1€ 
te 
se 
  
  
House Extraction Model 1 
Left Image Right Image 
tr 
Region Pair 
  
  
  
  
House Extraction Model 2 
LeftImage Right Image roof 
HL 
Region-Lines Pair 
  
  
  
  
  
  
House Extraction Model 3 
Left Image Right Image 
~~ 
J E 
~/ — 
   
  
  
Ground 
Line-Line Pairs 
  
  
  
Fig.3. House hypothesis extraction models. 
It is obvious that the possibility and reliability(P&R) of 
house extraction of above three models are different. 
Since the region-based process has the lowest ambiguity 
in stereo matching, model 1 has the highest P&R among 
above three models. 
5. HOUSE CHANGE DETECTION FOR 
GIS DATABASE REVISION 
As mentioned in Chapter 1, house changes are limited to 
two states in our current research : emergence and 
demolition. It means that we will not take into account 
the state of continuous change. 
5.1 New-Built House Detection Model 
These house hypothesis extracted from above three 
models which can not find their correspondences in 
existing database, are thought to be new-built houses. 
Since the new-built houses have no records in existing 
GIS database, the possibility and reliability(P&R) of 7 
new-built house detection is basically relied on the 
quality of house hypotheses extraction. Furthermore, as 
the house hypotheses captured from the reign-region pairs 
have the best quality and such hypotheses are generally 
783 
affected by the threshold values in edge and region 
segmentation processes, the iterative process is generally 
needed for improving the detection possibility and 
reliability. 
5.2 House Demolition Detection Model 
For an existing house in GIS database, if there exists 
corresponding house-like region or region pair and their 
overlayed area is larger than a threshold T, we may say 
that the house is still there. Otherwise, we say that such 
a house has already disappeared. It is then easy to find the 
demolition houses by comparing the house hypothesis 
extracted from model 1 and model 2 with existing 
database. 
However, it is unreasonable to try to find all demolitions 
from region pairs. Hence, 3D lines should be also an 
important source for demolition detection. A 3D line 
based extinction detection algorithm was then developed. 
It is summarized as follows : 
1) seeking all possible candidates of 3D lines relevant 
to existing house A in GIS database. The possible 
candidates here are these 3D lines that close enough to an 
edge of a house(e.g., average distance « 3 pixels) and 
having relative angle less than 25 degrees. 
2) matching each edge of existing house A with 
extracted 3D lines from step 1). In this matching scheme, 
the height of each 3D line is the major constraint for best 
match selection. 
3) shifting house A with a finite tolerance and repeat 
steps 1) and 2) to find the best matches. 
After above iterative process finished, for the existing 
house A, if there exists enough corresponding house-like 
3D lines, we can say house À is still there. Otherwise, 
we can get the decision that "house A has disappeared". 
6. EXPERIMENTAL RESULTS AND 
CONCLUSIONS 
Experiments were performed to examine the accuracy of 
the methods described above. As the page limitation, 
only one of the testing results is demonstrated in this 
paper. The original photographs used in this experiment 
were taken in 1988 scaled about 1:5000. The images 
utilized here are 1000X1000 pixels with an range of 
intensity level from O to 255. The sequential 
experimental results are illustrated in Fig.4. The results 
show that the possibility and reliability of house 
extraction and house change detection is affected by the 
threshold values in edge extraction and about 90-95% of 
the houses are extrected without interative processes 
whereas the extraction rate may be increased at rate of 3- 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996