141 
2 Fusion in Monocular Scene Analysis Systems 
Our work in man-made feature extraction from monocular 
views of a scene has used both edge-line intensity based 
techniques as well as shadow-analysis based techniques. 
BABE utilizes intensity edges to form comers, which then 
undergo structural analysis in order to generate plausible 
building hypotheses. These hypotheses are then evaluated in 
terms of size and line intensity constraints 3,4 . Figure 1 
shows a typical BABE result for a suburban area in 
Washington, D.C. 
SHADE is a building detection system based on a shadow 
analysis technique. SHADE utilizes a shadow intensity 
estimate generated by BABE to produce shadow regions, 
which are analyzed to locate shadow/building edges. These 
noisy edges are then smoothed and broken at corners by 
using an imperfect sequence finder 9 . The line segments that 
form nearly right-angled corners are joined, and the comers 
that are concave with respect to the sun are extended into 
parallelograms. Figure 2 shows a typical SHADE result. 
SHAVE is a system for verification of building hypotheses 
that examines the relationships between these hypotheses and 
the shadow regions in an image to rank the quality of these 
building hypotheses. SHAVE determines which segments of 
each building could cast shadows. Intensity walks are then 
performed for each pixel of these segments to delineate the 
cast shadows. Each segment is then scored based on the 
variance of shadow length along each segment. These scores 
can then be used to estimate the likelihood that a building 
hypothesis corresponds to a building, based on the extent to 
which it casts shadows. Figure 3 shows a typical SHAVE 
result. 
GROUPER is a system that utilizes shadows to cluster 
fragmented building hypotheses. GROUPER extends the 
shadow/building edges produced by SHADE along the sun 
illumination angle to form closed regions of interest in which 
man-made structures might occur. GROUPER intersects each 
building hypothesis with these regions of interest. 
Hypotheses that have sufficient areas of overlap with regions 
of interest are grouped together to form a composite building 
cluster. 
Figure 1: DC38008 industrial scene (smoothed) Figure 2: Histogram-splitting segmentation 
Figure 3: Refined S2 disparity map 
Figure 4: Extracted building regions