rela- 
ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
  
  
Figure 8: Three input images of an excavation layer at 
an archaeological site. The images are too far apart for 
our shape-from-video process to match features between 
the views. 
  
Figure 9: The reconstruction extracted from the relatively 
wide baseline images of fig. 8, with and without texture. 
ages have less structure than the ones of the town hall and 
are too far apart for our shape-from-video process to get 
its corner matching started successfully. Again, invariant 
neighbourhoods haven been matched and the PDE-based 
dense correspondence search succeeded in finding matches 
for most other pixels. A side view of the resulting 3D 
model is shown in fig. 9, with and without the texture. 
3 AUTOMATIC, CRUDE REGISTRATION OF 3D 
PATCHES 
3.1 Task description 
If partial 3D reconstructions have already been produced 
from different photo sets, model completion may better be 
done in 3D. Similarly, there are a new generation of struc- 
tured light techniques that generate partial, 3D patches 
from each picture that is taken. If there is sufficient over- 
lap between the 3D patches, they can be fitted together to 
build a single, complete shape model. This fitting together 
of patches is usually referred to as 'registration'. 
The state-of-the-art in 3D registration is similar to that in 
2D. Several, excellent methods have been proposed to pre- 
cisely fit together partial, 3D reconstructions from initial