planes in 
tor p. To- 
onstraints, 
he generic 
mage data 
ency. Nu- 
vhich best 
ithm finds 
ll NV cost 
(6) 
ction. For 
ted gradi- 
functions. 
gated gra- 
ints. Only 
int of new 
ice needed 
archically, 
st step the 
refined in 
f hypothe- 
angle be- 
legrees in- 
will return 
thesis has 
  
(amera- 
ion 
nts 
ement con- 
n the other 
atures like 
/o ways of 
the model 
numerical 
the model 
o the scene 
:d by mov- 
constraints. 
; used as a 
imation the 
roximation 
al steps the 
Constraints are hence exploited both for improved surface recon- 
struction and camera pose estimation, thus leading to a consistent 
and closed surface description from multiple viewpoints. 
6 Experimental Results 
This section deals with some experimental results that have been 
achieved with the system. In the first part we show the incremen- 
tal reconstruction of an existing appartment building. The second 
part deals with the view point registration that can be achieved 
with the system. 
6.1 Hierarchical Surface Reconstruction 
Figure 9 shows the input from the first of two camera viewpoints 
for the modelling process: The original camera image (a), a depth 
mep of the scene (b), which has a large number of dropouts, and 
a manually created segmentation of the scene (c). 
  
Figure 9: Input data for the modelling process: image (a), depth 
map (b), segmentation (c) 
In a first step the approximate shape is reconstructed consisting 
only of the main walls and the front roof (fig. 10 left). The knowl- 
edge base inserts the back walls although they are occluded in the 
first viewpoint. À number of edge, position and angle constraints 
are then imposed on the model, which serve to improve the initial 
model’s shape. 
   
Figure 10: Incremental surface reconstruction 
In the same optimization process the second viewpoint position 
597 
  
  
is estimated through edge and position constraints that link the 
model to the other camera position. After computing the basic 
form, the main parts are fixed and used as references for the in- 
cremental refinement of the model. As shown in fig. 10 (right) 
smaller parts like oriels are aligned through parallel constraints 
to the already determined main walls. 
Figure 11 depicts the completed wireframe model of the build- 
ing. In a last reconstruction step the original image information 
is backprojected onto the model, which can be seen in fig. 12. It 
has been generated from two stereo images, depth maps, a manual 
scene segmentation and constraints derived from the knowledge 
base. 28 surface elements are connected by 51 constraints. The 
model is completely closed and has been textured only from two 
original images. The camera position of the second camera has 
been determined together with the model's construction. 
  
Figure 11: Reconstructed Wireframe 
  
Figure 12: Textured Model 
6.2 Viewpoint Registration 
During view point registration new camera positions are added 
to the scene description. They can be used for incrementally re- 
fining the model. To show the performance of the system during 
viewpoint registration a reference object of well known geome- 
try has been used. Figure 13 shows two original images (front 
and right side). The estimated shape in form of a simple box is 
overlayed. 
  
Figure 13: Two views of the reference object using already regis- 
tered camera positions 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996