> XXXIX-B3, 2012 
ollowing lines after 
ents /(;) and /(i +1) 
t is inserted as /(i - 1) 
segments’ indexes are 
uation will be 
on (3) should be 
—1 (3) 
-]) which is initialized 
1e original parallel 
fining algorithm. 
uygon with the refined 
ited by the intersections 
RESULT 
okyo naming data2 was 
layed in Fig. 6 and Fig. 
| several stages of the 
a. There is one class 
nated. For comparison, 
| once at a time. Due to 
1, the building polygons 
lay random errors. The 
ntation (over 10 degree 
It with proposed one is 
Data2 
1 
16 
  
TOTS 
lerived automatically 
manually and display 
ome buildings on the 
not extracted or not 
modelling algorithm 
r the houses wholly 
re highly correct and 
[here are totally 186 
n this area. Only one 
th whole shape and 
€ are 6 models have 
models locate at the 
s. There are 4 models 
t edges of the houses 
sualization effect the 
is area are illustrated 
correctly extracted 
s were counted, as in 
its shape is taken as a 
nerging occurring in 
buildings and small 
y NDSM generation 
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 
algorithm. Experiments prove that no crossing false 
clusters occurred for any data at any parameters. The only 
possible question is more or less clusters. For the quality 
of the clustering considering their orientation correctness, 
by visually determination, a deviation of less than 10 
degrees is tolerable, and is not taken as an error. The 
orientation correct rate is defined as the number of 
correctly oriented houses divided by the extracted houses, 
as shown in Tab. 2. 
    
(d) Polygons modeled once at a time 
  
45 
       
3 <> CN v e» s AA 
(in blue) and manual polygons (in yellow) 
Figure 6 Stage results of the sample data 
(e) Models 
  
  
  
  
  
Figure 7 Corner and model details 
  
  
  
  
  
  
  
Datal Data2 
Extracted 111 163 
Error oriented 9 1 
Orientation Correct rate 91.89% 99.38% 
  
  
Tab. 2. Performance for the data sets 
6. CONCLUSIONS 
We developed a method of 2D building modelling based on 
human knowledge to the houses. It starts from DSM-image pair 
and end to polygon description. The whole work flow includes 
PPO grouping for building extraction, model hypotheses, 
feature detection, model refining. In each stage, some new 
technique or algorithm are developed. They make every step 
giving correct and accurate result. That is, the edges or lines fit 
that in the image very well. More work will be done for the 
refined modelling for a single house. 
Reference 
Baatz, M. and Schape, A. 1999, Object-oriented and multi-scale 
image analysis in semantic network. /n: Proc of the 2nd 
International Symposium on operationalization of Remote 
Sensing, August 16th-20th.