The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
438 
Local context objects are employed here as a means for gap 
filling in the extraction result for the primary object. The only 
allowed context relation is optional. Simple modelling for the 
context objects itself is in most cases sufficient, but a more 
detailed model, such as the one of Hinz (2004), for the context 
objects can also be incorporated into the extraction strategy, if 
necessary. 
3.4 Adaptation method 
The adaptation to the target resolution is carried out for the 
primary object model and the local context object model(s) 
independently applying the adaptation algorithm presented in 
(Heuwold et al., 2007), cf. section 2.1. 
Figure 3. Example image showing a road with local context 
The scale behaviour prediction is carried out using analysis-by- 
synthesis in the scale change analysis stage of the adaptation 
algorithm. This prediction encompasses all four possibly 
occurring scale events during resolution change and the 
attributes for the object parts as defined in the initial object 
model. Scale events are predicted by comparing blob support 
regions and extrema in initial and target resolution. Occurring 
scale events can have a severe impact on the resulting net, as 
they dictate the number of remaining object parts in target 
resolution. The object part attributes for shape and radiometry 
are predicted from the characteristics of the blob support 
regions in the target resolution. The scale behaviour prediction 
of attributes is essential for the adaptation of feature extraction 
operators, since the predicted attributes can serve as adapted 
operator parameters. As pointed out in section 2, feature 
extraction operators for local context objects usually rely on 
shape properties as well as radiometric characteristics. The use 
of colour implies the scale behaviour analysis of more than one 
intensity channel for the image. However, in the analysis-by- 
synthesis a simulation of the scene with maximum contrast for a 
single channel is sufficient, if no restriction for colour is implied 
by the model. It is irrelevant in which channel the maximum 
blob contrast occurs, but the maximum contrast complies with 
maximum existence in scale space. The maximum contrast 
object might still exist in coarser scale, while an object with less 
contrast might not be extractable any more. Note however, that 
this approach is only permitted for objects with optional context 
relation. 
As long as the context object is still predicted to be extractable 
in the target resolution, the type of context relation between 
primary object and context objects remains unchanged. 
4. ADAPTATION EXAMPLE 
This section gives an example for the adaptation of a high- 
resolution object model for a road (dual carriageway), which 
incorporates freight vehicles (trucks) as possible local context. 
The adaptation is demonstrated for an initial image resolution 
R 0 =0.03m to a target resolution /?,=0.20m. The object models 
are applied to a greyscale aerial image captured from Hanover 
region, Germany (see Figure 3). 
4.1 Model 
The road model including the context relations for the initial 
high resolution /?o=0.03m is given in Figure 4. The road model 
corresponds to the model in Heuwold (2006). The relation to 
local context objects is given by optional occlusion or optional 
shadow. As all road markings have a relation to local context, 
they all can be subject to occlusion or shadow by a vehicle. 
Lfciiansl: 
<>■■■■■■■ Part-«! 
Figure 4. Road object model with context relations for 
0.03m/pixel image resolution 
The initial object model for the vehicle, depicted in Figure 5, is 
relatively simple and not as detailed as e.g. the one of Hinz 
(2004), but despite its simplicity it should suffice for 
demonstrating the capability of the adaptation concept. A large 
freight vehicle is modelled comprising two object parts as 
rectangles for front and trailer. The attribute values are given in 
pixels. 
Legend: 
Front 
[dJ=20-30] 
► 
Trailer 
Rectangle 
Rectangle 
wid=65-85 
wid=65-85 
len=52-67 
len=100-535 
gv=1-256 
gv=1-256 
Part-of 
Spatial relation with 
parallel [d ( =] and 
perpendicular 
distance [di=] 
Figure 5. Freight vehicle object model for Æ^O-OSm/pixel 
The road model uses the operators for road markings and 
roadway, which are described in detail by Heuwold (2006). The 
operator for the vehicle parts uses region growing for 
segmentation. Suitable region candidates are selected by shape 
(convexity, compactness, excentricity, area) and radiometric 
properties (Meyer, 2006). The extraction results of the road 
model applied to the example image in high resolution yields an 
extraction gap for the road marking lane marking left at the 
context area. Figure 6 displays the results of the road extraction 
and of the vehicle extraction.