minimization of a 
multi-variate 
cost function 
  
Il 
V 
RC_SCENE 
| 
| part 
uem 
—..edpart 
minimization of a «--» RC. VIEW 
  
  
  
SPeCy RC. BOLT. 12mm 
  
multi-variate Foal RC_CAM_PARAM speg, RC ROUNDHEAD. \ 
Wide a a 
cost function À | part spc  w RC BOLT-- soup RC BOLT Z0mma 
/ [1,...,n] . mn & RC HEXAGONAL — ; RC. BOLT 36mni$ 
/ - N N 
| HEN / spec + ar 
| minimization ofa«----»RC OBJECT ^ m RC_CUBI m A “ 1 
multi-variate | | de « . €  9»RC 3HOLED BAR i 
cost function | | spec A. R G BAR Vw RC 5HOLED BAR a 
eon. | - spec RC. 7HOLED BAR 
Y | | D con 
| con con 4 
PE SCENE | Spec 
T | v PE 3HOLED BAR 
N | y PE.B AR-— V* PE 5HOLED BAR 
part | spec 5 "-PE 7HOLED BAR 
AL nl Y : M 
2 part <q 
PE_OBJECT .. Spec 
= > PE CUBE 
part 
spec x. a | 
. s: specy PE BOLT 12mm | 
P $- PE ROUNDHEAD: | 
* PE BOLT I ©. & PE_BOLT_20mm 
PE_HEXAGONAL =. PE BOLT 36mm 
> part part & a^ ; 
em PE BOLT HEAD PE BOLT THREAD , 
\ / con 
con | / 
\ " I FOCUS 
| cdpart \ y 
Y > 1 F A cdpart 
T'OBJECT I REGION 
^ ^ 
Il Il 
Il Ii 
V V 
| Local Linear Maps (LLM) polynomial classifier 
| neural network for for colored region 
| object recognition segmentation 
  
  
tries to detect the parts of a modified perceptual ob- 
ject as they are modeled in the semantic network. 
For our bolt example this yields in instances for *bolt 
head’ and ‘bolt thread’. Every instance has a con- 
crete of I REGION which is created by a polynomial 
classifier of sixth degree using intensity, hue, and 
saturation as features for classification. During this 
instantiation process restrictions for position, color, 
and shape are propagated in a model-driven way. 
Additionally, the restrictions of the current focus 
are taken into account. If a successful instance of a 
perceptual object is created then it is added as part 
of PE_SCENE which refers to all objects in the scene 
detected so far. After this step the focus is adapted 
according to the newly detected object and the next 
object hypotheses — created by the LLM-network 
718 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Figure 7: The hybrid knowledge base for object recognition and 3D-reconstruction 
and yielding into instances of I.OBJECT — are pro- 
cessed. In parallel, the corresponding concept on 
the 3D-reconstruction level can be instantiated acti- 
vating the individual reconstruction of the detected 
object. This process is described in the next section. 
5 Semantic Models for 3D Reconstruction 
and Camera Calibration 
Depth estimation is a well known problem in com- 
puter vision. Various approaches using different as- 
sumptions and heuristics try to reconstruct the miss- 
ing depth information from images. Model-based 
approaches use a priori known geometric object 
models. Model-based 3D reconstruction is a quan- 
titative method to estimate simultaneously the best 
viewpoint of all cameras and the object pose param- 
  
  
   
  
   
       
   
       
       
   
      
  
    
    
  
   
     
  
  
   
  
  
  
  
  
  
  
  
  
   
  
  
   
   
   
   
   
   
  
    
   
  
im 
in 
ba 
Fo 
ba 
is 
cia 
R( 
ist 
nei 
ob 
tec 
col 
the 
be! 
an 
SCE 
de] 
cal 
of 
are 
du: 
an 
jec 
fea 
an 
Th 
tio! 
the 
prc 
pre 
s(4 
e(4