RESULTS OF THE TEST ON IMAGE UNDERSTANDING OF ISPRS WORKING GROUP 11/3 
Monika Sester, Werner Schneider and Dieter Fritsch 
Institute of Photogrammetry 
Stuttgart University 
P.O.B. 106037 
70049 Stuttgart / Germany 
{monika.sester/werner.schneider/dieter.fritsch} @ifp.uni-stuttgart.de 
Commission Ill, Working Group 3 
KEY WORDS: data fusion, data integration, model based image interpretation 
ABSTRACT: 
The "Test on Image Understanding" was an initiative started by the ISPRS Working Group III/3 in March 1994. The idea was 
to get hold of and review the the state of art of image understanding techniques, especially in the domain of automatic recog- 
nition and reconstruction of cartographic objects. According to new tendencies in image understanding, as much information 
. as available should be used for the interpretation. In this context, the test provides different data sets of information which is 
generally available. The additional information includes GIS ground truth, but also a digital surface model, color and stereo 
aerial imagery. The scales of the imagery vary from 1:5000 to 1:12000, and the corresponding ground pixel sizes are in the 
range of 0.23 m to 2 m. The tasks slightly vary for the different data sets, but primarily focus on the detection and reconstruction 
of houses, streets, water bodies, field parcels. 
1 Introduction and Overview 
The given data sets represent realistic information, what 
should be generally available to date. Thus the overall ob- 
jective of the test aimed at information interpretation and ver- 
ification. In particular interests were directed to the methods 
integrating further (external) information. Therefore the idea 
behind the test was twofold. On one hand researchers should 
have the opportunity to test and compare their algorithms on 
a standardized, realistic test data set. In this way results are 
getting comparable, and individual approaches or methods 
can be improved. The second issue was to stimulate re- 
search by providing a broad set of different data sources. 
Due to the fact that different knowledge sources for the in- 
terpretation are available, another fundamental issue of the 
whole test comes up: in order to evaluate the results and the 
applicability of the approaches, the underlying strategy and 
the models used have to be made transparent. This implies 
to make explicit the knowledge used - namely the object mod- 
els and the strategy of the algorithms. Concerning the object 
models, some interesting questions arise, e.g. whether spe- 
cific object models are needed or or whether they can be 
generic, whether a 2D-cue of a 3D-object is sufficient for its 
detection (and for its reconstruction, resp.) ? In this way all 
the assumptions the program uses should be clearly sepa- 
rated and not hidden in program code. 
Having an exact description of the procedures applied and 
768 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
the knowledge involved, extensions to other data sets and 
also to other objects seem to be possible. Ideally a separate 
knowledge base consisting of object models and correspond- 
ing strategies helps to handle different kinds of problems. In 
order to clarify this aspect, every participant had to report an 
extensive description of the approach according to a detailed 
questionnaire. The tasks concerning the individual data set 
have not been specified too strictly. The underlying reason 
was to leave open a broad spectrum of possible operations 
on the data. In particular, the data set flat consists of a stereo 
image pair and a DEM generated by shape-from-stereo. The 
objects (buildings) contained in the data sets can either be 
reconstructed by DEM-analysis, by stereo reconstruction or 
by monoscopic interpretation. 
2 Test Setup and Responses 
The test setup has been reported in Fritsch, Sester & Schenk 
[1994]. Although there has been a great response to the data 
(over 300 people have accessed our ftp-server), only a small 
number of scientists actually handed in their results. The fol- 
lowing table gives an overview of the tasks the individual par- 
ticipants of the test solved. The table reveals that the data 
sets including range data (flat, suburb) seemed to be a chal- 
lenge for most of the interested scientist. 
   
  
  
   
  
   
  
   
   
   
  
     
    
    
   
   
   
     
     
   
   
   
  
    
   
  
    
   
  
    
    
     
   
   
    
   
   
    
  
  
    
   
  
  
  
  
  
In the 
the in 
One | 
DEM: 
one ( 
stal ( 
& Ro: 
a follc 
tasks 
be tre 
tion c 
using 
All th: 
on tol 
an op 
2.1 
Data 
Objet 
being 
areas 
Prior 
ing th 
tion a 
used 
jects i 
Strate 
of the 
straig| 
analy: 
Startit 
jects | 
struct