to the same 
h as building 
n material and 
a significantly 
ges. For these 
yresent, to use 
. cues for the 
N images for 
ows a block 
most column 
bed above. If 
extracted from 
rmation can be 
formation and 
s analyzed to 
1 in the process 
on for MVS. 
detail next. 
e 22 (displayed 
heir boundaries 
n DEM images 
as being higher 
  
  
  
he magnitude of 
ith a Laplacian- 
ions bounded by 
nt stages of the 
hypotheses thal 
f a line segment 
n we project the 
gion. Figure 19 
e near the DEM 
any of the lines 
onal complexity 
id thus including 
/erify promising 
r hypotheses caf 
‘rease during the 
vidence and wall 
Edward M. Mikhail 
  
evidence. When no evidence of walls or shadows is found, we require that the DEM evidence (overlap) be higher, to 
validate a hypothesis. The 3-D Models constructed with DEM support from the validated hypotheses are shown in 
Figure 24. Comparing it to Figure 20 shows that false detections are eliminated with DEM cueing. Also, the building 
components on the top left and on the lower part are not found without DEM support but found with it. Once the 
buildings have been detected, the DEM can also be improved by replacing parts of the DEM with building models. 
3.3.3 MVS Supported by Thematic Data 
The HYDICE image strip shown in Figure 10 was classified as described in section 3.1 and rectified, thus producing a 
useful thematic map. To extract cues we first extract the roof pixels from the thematic map. Many pixels in small 
regions are misclassified or correspond to objects made of similar materials as the roofs. The building cues extracted 
from this image are the connected components of certain minimum size. 
HYDICE cues are used, in ways similar to those for the DEM cues described above, at different stages of the hypothesis 
formation and validation processes. The linear segments near HYDICE cues, are very similar to those shown earlier in 
Figure 23 with an increased reduction in the number of lines (84%). As with the DEMs the HYDYCE evidence helps 
simplify the hypothesis selection process. The evidence consists of support of a roof hypothesis in terms of the overlap 
between the roof hypotheses and the HYDICE cue regions. The hypotheses are constructed from matching features in 
multiple (two in our Ft. Hood example) images and are represented by 3-D rectilinear components in 3-D world 
coordinates. We can therefore project them directly onto the HYDICE cues image to compute roof overlap. The system 
requires that the overlap be at least 50% of the projected roof area. 
  
   
Figure 25. Building components extracted using 
HYDICE cues 
Figure 24. Building components extracted using 
DEM cues 
  
  
  
  
  
Figure 25 shows the detected buildings using the HYDICE cues. This result shows no false alarms. Once the buildings 
have been detected, the roof class can also be updated. The performance of the MVS system is very similar using DEM 
or HYDICE cues. There will be many cases where the quality of the cues from one sensor may be higher. It is 
appropriate to characterize this quality and combine the support from various sensor modalities. This is the subject of 
our current work. 
3.4 Road Grid Extraction 
The system uses a simple three-dimensional road segment and intersection model and known camera parameters, which 
allows the use of either nadir or oblique views. Roads are assumed to have visible edges without significant occlusions. 
Since we use the geometric structure of the road and the intersection, vehicles and markings on the road are not a 
serious drawback. Indeed, for verification, they may be an important feature. We also assume a regular street grid but 
that the program must detect where the regular grid ends. Some variations in the grid are detected by using a grid model 
that is smaller (e.g. 1/2 or 1/3) than the actual street grid. 
The system requires only a few interactive steps, which could be performed by imperfect automated techniques that 
have been suggested in the literature. By delaying total automation, we are able to focus on the important issues of 
using context and grouping for street grid extraction. The only inputs from the user are three points (i.e. the center of 
three intersections) that give the location, direction and spacing of the street grid. This step can be replaced by 
automatic methods to find dominant direction and spacings, but these are less reliable and not a focus of this current 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 603