In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C., Tournaire O. (Eds). 1APRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
the LiDAR measurement, and the 20cm RMSE of residuals at 
checkpoints from image orientation, reflecting the absolute 
accuracy from ray intersection. 
20 
£ 10 
□ side 1 
UllJMlail 
RFRFRFRFRFPFGEGEGR 
Original RoofType 
Detected 
RoofType 
Roof type: RF - rectangular flat roof; 
PF - pentagonal flat roof; 
GE - gable roof with detected facade at bottom edge; 
GR - gable roof with detected facade at the ridge; 
OO - other roof type (detected result) 
Figure 5. Histogram of Rr of all 17 tested buildings on both 
sides. When Rr on one side is above 5, the building 
is detected as rectangular flat roof building. 
As predicted, pentagonal flat roofs could not be distinguished 
from gable roofs. This was probably due to the background 
included in the rectangle segments outside the pentagon. 
However, in the histogram of a pentagonal flat roof (not shown 
in this paper) there was a rise in the curve, but wider and lower 
than the peak for a rectangular flat roof. Therefore, detection of 
buildings of this type could be improved by exploring the use of 
the histograms, or by adjusting the segmentation to different 
shapes. 
5.2.2 Refinement of Roof Height: For each identified 
rectangular flat roof buildings, its height was refined by plane 
sweeping with shorter searching step, and then compared with 
its height measured from LiDAR data. 
ID 
Coarse 
Height 
(m) 
Fine 
Heigh 
t (m) 
LiDAR 
Height 
(m) 
Difference 
(Coarse - 
LiDAR) 
(m) 
Difference 
(Fine - 
LiDAR) 
(m) 
1 
61.06 
61.11 
60.66 
0.40 
0.45 
2 
57.28 
57.23 
57.20 
0.08 
0.03 
3 
65.07 
64.97 
64.97 
0.10 
0.00 
4 
53.40 
53.55 
53.60 
-0.20 
-0.05 
5 
54.72 
54.67 
54.60 
0.12 
0.07 
6 
50.14 
49.99 
50.20 
-0.06 
-0.21 
7 
51.44 
51.34 
51.52 
-0.08 
-0.18 
8 
76.00 
76.20 
76.25 
-0.25 
-0.05 
9 
59.38 
59.38 
59.53 
-0.15 
-0.15 
RMSE 
0.19 
0.19 
Table 6. Comparison between coarse heights, refined heights 
and heights measured from LiDAR data. 
Table 6 presented the coarse and fine heights detected from 
oblique imagery, and their comparison with the measured 
height from LiDAR data for each of the nine buildings. The 
absolute differences between coarse heights and the LiDAR 
heights were from 0.06 to 0.40m, whilst it ranged from 0 to 
0.45m between the refined heights and LiDAR heights. The 
RMSE of both were the same, showing that there was no 
significant improvement on the accuracy of the detected height 
by changing the searching step length from 0.5m to 0.05m. 
However, the final RMSE value of 19cm was reasonable, given 
the nominal accuracy (standard deviation) of around 10cm for 
5.3 Result from Roof Outlining 
Roof outlining was earned out at the detected height, Figure 7 
showed the steps for roof outlining on two example buildings. 
Lines were extracted from the roof area and validated in object 
space using four selected images. The initially extracted lines 
usually included ones from roof structures or some noise, and 
they did not connect to each other (Figure 7(a)). So a rectangle 
was used to fit the bounds of the extracted roof lines (Figure 
7(b)). Then facades were generated by extruding the roofs 
borders to the ground plane (Figure 7(c)). Due to the absence of 
the information on the height of the ground, an assumption of 
40 meters was made, which was the approximate ground height 
of the study area. 
Figure 7. Detection of cubic buildings, (a) image of the original 
building © Blom; (b) initial detected roof lines; (c) 
fitted rectangular roof outline; (d) building cube 
6. CONCLUSIONS AND FUTURE WORK 
The method shown in this paper requires no previous 
knowledge to detect buildings. By combining geometric and 
radiometric features of multi-view oblique imagery, outlines of 
simple cubic buildings can be successfully detected. The height 
accuracy of the detected rectangular flat roof buildings is 0.2m, 
which is acceptable. 
Detection of buildings with flat but not rectangle roofs can be 
achieved by looking for new clues on the histogram from plane 
sweeping, or by adjusting the shape of segmentation. Referring 
to the work by Baillard et al. (1999), buildings with gable roofs 
or inclined roofs will be tested with improved plane sweeping 
strategy. Instead of using only height as the parameter, angle 
will also be used. 
ACKNOWLEDGEMENT 
I would like to thank BLOM Aerofilms for providing the 
Pictometry dataset. Then 1 would like to thank Adam Patrick 
Nyaruhuma for providing the code for image matching and 
David Rossiter for the revision. 1 would also like to thank the 
anonymous reviewer for their precious comments. 
REFERENCES 
Baillard, C., Schmid, C., Zisserman, A. and Fitzgibbon, A., 
1999. Automatic line matching and 3D reconstruction of 
buildings from multiple views. In: International Archives of the 
255 
»- 
■H