4(c)-(e)). 
51