according to the number of line segments used to 
determine it, else it is introduced as a new point. No 
further action due to the discontinuity is taken at this 
point. In the consistency check following the parsing it 
will be investigated if the discontinuity can be 
removed. Some rules of the parsers require introduc- 
tion of line segments neither vertical, nor horizontal. 
In these cases arbitrary heights are introduced for later 
correction. 
When the parsing is initiated using the region 
segment parser, the region segments are matched to 
an aspect prototype and introduced into object space 
using the known labels of the line segments in the 
prototype, placing a ground line segment at an 
arbitrary datum and otherwise following the same 
procedure as for the line segment parser. The 
meaning with the region segment parser is therefore 
mainly a preassigning of certain labels to specific line 
segments in the input. 
5 PARSING THE SEGMENTATION 
OUTPUT 
Details of the segmentation output are to be 
interpreted as details in object space using a generic 
model for possible objects and a parser. This model 
describes the building in terms of roofs, walls, doors, 
windows and garden regions, all described by geomet- 
rically simple regions and relations between these. 
There is no reason to insist in a parser that succeeds 
with an appropriate analysis of any possibly 
incomplete segmentation of any possible building. If 
at all possible to construct, such a parser will be very 
complicated and probably quite slow. What is needed 
is instead a procedure which can tell when simple and 
fast parsers have failed. If the description of the object 
is accepted, a successful parse has been performed in a 
short time, if not, another parser is chosen. Using a set 
of parsers of moderate success rates will generate a 
compound procedure with a high success rate. The 
procedure described below is an example of such a set 
of parsers, designed to take care of the fact that images 
of horizontal line segments are not easily identified. 
The parse can be performed either using line 
segments or region segments. There are two 
advantages using line segments: First, the probability 
of a region segment being erroneous is high as one 
missing or misinterpreted line segment makes the 
whole region segment erroneous. Second, the fact that 
all vertical line segments point towards the nadir 
point gives a simple criterion for the interpretation of 
these line segments. There are, however, also 
disadvantages: Vertical line segments of objects close 
to the nadir point are short or missing in the 
segmentation output. Two different kinds of parsers 
are therefore used in the procedure below, which 
starts with the set of line segment parsers and 
continues with the region parser if the first procedure 
breaks down. 
5.1 Object recognition 
The input to the parsers consists of the segmentation 
output together with a window identifying what part 
of the segmentation should be interpreted and 
described as a building. The window is provided by 
the operator or given by a recognition procedure as for 
instance the following one: 
A simple scan of the segmentation represented as 
region segments gives locations where the group 
index (see section 3.2) is minimum. These locations 
are used as probable locations of the objects looked for. 
In the case of buildings, the interest index was chosen 
so that building roofs should tend to give minimum 
group indices. The reference point of the segment, 
together with a window size determined from neigh- 
bouring object indices, determines a window to adjoin 
the segmentation presented as input to the parser. 
5.2 A generic model for buildings 
Complex objects like buildings are difficult to model 
using specific models. Although it is possible in 
principle to cover the possible logical structures using 
parameters, this is impractical. An alternative 
approach using generic models has been used by 
several investigators, e.g. by (Fua and Hanson, 1988). 
Here, a generic model is used, defining buildings as 
objects present in object space and consisting of sub- 
objects with certain properties and relations. A sketch 
of this model is given below, the details of which are 
given implicitly in the parsers in sections 5.3 and 5.4. 
buildings: connected set of roofs, walls, floors, windows, doors, 
chimneys 
properties: 
ground: set of nonbuildings 
properties: 
roofs: geometry: polygon(closed, plane, nonvertical) 
neighbours: roofs, walls 
properties: a neighbour at a sloping border is a 
roof 
walls: geometry: polygon(closed, plane, vertical) 
neighbours: ^ roofs, walls, ground 
properties: a neighbour at a horizontal border 
is not a wall 
floors: geometry: polygon(closed, plane, horizontal) 
neighbours: walls 
properties: at least one neighbour is beneath 
the floor 
windows: geometry: polygon(closed, plane, vertical) 
neighbours: walls 
properties: one surrounding neighbour 
doors: geometry: polygon(closed, plane, vertical) 
neighbours: ^ walls, roofs 
properties: ^ one semi-surrounding neighbour 
chimneys: vertical structure, part of roof 
polygons: connected set of line segments 
line segments: 2 points 
points: (x,y,z) 
Table 1. Sketch displaying principle of generic model 
for buildings 
5.3 The parser for line segments 
The most conspicuous properties of images of 
buildings are that they can be described using line 
drawings, where some of the line segments represent 
vertical or horizontal lines. A natural approach when 
parsing a segmentation including buildings is then to 
identify vertical line segments as those pointing at the 
730 
ra ra 
A d "P o6(C om nn bie mie fd A) RI 
ES (QU Mo m. 
4) 
5)