Proceedings of 2011 NSF Engineering Research and Innovation Conference, Atlanta, Georgia Grant #CMMI-0856558
Figure 5. Context-based modeling results on some challenging
cases. The algorithm correctly distinguishes items that look
significantly like clutter, such as large cabinets (a-b) and
bookcases (c-d). Reflectance images are shown in the left
column, and the corresponding patches are shown in the right
column. Green patches are labeled as clutter, Blue patches are
walls, and yellow patches are floors.
encode information about a patch’s neighborhood
configuration. We investigated various relationships,
including orthogonal, parallel, adjacent, and coplanar.
We use a machine learning model known as conditional
random fields (CRF) to combine these two types of
features in an optimization framework. By maximizing
the likelihood of the labels assigned to each patch (e.g.,
wall, ceiling, floor, or clutter) given the local and
contextual feature values, the algorithm finds the
optimal labeling for the patches, taking into
consideration all of the labelings simultaneously. For
example, a patch that is coplanar with other patches that
are likely to be a wall will be more likely to be a wall as
well.
We conducted experiments using data from 26 rooms of
a school building that was professionally scanned and
modeled. We divided the data into training, validation,
and test sets and then evaluated the performance of the
algorithm at labeling the detected planar patches. We
found that the use of context does improve the
performance of the recognition process. We compared
the context-based algorithm to one that just uses local
features, and we found that the context-based algorithm
improved performance from 84% accuracy to 89%
(Figure 5). The main failures of the algorithm are in
unusual situations, such as the tops of the interiors of
short closets (which are considered ceilings in the
model). Our approach is effective at distinguishing
clutter from surfaces of interest, even in highly
cluttered environments. Finally, we found that the
coplanar relationship is very helpful for addressing
fragmentation of wall surfaces due to occlusions and
large window and door openings.
4. Algorithms for detailed modeling of planar
surfaces. One of the goals of our research is to
explicitly reason about occlusions in the data in order to
avoid problems caused by missing data. Most previous
work on modeling building interiors focuses on
environments with little or no clutter. One reason for
this is that clutter causes occlusions, which makes
modeling the surfaces of interest more difficult. For
automated creation of BIMs to be useful in real-world
environments, algorithms need to be able to handle
situations with large amount of clutter, along with the
resulting occlusions. For example, it is not practical to
move all furniture out of a building before scanning it
for creating an as-built BIM.
In this research, our goal is to model wall surfaces at a
detailed level, to identify and model openings, such as
windows and doorways, and to fill occluded surface
regions. Our approach utilizes 3D data from a laser
scanner operating from one or more locations within a
room. Although we focus on wall modeling, the method
can be applied to the easier case floors and ceilings as
well. The method consists of four main steps (Figure 6):
1) Wall detection – The approximate planes of the
walls, ceiling, and floor are detected using projections
into 2D followed by a Hough transform. 2) Occlusion
labeling – For each wall surface, ray-tracing is used to
determine which surface regions are sensed, which are
occluded, and which are empty space. 3) Opening
detection – A learning-based method is used to
recognize and model openings in the surface based on
the occlusion labeling. 4) Occlusion reconstruction –
Occluded regions not within an opening are
reconstructed using a hole-filling algorithm.
The primary contribution of this research is the overall
approach, which focuses on addressing the problem of
clutter and occlusions and explicitly reasons about the
missing information. Our approach is unique in that it
distinguishes between missing data from occlusion
versus missing data in an opening in the wall. Secondly,
we propose a learning-based method for detecting and
modeling openings and distinguishing them from
similarly shaped occluded regions. Finally, we propose
and use methods for objectively evaluating
reconstruction accuracy, whereas previous façade
modeling work has focused on primarily on subjective
visual quality.