Nevatia - 6
set, say a pair of features, to which we add more features that are consistent with the given
constraints. Thus, it may be possible to limit the complexity to be only linearly dependent on
number of features, m, (i.e. 0(m)), in the image.
Use of constraints from geometry is a key component in this approach to reducing
complexity. The stronger the relations between desired features, the fewer candidates we will
generate. Knowledge of viewing geometry, such as that of vanishing points (or the direction of
verticals for parallel projection) can be very important. In our implementation for building
detection, we can not, in general, estimate the orientation of the sides of the building a priori ;
however, given one side, the viewing geometry allows us to constraint the direction of the other
side (assuming a flat roof). Fig. 6 shows the hypotheses generated from the image of Fig. 1(a) by
the USC building finder described in [6]. The number of hypothesized parallelograms is 2339,
while the number of segments in Fig. 1(b) is 9344. Even though 2339 may appear to be a large
number, note that it is not anywhere as large as 9344 4 .
Figure 6. Roof hypotheses from Figure 2(b)
b) Validity : A perceptual grouping system must make many subtle distinctions at every stage of
processing. Observed features in an image are fragmented due to characteristics of the images and
of the feature detectors. Their may be unmodelled distortions and noise present. Thus, each stage
of grouping must allow for some “tolerances” from expected properties. Also, we are likely to
generate many alternative hypotheses for explaining the observed evidence and must select
among them. This process can be difficult and the right set of rules and parameters is often chosen
by a trial and error process. Our approach has been to not discard too many hypotheses at an early
stage but rather to make the distinctions between likely hypotheses only when sufficient data is
available to make a reliable distinction.