The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
437 
lying relatively close together a scale event may occur for a 
certain resolution reduction AR, another object constellation of 
A and B with larger distance between each other may experience 
a different scale event or no scale event at all for the same 
change of resolution AR. Thus, the scale behaviour of a 
combination of objects strongly depends on the relative position 
of the objects, making an unambiguous prediction of the scale 
behaviour for local context objects with the primary object 
impossible. The result of the scale-dependent adaptation would 
be a number of different concept nets in the target resolution - 
one for each constellation. Hence, for a complete scale 
behaviour analysis an explosion in the number of target 
resolution object models can occur, making it unusable for real- 
world applications. 
Nevertheless, several solutions regarding scale behaviour 
prediction for this variability problem are possible: 
separated models for the primary object (road in our example) 
and one for each context object are assumed. 
The connection between the primary object model and the 
context model(s) is defined by the relation optional, i.e. the 
local context object may be but does not have to be present. The 
context object can occlude or cast a shadow on object parts of 
the primary object, represented by the context relations occludes 
and casts shadow on. 
Additionally, the incorporation of more than one context object 
is possible. The priority of several context objects for the 
extraction process can be defined by an ordering scheme for the 
individual context objects with their relations to the primary 
model. Thereby, a sequence for applying the context object 
models to the image for refining the extraction result is defined. 
3.3 Extraction strategy 
1. Highest probable constellations approach: The scale 
behaviour of the local context objects together with the 
primary object is predicted only for certain (a few) 
constellations with highest probability. The benefit is the 
exact prediction of the scale behaviour under 
consideration of the surrounding primary object, but the 
variability of the position of local context objects is not 
fully considered in the adaptation. This location constraint 
poses a contradiction to the inherent function of local 
context objects, as their occurrence and their resulting 
implications on the primary object are naturally extremely 
variable by their definition. 
2. Statistics approach: The probability functions for the 
position of the individual local context objects are 
modelled from experience or derived by statistical 
investigations from training data. The scale behaviour for 
each constellation of objects is then predicted with the 
respective probability for their occurrence. The result is a 
set of concept nets in target resolution with statistical 
information. 
3. Separation approach: The scale behaviour of the primary 
object and the local context objects is analysed separately. 
Thereby, the variability of the context object position is 
preserved on the cost of an exact combined scale 
behaviour prediction, because no embedding into the 
primary object takes place during the scale change 
prediction. The relation to the primary object is thereby 
partially neglected. However, most image analysis 
algorithms for context object extraction have to work in 
diverse surroundings due to the context object location 
variability making the consideration of all details of the 
primary object impossible. Thus, in most cases this 
approach will be sufficient for the adaptation of image 
analysis models for primary objects and context objects to 
a lower resolution. 
Our adaptation concept uses the separation approach because of 
the advantage of preserving the flexibility of the context relation 
in the model, which we consider as a crucial aspect for realistic 
object extraction. Although the statistics approach represents a 
more exact solution, the analysis is not straightforward and its 
research effort would not be in accordance with our current 
resources. 
3.2 Context Modelling 
In our adaptation approach the landscape object one is 
interested in (in the following denoted as primary object) and 
the local context objects are regarded as separate objects. The 
local context objects are not incorporated into one single joint 
object model for the primary object and context objects, but 
Figure 2. Extraction strategy with local context objects 
It is assumed for our adaptation approach that the strategy for 
the extraction of a landscape object considering local context 
follows a two-step process. Figure 2 depicts the extraction 
strategy with local context objects. At first, the extraction of the 
primary object with its respective object model without 
considering local context is carried out. If this extraction partly 
fails, then we try to explain the emerged gap in the extraction 
result. The reason for the failure can be a local context object. 
Thus, an extraction of the local context using the corresponding 
object model(s) is carried out at the image area, where the gap 
occurred. If a local context object is found in the extraction gap 
area, the primary object is regarded as completely existent at 
this location, since the extraction failure can be explained by 
local context. If none of the context objects is verified in the gap, 
the extraction of the primary object remains disconnected in that 
area.