4. The idea was 
iutomatic recog- 
iuch information 
‘mation which is 
olor and stereo 
sizes are in the 
d reconstruction 
r data sets and 
eally a separate 
and correspond- 
of problems. In 
had to report an 
ing to a detailed 
lividual data set 
derlying reason 
sible operations 
sists of a stereo 
rom-stereo. The 
s can either be 
construction or 
nses 
ester & Schenk 
onse to the data 
er), only a small 
results. The fol- 
e individual par- 
Is that the data 
ed to be a chal- 
  
  
data set remstal 
  
  
  
  
  
  
Participants 
data set glandorf data set || data set 
street | field | house | railway || street | field | brook | house flat suburb 
  
  
  
  
  
  
  
    
a 
eianer 
e 
  
In the following the prerequisites, strategies and methods of 
the individual participants are described in some detail. 
One participant (8) used the stereo image pair as a test for a 
DEM-generation program ([Lotti & Giraudon 1994]); another 
one (7) tested his line-tracking program on the data set rem- 
stal (cf. [Trinder & Li 1995]). In the approach (9) of Fierens 
& Rosin [1994] GIS data is used to define training regions for 
a following classification process. Due to the fact that these 
tasks did not exactly match the scope of the test, they will not 
be treated in detail here. However, the focus of the evalua- 
tion concentrates on the reconstruction of man-made-objects 
using prior information (participants 1 to 6). 
All the results reported back to the data provider are based 
on totally automatic strategies which involve no interaction of 
an operator. 
2.1 Uwe Stilla 
Data Source: Stereo image pair, data set flat 
Object Model: The generic model describes a building as 
being composed of two roof parts, namely two rectangular 
areas in 3D. 
Prior Knowledge: Prior knowledge is introduced concern- 
ing the camera parameters and the thresholds in the extrac- 
tion and grouping procedure. The common sense knowledge 
used mainly concerns the scene model, especially the ob- 
jects in a scene: 
> Buildings are rectangular and have a length |_house 
(Lhouse.min « l.house « l.house.max). 
> The two areas of a gabled roof enclose an angle 
gamma (gamma.min « gamma « gamma.max) 
Image related information: 
> Type of primitive objects for structure approximation 
(Type=LINE) 
> The areas of a roof appear as parallelograms 
> The small angle in a parallelogram is alpha (alpha_min 
< alpha < alpha_max) 
> The edges of the roof (connected with the gable) have 
length I_side (l_side > |_side_min) 
> The shorter side of two opposite sides of a parallelo- 
gramm must be at least half as long as the longer side 
Strategy: In a preprocessing step a symbolic description 
of the images is generated, which consists of a collection of 
straight lines (LINE). In both images preprocessing and 2D- 
analysis is carried out independently. 
Starting with the object primitives LINE more complex ob- 
jects ANGLE, U_.STRUCTURE, PARALLELOGRAM are con- 
structed by grouping. In lower levels there is no decision yet 
769 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
if an object is part of a target object or not. Thus, a lot of 
alternative objects are produced. 
The 3D-analysis attempts to find pairs of 2D-objects 
(U_STRUCTURE or PARALLELOGRAM) which are projec- 
tions of the same 3D surface. This is done by selecting pairs 
and examining rays originating at the centre of the projec- 
tion and passing through the vertices of the 2D-objects. The 
2D-objects will be called NOT CORRESPONDING if the dis- 
tance between the rays of pairs of vertices is greater than a 
given threshold. In 3D-domain more complex objects (gabled 
roofs) are constructed, if the conditions in space are fulfilled 
(neighbourhood, location, orientation). 
Pseudo code of the program is given by the set of production 
rules. 
L /NL (angle-shaped) -> A 
A /\ A (u-shaped) -> U 
U /\ L (parallelogram-shaped) -> P 
U /NU (corresponding in 3D) -> CA 
P /NU (corresponding in 3D) -> CA 
P. /\ P (corresponding in 3D) -> CA 
CA /\ CA (building an edge in 3D) -> CE 
(L) LINE, (A) ANGLE, (U) U_STRUCTURE, 
(P) PARALLELOGRAM 
(CA) PART OF ROOF, (CE) HOUSE ROOF 
More details of the procedure can be found in [Stilla 1995] 
and [Stilla, Michaelsen & Lütjen 1995]. 
Results: Detection and reconstruction of 14 buildings (from 
17 buildings in total). 
2.2 Uwe Weidner 
Data Source: Range data, data set flat 
Object Model: The approach bases on generic object 
models, i.e. that buildings are usually higher then their sur- 
rounding topographic surface, that the ground plan of the 
buildings consists of straight lines and that these straight lines 
form polygons, which have edges being orthogonal, parallel, 
and collinear. Furthermore, parametric building models are 
used, namely rectangular buildings with either flat or sym- 
metrically sloped roofs. 
Prior Knowledge: The assumption of the minimal size of 
the buildings and their minimal height is enough to fix control 
parameters for the subsequent segmentation steps. Build- 
ings are assumed to be separate from each other.