When the parsing has been initialized according to 
these rules, the rest of the parse is continued using the 
line segment parser and starting at rule 6). 
5.5 Alternation between parsers 
In the procedure described above, the parsing begins 
with trying a line segment parser. If the parse 
produced is unsuccessful, other versions of the same 
parser are used trying the various strategies in rule 4. 
If no parse passed the consistency test, the region 
segment parser is used. If also this attempt is a failure, 
the user must be asked for help. 
As images of buildings never display the complete 
structure, the consistency test will be performed in 
such a way that also partially parsed buildings are 
accepted. A typical result of parsing the segmentation 
of a single image is therefore a partial building 
structure in object space plus a set of line segments in 
the segmentation, where the parsing has failed. These 
remaining line segments might belong to the building 
or they might describe the ground surrounding it. 
In order to take care of the line segments which 
belong to the building, the complete parsing 
procedure is started again, now skipping the line 
segments already included in object space. A new 
vertical line segment connected to horizontal line 
segments is selected according to rule 5), the sequence 
of line segment parsers and region segment parser 
being carried through until no more line segments are 
parsed. Again, this complete procedure is repeated 
until no more line segments are parsed. In this way 
weakly connected parts of the same building, and also 
different buildings contained in the window to be 
parsed, are introduced into object space at 
approximately correct locations. 
Remaining line segments or polygons, which can be 
connected to a ground point of the reconstructed 
building are assumed to be ground line segments in a 
horizontal environment. They can be introduced 
using the transformation (1). If there still are line seg- 
ments left, they are unconnected to anything parsed. 
5.6 Consistency check for buildings in object space 
The general procedure given above for interpretation 
of the input segmentation is to use a series of 
moderately successful parsers and terminate the 
procedure when a consistency check indicates an 
acceptable object description. This approach rests on 
the assumption that it probably is easier to design a 
high quality consistency check than a well performing 
parser. The consistency check is designed so as to 
check all rules contained in a generic model. As a 
consequence, this procedure is an interpretation 
process. Substructures, in the form of 2-D regions in 
object space, which have passed the consistency test, 
are ascribed labels (wall, roof, etc). 
As the parse to be checked usually will contain 
inconsistencies which can be removed, the 
consistency check should include procedures for such 
removals. Examples are discontinuities due to 
erroneous parses of individual line segments and also 
missing lines which sometimes must be introduced in 
order that the the 2-D regions in object space be plane. 
5.7 Introduction of missing lines 
Check that the mappings in object space of all parsed 
region segments are planar. Denoting the points r, = 
GGoyiZk), k=1,...,n in the associated closed polygon, 
check that the following determinants are zero: 
Xk yk zx 1 
XU] £dd 4 : 
Qn cul" 0; k=4.n 
X9 1y3 73-4 
Segments not satisfying this requirement are assumed 
to be composed of several segments divided by line 
segments missing in the segmentation. Subsegments 
are produced and tested for planarity by introducing 
line segments Ly; = (Ty, fj ), kz1, j23,...,n-1; kz2,...,n-2, 
j=k+2,...,n. When a planar subsegment has been found 
the line segment is introduced into object space. The 
remaining subsegment is again tested for planarity, 
continuing the procedure until no segment is left. 
5.8 Interpretation 
After checking that all polygons are closed and planar, 
they are labelled walls, roofs, floors, windows, etc. 
according to the properties of these subobjects as 
specified by the generic model. This amounts to 
checking surface orientation and neighbours. 
* Starting with identifying all vertical region seg- 
ments, these are checked for generating horizon- 
tally connected structures. If unconnected vertical 
boundaries exist, this is due to effects of perspective 
giving rise to hidden regions. All region segments 
participating in such structures are labelled walls. 
* Sloping regions above walls are labelled roofs. 
* Horizontal regions above walls are labelled floors, 
below walls are labelled ground. 
* Unclassified vertical regions located in the interior 
of walls or roofs are interpreted as doors or 
windows, depending on them sharing the bottom 
line segment with the parent region or not. 
5.9 Miscellaneous tests 
Besides the rules given in the generic model, certain 
criteria must be fulfilled due to the projection under 
which the image was made. Also other tests can be 
performed in order to ensure a correct parse. Examples 
of such rules are that all line segments from the given 
segmentation are visible, i.e. as seen from the sensor; 
they cannot lie behind surfaces generated by other line 
segments. Also no line segments can cut through 
building surfaces of any kind. 
5.10 The decision of acceptance 
After running the consistency check on a given parse, 
the final decision of accepting the parse or not has to 
be made. Local failures can be accepted as well as local 
errors, which are just removed. The kind of parsing 
errors which should lead to rejection are global: e.g. 
when the line segment parser chooses to start with a 
line pointing at nadir and this line in fact is non- 
vertical. Criteria for this and other fatal mistakes are 
yet to come. 
333