417 
ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS’’, Bangkok, May 23-25, 2001 
fl 
thematic and physical characteristics, these aspects are not 
discussed here. Since the rendering system visualises virtual 
objects (real objects are considered in case of occluding), colour 
and texture of real objects are not of interest and are not to be 
maintained. More details related to physical properties of a 
spatial object can be found in [15]. 
The new element in the 3D topological model is the organisation 
of the data for positioning. The line features are encapsulated 
with their co-ordinates and stored as a separate data set, i.e. 
lines. Each line is considered as a strait line represented by two 
sets of co-ordinates. 
The IFO semantic data model is used to represent the spatial 
relationships between all the objects. The motivation is based on 
the fact that the IFO semantic model uses object-oriented 
notations to represent objects and their relationships and 
provides a mechanism to map them into a relational model. 
Three kinds of object-types (abstract, printable and derived) are 
distinguished by the model (see [1]). An abstract type is an 
object that cannot be shown printed, mapped, etc., e.g. a 
building, a street. A printable type of object is an object that can 
be printed, shown, etc. e.g. co-ordinates of a building or a street. 
The objects that are composed of some printable objects or 
other atomic objects are derivable. From atomic (printable or 
derivable) objects complex objects can be derived by 
aggregation and grouping. The principle difference is that 
aggregated objects contain elements from different types, while 
grouped objects contain objects only from one type. IS-A 
relationships are utilised to define sub-types (specialisation) and 
super-types (generalisation) of objects. Figure 5 shows the 
graphical notations for the objects, relationships and principles 
of construction. 
a) types o» objects: 
o- C3 i i printable 
b) functional relationships: 
h»(1:m) has (1:1) is-a 
Figure 5: Notations of IFO model 
Applying these graphical notations we obtain the schema of the 
model shown on Figure 6. The four abstract objects named 
geometric objects (GO), one explicit spatial relationship (GR) 
and two constructive objects (CnsO) attempt at representation of 
all the objects. 
Figure 6: The proposed 3D topological model 
The topological model is similar to the ones presented in [3], [5] 
and [9] but differs in number of construction elements, i.e. only 
two. The 1 D-cell, often called arc or edge (see [5], [9], [10]), is 
omitted. The arc in 2D space have the unique feature of defining 
1:2 relationships with faces and nodes, i.e. an arc has two 
neighbouring faces and nodes. This feature is only partially true 
in 3D and therefore the explicit storage of arcs does not bring a 
significant facilitation. Such representation, i.e. without arcs, has 
shown speed acceleration in many data retrieval operations (see 
[15]). Moreover the computational complexity of writing the 
VRML file is largely reduced, since the representations of 
polyhedron and surface are similar to description of irregular 
shapes in the VRML node IndexedFaceSet (see [14]). 
Figure 7: Modelling of complex spatial objects 
4 DATA COLLECTION 
The 3D reconstruction of urban areas is still a rather complex 
process involving quite a lot of time and manual interactions. 
Several different methods are applied to provide data with the 
appropriate accuracy and resolution and to obtain the 3D 
topology. The utilised methods can be subdivided into three 
major groups: manual, semi-automatic and automatic. Whilst 
manual and semi-automatic methods are used in the case of 
relatively limited amounts of object elements with high accuracy 
and 3D topology are required (i.e. to build the topological 
model), automatic methods are applied for collection of many 
elements with non-topological organisation (i.e. the line 
features). Furthermore objects (e.g. buildings) with complex 
shape and variety of details (balconies, overhanging roofs) are 
modelled manually from images taken at street level. The 
images are processed in a 3D modelling software (i.e. 
PhotoModellei) that ensures accuracy and resolution within 
required values. A snapshot from CosmoPlayer shows some of 
the reconstructed objects of this group (see Figure 7). 
Figure 8: Modelling of buildings without overhanging roofs 
Simpler objects such as buildings with flat or gable roofs are 
reconstructed applying semi-automatic methods. For example 
the reconstruction of buildings with flat roofs and roof outlines 
that can be projected onto the footprints (i.e. lack of overhanging 
roofs) are reconstructed by a procedure utilising existing DTM 
and manually digitised roof outlines (see Figure 8). The walls 
and the corresponding 3D topology are obtained fully 
automatically (see [16]). Terrain objects as streets, bicycle or 
c) constructs: 
/ X) ND 
grouping 
colouA texture