different scenes. One is an outdoor scene of building facades
(see Figure 6a), which is the terrestrial laser scanning point
cloud from the large-scale point cloud classification benchmark
datasets published by ETH Zurich (Hackel
et al.
, 2017). The
test area of this scene is approximately 700 m
2
. The other scene
represents a construction site (see Figure 6b) located in the
downtown area of Munich, Germany, with both laser scanning
and photogrammetric point clouds (see Figures 6c and 6d) ac-
quired. Its testing area is around 320 m
2
, including foundation
pits, ground objects, wall surfaces, equipment, etc. The ter-
restrial lidar point cloud was acquired using a Leica HDS 7000
scanner, whereas the photogrammetric point cloud was created
from a structure from motion (
SfM
) system and multi-view
stereo matching method (Tuttas
et al.
, 2017), using a Nikon
D3 DLSR camera with 105 images. In Figure 7, we provide an
illustration of the configuration of images for generating the
photogrammetric point cloud of our test scene, involving 43
images. Scanning positions of the laser scanner are also given
in Figure 7. Moreover, prior to the major processing, statistical
outlier removal filtering (Rusu and Cousins, 2011) was applied
to these point clouds. The lidar and photogrammetric point
clouds are both downsampled to around nine million points.
Evaluation Metrics
For evaluating the performance of the proposed methods, four
representative segmentation algorithms, including the Euclid-
ean distance and difference of normal (
DON
) based clustering
(Ioannou
et al.
, 2012) method, the smoothness based region
growing (
RG
) (Rabbani
et al.
, 2006), and the Locally Convex
Connected Patches (
LCCP
) technique (Stein
et al.
, 2014) are
used as baseline methods. Here, the
RG
and
DON
methods are
renowned and widely-used point-based segmentation algo-
rithms, while the
LCCP
method is a popular supervoxel-based
segmentation method, which also adopts the voxel structure
and uses the convexity as the segmentation criterion. In
experiments, all methods are implemented in C++ and run on
an Intel i7-4710MQ CPU @ 2.5GHz and with 16.0 GB RAM.
All baseline methods are implemented as part of
PCL
(Rusu
and Cousins, 2011).
The quantitative evaluation is conducted by comparing the
obtained segments against the manually segmented ground
truth (see Figures 8f through 8j) using approaches described
in (Awrangjeb and Fraser, 2014) and (Vo
et al.
, 2015). Spe-
cifically, in the evaluation process, a pair of segments (one
is from the obtained segmentation result and the other one
is from the ground truth dataset) is found and considered
as having correspondence, if the points of them have the
largest overlap when compared with other possible ground
truth segments. Once a pair of segments with correspon-
dence was found, the numbers of overlapping points and
non-overlapping points between the obtained segment and
the ground truth segment are computed by checking whether
the positions of points match or not. The true positive (
TP
),
the false negative (
FN
), and the false positive (FP) values are
computed from the numbers of overlapping and non-overlap-
ping points. Then, three standard metrics, Precision, Recall,
and F1 measure, are calculated using
TP
,
FN
, and FP values,
and then used to assess the quality of segmentation (Xu
et
al.
, 2017b). Here, three sample areas are selected from two
scenes and manually segmented as ground truth (see areas in
dashed boxes in Figures 6a and 6b). Note that for the scene of
the construction site, ground truth is sampled from both laser
scanning and photogrammetric point clouds. The manual
segmentation follows the rule that each segment should cor-
respond to a semantic object of building components. For
instance, planar segments represent wall or ground surfaces
in the scene, while line segments are related to frames of win-
dows in the facade. The number of segments in ground truth
datasets is listed in Table 1.
Table 1. Number of segments in ground truth datasets.
Ground truth
Number of segments
Laser scanned
Photogrammetric
Sample 1
33
—
Sample 2
66
37
Sample 3
74
50
Figure 5. Aggregation of supervoxels: (a) Definition of local contextual area of supervoxel V, (b) Local contextual graphs of
V
i
and
V
j
. (c) Partition of local adjacency graphs, (d) Determination of connection relations of supervoxels, and (e) Connection
based clustering.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2018
383
