SCALE-DEPENDENT ADAPTATION OF IMAGE ANALYSIS MODELS 
INCORPORATING LOCAL CONTEXT OBJECTS 
Janet Heuwold, Kian Pakzad, Christian Heipke 
Institute of Photogrammetry and Geolnformation, Leibniz Universität Hannover, 
Nienburger Str. 1, D-30167 Hannover, Germany - (heuwold, pakzad, heipke)@ipi.uni-hannover.de 
Commission III, WG III/4 
KEY WORDS: Multiresolution, Scale Space, Modelling, Image Analysis, Interpretation, Urban, Method 
ABSTRACT: 
Local context objects have significant impact on object extraction from real aerial or satellite images, since occlusions or shadows 
can substantially hinder a successful extraction of a particular landscape object one is interested in. This paper presents an adaptation 
concept for image analysis object models considering local context objects to a lower image resolution. The scale-dependent 
adaptation of local context poses a severe problem for an unambiguous scale behaviour prediction, as the exact position of local 
context objects is generally unknown. However, by adapting the object models for the landscape object of interest and the local 
context objects separately, the influence of this contradiction is avoided while the loss of the exact combined scale behaviour 
prediction remains acceptable. An example for road extraction with vehicle as local context object demonstrates the capability of the 
adaptation approach. To conclude, the assumptions for the adaptation methods and its limitations are discussed. 
1. INTRODUCTION 
A main prerequisite of knowledge-based image analysis from 
aerial and satellite data is the availability of suitable models for 
the extraction of relevant landscape objects. In the last decade 
more and more sensors from space with very high spatial 
resolution capabilities became available. The range of available 
image resolution, thus, has been enlarged. An image scene 
changes with different resolution: area-type features change to 
lines or points, some details may get lost or small features 
disappear, while other features may become more dominant and 
important for the extraction of a particular object. As landscape 
objects change their appearance in images of different image 
resolutions, models for the extraction of landscape objects are in 
general not applicable to another image resolution. Automatic 
object extraction can require substantially different models for 
different image resolutions, because of the large amount of 
various details that are not visible or have merged with adjacent 
objects in coarser image resolution. Therefore, the need for 
different models for object extraction in different image 
resolutions is growing as well. 
Furthermore, the extraction process of a landscape object may 
be severely influenced by local context objects, i.e. other near 
by objects in the vicinity occluding the landscape object one is 
interested in. For example, roads are often partly obstructed by 
vehicles, trees, buildings or their shadows (cf. Figure 1). Then, 
the extraction of the object of interest can fail, since some object 
parts cannot be completely extracted due to the influence of 
local context objects. Hence, modelling of a landscape object 
only by its inherent features is often not sufficient for reliable 
extraction results, but requires the incorporation of local context 
objects. 
A methodology for the automatic adaptation of image analysis 
object models for 2D landscape objects represented in form of 
semantic nets was developed previously (Heuwold et al., 2007) 
using linear scale-space theory. This procedure is extended here 
to incorporate additional local context objects. The context 
objects are modelled as a semantic net as well with attributes 
describing their appearance and their specific feature extraction 
operators. While from the object models treated so far all parts 
of the landscape object of interest and their relative position was 
known, the presence and location of local context objects is 
usually not. The location of local context objects can be 
extremely variable and can mostly only be determined roughly, 
if at all, in the image analysis model. This variability is 
contradictory to a non-ambiguous scale behaviour prediction 
based on position and needs to be accounted for in the 
adaptation method. The extension of the existing method for the 
scale-dependent adaptation of object models towards image 
analysis considering additional local context objects is the topic 
of this paper. Due to the high relevance for database update in 
urban areas, our study focuses on road extraction with vehicles 
as local context objects. However, the presented adaptation 
concept enables the application to other landscape objects as 
well. 
The remainder of this paper is organized as follows: The next 
section reviews related work concerning the automatic 
adaptation of image analysis object models and the 
incorporation of local context into object extraction. Section 3 
derives the developed concept for the adaptation of object 
models considering local context, while section 4 gives an 
example for the adaptation of a road model incorporating 
vehicles as local context. At the end of this paper, the results of 
the example are discussed and conclusions are drawn.