andi- 
)olar 
atch 
ality 
late, 
lage 
radi- 
ang- 
reas 
G| > 
The 
:om- 
tion. 
n be 
irec- 
nea- 
ality 
pon- 
d: a 
arity 
con- 
  
  
  
  
a) left original image 
  
c) disparity map ( dark = far from camera, 
light = near to camera) 
Fig. 2: Correspondence analysis of image pair 
tains the combined quality measure C, for each can- 
didate. Fig. 2 demonstrates the correspondence ana- 
lysis for an image pair of the sequence "house". The 
image sequence "house" consists of a series of 90 
views of the house where the house is rotated 4 de- 
grees around the vertical axis in each view. The cam- 
eras are displaced 15 mm in horizontal direction with 
parallel optical axes and the house is placed approxi- 
mately 400 mm from the camera origin. Fig. 2a and 
b show the left and right input image, Fig. 2c the dis- 
parity map and Fig. 2d the corresponding confidence 
map. Black regions are regions where no disparity 
could be measured. The measured disparity values 
are between 30 and 50 pixel. It can be seen that some 
regions in the area of the roof have false disparity 
values. This areas correspond to regions with low 
confidence because no surface structure is available 
to uniquely select a candidate. 
Scene segmentation 
The correspondence analysis yields a disparity map 
based on local depth measurement only. These mea- 
surements are uncertain and must be merged to re- 
gions that describe physical object surfaces. Based 
on similarity measures of scene depth the segmenta- 
  
  
  
  
  
b) right original image 
  
d) confidence map ( dark = low confidence, 
light = high confidence) 
"house". 
tion divides the viewed scene into contiguous sur- 
faces and merges all disparity measurements of one 
object surface. The object boundaries are corrected 
from the grey level image with a contour approxima- 
tion by assuming that physical object boundaries 
most often create grey level edges in the image. 
Once the depth map is segmented into object regions 
all measurements of one region are interpolated by a 
thin plate surface model that calculates the best qua- 
dratic surface approximation of the disparity map 
based on the uncertain depth measures. A multi grid 
surface reconstruction algorithm described by [Terzo- 
poulos,1988] was chosen to calculate the interpola- 
tion with a finite element approximation. The interpo- 
lation fils out gaps in areas where no disparity 
calculation is possible. The process of segmentation 
and model building is shown in Fig. 3 for a pair of the 
image sequence "house". In Fig.3a the segmentation 
of the object "house" is marked grey, the residual 
background is marked white. Black regions are areas 
that cannot be analyzed at all because these areas 
are visible in one camera only. Fig. 3b displays the 
interpolated disparity map of the object house that is 
converted into a depth map for model building.The 
segmentation could be improved if not only the outer 
429