The Monocular Automated Building System (MABS)
Our work is based on USC-MABS: it requires only one image with tilt
angle, swing angle and sun direction as input and assumes weak
perspective. The building model which is used here consists of
rectangular parallelepipeds that can be composed to, for example, L,T
or I-shapes. After a linear
feature extraction roof-hypotheses are formed by an hierarchical
perceptual grouping (lines, folded lines, parallels, U-contours,
parallelograms). This step creates a large number of parallelograms,
allowing the establishment of also weak hypotheses. Next the
parallelograms are selected by local and global criteria using domain
knowledge and shadow- and wall information. Finally a verification is applied
to validate the selected
hypotheses using shadow and wall evidence. This step results in a
three-dimensional model
of the building, which is determined by the associated shadow and wall
evidence.
Click on the boxes below to get example images.
The system performs reliably on images where the buildings are
separated, wall-, shadow- or roof-lines are at least partially well
detected. Otherwise, it may fail to derive a three-dimensional
description of a building, even if there is information extracted that
at least partly contain the desired result. A rejection of a correct
hypothesis can occur during verification or during selection. A
correction here would just be to specify this particular
parallelogram. It is also possible that the correct hypothesis is not
generated due to lack of existing features in the image. Still partly
correct hypotheses are likely to exist and the missing feature(s) have
to be added to obtain a correct description of a building.