a) Scene segmentation mask 
  
  
  
  
C) textured object "house" with triangular surface 
mesh superimposed 
Fig.3: 
boundary of the house but also the inner edges were 
known. By simple human interaction these edges 
could be classified and would greatly improve surface 
modeling. A human interaction module that intro- 
duces knowledge about the objects will therefore be 
of great help. 
t — 
The interpolated disparity map contains the visible 
scene geometry measured from the camera view 
point. When the scene contains occluded surfaces 
then the camera must be moved around the objects 
and the measurements from multiple view points 
must be included. For that purpose the 2D disparity 
map is first converted into a 3D surface description 
that can be modified to include hidden surfaces. The 
disparity map can be transformed into a depth map 
containing absolute scene geometry when the binoc- 
ular camera geometry parameters are known. This 
  
Scene segmentation and model building. 
430 
   
b) interpolated disparity map (dark = far from camera, 
light = near to camera) 
  
  
  
  
d) Synthized view of object "house" after integration 
of depth measurements from 6 view points, 
each view rotated 4 degrees 
depth map is sampled and a 3D-surface for each 
contiguous object is constructed by spanning a trian- 
gular wireframe in space. The samples of the depth 
map generate control points in space that are con- 
nected through plane triangular patches. The triangu- 
lar mesh was chosen because itis capable to approxi- 
mate arbitrary surface geometries without 
singularities. On the surface of each triangular patch 
the object surface texture can be stored in a texture 
map from which a naturally looking view ofthe original 
objects can be synthesized with computer graphics 
methods. In Fig. 3c a synthesized view of the 3D 
object "house" is shown with the triangular surface 
mesh superimposed as black lines. The surface ge- 
ometry was calculated from the interpolated disparity 
map while the surface texture was taken from the left 
original image. In Fig. 3d a first result of the informa- 
tion fusing stage is shown which will be described 
below.