desl. 
ge 2 
ie matched 
ned above, 
stant points 
fore all the 
mages (see 
s step, i.e. 
led to 3 m. 
ular size of 
ade. 
  
  
  
Figure 12 shows some of the lines longer than 3 m. 
Predominantly those are a result of horizontal edges, which 
interpretation planes intersect in a very small angle and thus 
often far away from the facade they belong to. Since these lines 
are eliminated from the final set, vertical lines are the only ones 
visible on Figure 11. For example, in case 7 (despite the good 
match) only one line is left after the intersection (i.e. the vertical 
line in Figure 10e). 
    
    
  
a) case 4: buff-5, end-points=1, length=0.8, angle=1 
b) case 6: buff-3, end-points=2, length=0.8, angle=1 
Figure 11: Vertical 3D lines back projected on the images 
Em EB ime Gp Dower Heb 
  
  
Projection center 1 
  
Figure 12: Intersection of interpretation planes. 
4. DISCUSSION 
Currently, the experiments are concentrating on tuning 
thresholds to reduce the number of possible candidates for 
matching and improving the intersection of the matched 
candidates. 
4.1 Edge matching 
The experiments clearly showed that the utilisation of a rough 
3D model (e.g. façades of interest) significantly improves the 
quality and quantity of candidates for matching. The benefits of 
this approach can be summarised as follows: 
  
  
IP. Deviant: Doe 2 
e The number of edges to be processed is limited to 
those that do belong to the facade of interest. 
e All the edges from the first image can be transformed 
to approximately the corresponding position on the 
second image 
e The search of candidates for matching can be 
conducted in a very limited area of interest around 
detected edges. Compare to the epipolar line match, 
which fails to match edges of which the endpoints are 
in the same epipolar plane, the algorithm successfully 
finds matches regardless the direction of the edge. 
e The angles of interest allows to eliminate fake edges 
detected in one of the images like shadows, 
reflections, branches of trees (a very common case for 
images taken from street level), etc. 
However a large number of edges detected on one image (that 
may be considered as real 3D features) still cannot be matched 
due to a number of reasons: 
e Lack of visibility (e.g. the facade is only partly visible 
on the second image) or occlusion, possibly by other 
objects (e.g. trees). 
e The edges are not detected on the second image due 
to lower contrast. 
e The position of the feature changed while the images 
were taken (e.g. a window or door is opened or 
closed). 
e The accuracy of the rough model used for the depth 
assessment. Features that are in front of or behind the 
used plane of the facade are systematically shifted to 
the right or left on the second image. This shift may 
appear larger than the interest area used for finding 
candidates. 
e The area of interest (buffer) depends very much on 
the size of the features that can be expected and has to 
be tuned very carefully by many experiments with 
different images. 
e Since the visibility of edges is not equal on the 
different images, the same edge (even well visible) 
may appear with different length (covering even two 
features). For example, the edge on the upper-right 
window (Figure 8a) is wrongly detected as a very 
long edge and it will be matched with two edges 
(Figure 9c). The two constraints, i.e. end points and 
difference in the length of the candidates, will most 
commonly eliminate such edges (compare with Figure 
10), although real 3D line features have to be 
encountered there. An eventual solution could be 
found by tuning the parameters that are used for the 
edge detection. 
4.0 Interpretation plane intersection 
After matching the edges of the two images, the interpretation 
planes of the two corresponding edges are intersected to obtain 
the parameters of the related line in space. However, the quality 
of the parameters of this 3D line depends on several factors: 
e The quality of the match. Are the matched edges 
indeed projections of the same object edge? 
e The precision of the two interpretation planes that 
depends on: 
o The precision of the parameters of the 
edges in the images (again dependent on 
-237- 
RR SR RNS 
ER AL i