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 
the point cloud was achieved by fusion of several flights. This 
has introduced inconsistency in the dataset. 
The second test site is part of the campus of the University of 
Melbourne, Australia. The data was collected in a recent 
campaign by Optech ALTM Gemini with 4-5 pts/m?. The scene 
contains larger buildings. Due to low density of the point cloud, 
only a few building walls were illuminated by the LIDAR 
sensor. 
4.1 Parameter setting 
Adaptive scale factor for searching neighbours is mainly 
depended on the initial À selection. Due to the final scale factor 
is three times less than initial scale factor, the searching region 
is nine times less than initial one. Initial is estimated based on 
point density and visual inspection for two datasets and selected 
as 90 and 40 empirically. 
Ideally, the local surface variation index should be zero for a 
planar point. Thus, a small value can be assigned as the 
threshold. Figure 4 shows the results of a portion of the 
Enschede dataset with threshold values (Ij,,) of 0.005, 0.01, 
0.015 and 0.02 respectively. It can be seen roof points were 
largely misclassified with 0.005 and 0.01. This is because the 
point cloud was a fusion of several acquisitions with 
discrepancies (up to 3-6 cm by manual inspection). A larger 
value of the threshold can account for the quality of such data, 
as shown in the results using 0.015 and 0.02 as the threshold. 
However, the result of value 0.02 led to misclassification of 
vegetation points as planar points. Therefore, the optimal 
threshold has been set to 0.015 in the reported testing. 
  
: ER 
CREE SEEN 
Figure 4. Detected planar points using threshold of 0.005 (a), 
0.01 (b), 0.015 (c) and 0.02 (d) for local surface 
variation index. Red indicates planar points while 
green for non-planar points. 
To set an appropriate tolerance value T4, for segment 
generation, four training samples of flat surface were selected 
from the Enschede dataset. The standard deviations of these 
four sites were calculated and listed in Table 1. 
  
  
  
  
  
  
  
  
  
Flat site 1 2 3 4 
Standard 0.053 | 0.032 | 0.046 | 0.064 
deviation (m) 
Number of 1446 2462 | 2711 | 804 
points 
  
Table 1. Standard deviations of flat plane 
Sites 1 and 4 were captured in multiple flight lines and thus 
have larger standard deviation. A confidence interval of 26 was 
defined as the tolerance. Taking the sample size and the RMS 
into account, the tolerance distance value for the Enschede data 
116 
was set to 0.09 m. A similar procedure was applied to the 
Melbourne dataset, and the optimal local surface variation and 
tolerance distance toward plane as 0.02 and 0.10m, respectively, 
were selected. 
4.2 Verification 
Correctness as well as completeness for verification of 
extraction result was performed by manually outlined building 
facades from point clouds. A buffer from reference wall facades 
was preformed for evaluating correctness and completeness. 
The correctness was calculated as the length of the extracted 
lines inside the buffer divided by the length of all extracted lines. 
The completeness was defined as the length of the extracted 
lines inside the buffer divided by the length of the reference 
lines. The width of buffer is selected as 40cm which we believe 
the true position of wall facade should inside in this research. 
  
  
  
  
  
  
  
Site 1 2 
Correctness 0.75 0.65 
Completeness 0.62 0.32 
  
Table 2. Correctness and completeness of verification 
4.3 Result of wall reconstruction 
The datasets were processed with the methods described in the 
previous section. A portion of raw point cloud of the Enschede 
data is shown in Figure 5. The seed points were extracted and 
the segments were coded by colours. It can be seen that the 
building roofs were successfully extracted as planar segments 
while the vegetation points were treated as non-planar (shown 
in white in the Figure 5). Since the site in Enschede is quite flat, 
the ground surface was also clustered into several large planar 
patches. The detected walls are presented in Figure 6. On the 
left of the figure, walls are shown in 2D while the 3D wall 
facades are shown on the right. Some gaps exist in wall facades. 
This is caused by absence of LIDAR points. It was also noticed 
that a few small facades were detected in vegetation area. These 
false detections can be removed easily due to their sizes are 
small. 
    
    
Figure 5. Segmentation result in Enschede site. 
(non-segmented points are represented in white) 
  
VA 
Figure 6. Extracted wall facades in the Enschede data. 
Fina 
of b 
ridg 
Seve 
in F 
outli 
buik 
corn 
full | 
  
  
Fig 
Fig