s for the first image 
ctors. For selected 
manually to see the 
es running out of the 
    
  
  
  
  
  
50/ 200 / 
1.0 4.0 
6 | 332 15.7 
7.44735 [24% 
8 1614 | 385 
7133.81 3123 
$ | 645 .| 424 
  
  
  
' homologous points 
ner points and SIFT 
alucaltion of trifocal 
e bundle adjustment 
the position of the 
ited image sequence 
iologous points. The 
es and the estimated 
1ence can be seen in 
ady visible. Most of 
indows and grouped 
le adjustment of the 
shows smaller errors 
ints due to the weak 
window regions for 
  
ce along a group of 
estimated camera 
sequence orientation 
ig model (Fig. 4). A 
matching to remove 
trees. The local 
derive an estimated 
point is assigned to 
ind similar normal 
direction that differs 
cted. The remaining 
matching with the 
    
  
  
Additionally, a line matching (Frueh et al, 2004) in the image 
space is done to refine the exterior orientation of every image 
(Fig. 5). 
  
Figure 4. Grid model of the building with point cloud (light 
grey) and camera path: light grey: GPS path, dark grey 
transformed estimated camera positions 
  
Figure 5. Image overlayed with the grid model of the building. 
Left: before line matching, Right: after line matching 
3.4 Comparison of extracted Textures 
For every surface of the model, partial textures are generated 
one from every image where the surface is visible. These 
textures normally do not cover the hole surface (Fig. 6). Due to 
the recording configuration, the geometric resolution decreases 
to the roof and the right and shows only a small part of the 
facade on the left. 
  
0000000 
Figure 6. Partial texture of one facade extracted from one IR 
image 
    
In a first step, the partial textures of one surface from one 
sequence are combined (Fig. 7). The resulting combined 
textures show a good fitting in the middle of the images but 
disturbances at the roof and especially on the ground. The roof 
disturbances seem to be caused by the viewing angle and the 
low resolution in all images, the errors near the ground are 
caused by occlusion. 
  
Figure 7. Surface texture generated from the complete image 
sequence 
If a texture is associated with a building surface, we are able to 
compare different textures from different recording times and 
conditions. A straight forward way is to overlay these textures 
(Fig. 8). In this example two infrared textures, one from a 
sequence in the evening and one from a sequence in the next 
morning, are combined. The blue color indicated a cooling 
effect over night. One can see, that the position of the windows 
in the first and second floor are fitting very well, whereas as the 
third floor seems to be blurred. One can also see small diagonal 
lines. These lines are the result of the combination of partial 
textures of one sequence. This combination are not exact the 
same for to textures from different sequences and thus can show 
small differences in the intensity. 
  
Figure 8. Temperature differences of two textures from 
different image sequences. 
4. DISCUSSION AND CONCLUSION 
The matching of the point cloud and the building model shows 
different behaviour depending on the building geometry. In 
many cases, occlusions reduce the number of visible facades in 
the image sequence and thus no 3d points for the matching 
exists. This sometime leads to remaining shifts of the point 
clouds in the facades. These shifts can be reduced by checking 
the edges of the building model against the images using 
projected voxels of the facade borders. It can be seen, that this 
method allows to handle geometric details on the facade which 
have not been modelled in the 3d building model itself. The 
geometry of the recording leads to an quite unbalanced 
distribution of feature points compared to the model surfaces.