3. A CONCEPT FOR OBJECT RECONSTRUCTION 
3.1 General Strategy 
    
    
  
    
   
   
   
Orientation 
parameters 
           
   
Coarse ob- 
ject model 
Pyramid 
generation | 
i=N 
d 
  
Approximate 
values 
v 
pyramids 
  
  
/ Models 
/ Parameters, 
Matching at level i 
  
  
  
  
  
  
Figure 2: General matching strategy 
The general strategy of our concept for object 
reconstruction is shown in figure 2. Two or more digital 
images, their orientation parameters and a coarse model 
of the object surface (e.g. a tilted plane or a cube) provide 
the input for the algorithm. In a first step, image pyramids 
have to be generated from the original images. The 
matching algorithm which is controlled by a small number 
of parameters and by pre-defined object models, is first 
applied to the upper level (level N) of the image pyramids 
with approximate values derived from the coarse object 
model. The matching result is a cluster of points 
supposed to be on the object surface, possibly together 
with some information about surface discontinuities. A 
triangulation of these points delivers the description of the 
object surface at the upper level of the image pyramids 
which is now used as an approximation for the next lower 
level. 
Matching and triangulation are now iteratively applied to 
each level i of the image pyramids; the results of level i 
provide the approximate values for level i-1. The process 
is stopped as soon as the lowest level of the image 
pyramids (i.e. the level with the highest spatial resolution; 
i 2 0) has been reached. 
The matching algorithm at a certain level i of the image 
pyramids is described in more detail in figure 3. Matching 
  
694 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
   
starts with the extraction of features and their mutual 
relations under the assumption of a certain image model 
(section 3.2). The topological relations between 
neighbouring features are described by a feature 
adjacency graph. Basically, feature extraction delivers 
both point and line features as well as homogeneous 
image regions. In our concept, we only use points to 
represent surfaces for the time being. However, the 
topological relations between points, lines and 
homogeneous regions will be used in the course of 
matching. Additionally, the concept can be extended to 
matching of line segments in the future. 
  
/ Image model L» Feature extraction 
! 
! 
I 
1 
1 
| 
I 
| 
I 
I 
x i 
1 
I 
| 
| 
| 
| 
I 
Feature adjacency graph 
Attributes 
v VY 
Geometry of Generation of correspondence 
image ; hypotheses 
Hypotheses for 
object points 
  
  
  
  
  
  
  
  
  
VA 4 
    
      
  
  
  
Local surface 5 Evaluation of correspondence 
model | hypotheses 
\ | 
ILL 
    
  
Consistent object points 
quality of fit 
  
    
  
TL 
OK ? 
[iY 
v 
  
Triangulation 
  
  
  
Figure 3: Matching algorithm 
Having detected point features in two or more images, 
correspondencies between homologous features from 
different images have to be found (section 3.4). Finding 
such correspondencies comprises two steps (Gülch, 
1994): 
e The generation of correspondence hypotheses which 
makes use of approximate values and the orientation 
parameters under the assumption of a model of 
image geometry. 
e The evaluation of these hypotheses under the 
assumption of some (pre-defined) local surface model 
in object space. Only hypotheses consistent with the 
model will be accepted. 
   
cons 
the fi 
make 
fit ol 
indic 
anotl 
evall 
topic 
3.2F 
Many 
gram 
oper 
(Krzy 
desc 
featu 
conte 
this 
infor 
extra 
the r 
We | 
prop: 
desic 
the 
homi 
lines 
blurr 
addit 
spea 
reas 
(figui 
iF 1 
Fi 
graf 
Each 
home 
to a 
meas 
textu 
of th 
imag 
grapl 
betw 
1995 
At tl 
creat 
regio 
avere 
parat 
attrib