inbul 2004 
10my and 
pondence 
er Vision 
DAR Data 
> Descrip- 
)gramme- 
nces, Vol. 
etermina- 
qm: ISPRS 
Mapping 
Teller, S. 
rumented 
plications 
Analysis 
Imagery. 
ichine In- 
. A Gibbs 
ly Sensed 
ion 57(2), 
rees from 
> Interna- 
e Sensing 
pp: 157 
quisition: 
incaise de 
56-63. 
riteria for 
nd Pattern 
\utomatic 
tal Eleva- 
y and Re- 
echniques 
rom Pho- 
Computer 
m for Au- 
ion of Im- 
metry and 
SEMI-AUTOMATIC OBJECT EXTRACTION - LESSONS LEARNED 
E. Gülch *, H. Müller^, M. Hahn? 
“ Stuttgart University of Applied Sciences, SchellingstraBe 24, 70174 Stuttgart, Germany — 
cberhard.guelch@hft-stuttgart. de, michael.hahn@hft-stuttgart.de 
° Snowflake Software Ltd, 8a Carlton Crescent, Southampton, SO15 2EZ, United Kingdom — 
hardo.mueller@snowflakesoftware.co.uk 
IC WG IVIV 
KEY WORDS: Automation, Extraction, GIS, Imagery, Object, GML, High resolution 
ABSTRACT: 
This paper summarizes experiences and views gained through development and application of semi-automatic object extraction and 
gives recommendations for further developments. The last decade has seen a lot of research efforts in the field of object extraction 
from aerial imagery and digital surface models from airborne laser scanning. Despite these tremendous efforts, very few approaches 
are commercially available and used in practice. There are probably many reasons for this. Amongst others we can find: economic 
crisis of many states and cities and a general decrease in photogrammetric applications in many countries, which are both external 
factors and hardly can be changed, but which increase the need to look for new markets. On the other hand, we can observe internal 
problems as well: too high expectations which are not fulfilled, quickly changing focus in developments on automated feature 
extraction modules for real world applications, lack of standards, few empirical comparisons etc. In this respect there is a need to 
bundle expertise and resources and ISPRS and EuroSDR (European Spatial Data Research, formerly OEEPE) are two organizations, 
that are capable of providing the necessary platform to give a new push to research and development in reliable automated feature 
extraction. 
1. INTRODUCTION 
In practice the extraction of topographic objects from images 
for generating and updating GIS databases is carried out 
interactively based on mono or stereo plotting. By the 
tremendous success of mobile phones the need to have real 3D 
data in city areas had put a big push to photogrammetric 
applications and to research and development. Having made 
digital orthophotos, aerial triangulation and to a certain extent 
also the DTM matching more or less automatic, the automation 
of feature extraction is still unsolved and thus interesting for 
research. 
Numerous efforts have been made in the past to automate the 
acquisition of point, line, and area features in aerial imagery. 
Overviews are e.g. given in (Gülch 2000, Grün & Baltsavias 
2001, Ohlhof et al. 2000). Fully automated (autonomous) 
Systems, however, are until now in the research stage or can be 
only used for limited purposes (Grün& Baltsavias 2001). Semi- 
automatic systems are e.g. described in (Gülch et al. 2000, 
Giilch&Miiller 2001, Inpho 2004, CyberCity 2004, Ulm 2002). 
This paper summarizes experiences and views gained through 
some years of development and application of semi-automatic 
object extraction. The program inJECT has been introduced by 
Inpho GmbH some years ago to open the field of semi- 
automatic object extraction with a focus on the extraction of 
buildings from aerial imagery. This software has lately been 
enhanced substantially to include other objects as well. A 
specific application is described in (Ohlhof et al. 2004) with 
very advanced automated tools. 
We will present some of the basic features of this software and 
include them in our discussion on lessons learned. Finally, 
conclusions are drawn and recommendations are given for 
further developments in this field. 
189) 
VD 
2. SYSTEMS AND DEVELOPMENTS 
We have seen more than a decade of development in automated 
feature extraction in photogrammetry with numerous efforts on 
a broad range of applications, but despite high expectations 
after the good experiences with e.g. automatic aerial 
triangulation, there are practically only three or four 
commercially available systems on the market that are 
specifically designed to deal with photogrammetric automated 
feature extraction for a wider range of applications and a wider 
range of objects (cf. Gülch et al. 2000, Gülch&Müller 2001, 
Ulm 2002). Some specific semi-automatic object extraction 
systems are used by institutions for own production purposes 
and are not for sale. Concerning the automated feature 
extraction there is a tendency for stand-alone modules, which 
are not depending on stereo viewing capabilities. Lately there is 
a trend to support the extraction of a variety of features, not 
only for buildings, but also linear or area features. With some 
few exceptions we can not find classical Digital 
Photogrammetric Workstation extended with substantial 
modules to support semi-automatic feature extraction. This puts 
several questions: Is there no market, are the tools not good 
enough, or? 
LESSON i: Despite- the feature 
extraction has so far not attracted enough attention or even 
succeeded in practice like e.g. automatic aerial 
triangulation. The reasons for this can be partly found 
below. 
huge expectations, 
2.1 Where are the users? 
One aspect is certainly the decline in economics for many cities 
and states. There are also the current problems of major players