In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
our five unknowns for a certain building i has been formulated as
follows:
em®<) = Y.
¿=1
(hrm + h T j (x) - ?f(x)) 2 dx (4)
Each prior that has been selected for a specific region is forced
to acquire such a geometry so as at every point its total height
matches the one from the available DEM. It’s a heavily con
strained formulation and thus robust. The introduced, here, recog
nition driven framework now takes the following form in respect
to (j), %, L and ©,:
Etotai — E seg ((j)) + [xE2d{4>i Ti, L) + ¡iE3i){Qi) (5)
2D Quantitative Measures
Completeness
Correctness
Quality
0.84
0.90
0.76
3D Quantitative Measures
Completeness
Correctness
Quality
0.86
0.86
0.77
Table 1 : Pixel- and Voxel-Based Quality Assessment
The energy term E seg addresses fusion in a natural way and
solves segmentation 0 in both J(x) and H(x) spaces. The term
E2D estimates which family of priors, i.e which 2D footprint i,
under any projective transformation T, best fit at each segment
(L). Finally, the energy E^d recovers the 3D geometry ©, of
every prior by estimating building’s h rn and h r heights.
4 QUALITATIVE AND QUANTITATIVE ASSESSMENT
OF THE PRODUCED 3D MODELS
The quality assessment of 3D data ((Meidow and Schuster, 2005),
gent et al., 2007) and their references therein) involves the assess
ment of both the geometry and topology of the model. During
our experiments the quantitative evaluation was performed based
on the 3D ground truth data which were derived from a man
ual digitization procedure. The standard quantitative measures of
Completeness (detection rate), Correctness (under-detection rate)
and Quality (a normalization between the previous two) were em
ployed. To this end, the quantitative assessment is divided into
two parts: Firstly, for the evaluation of the extracted 2D bound
aries i.e. the horizontal localization of the building footprints
(Figure 3) and secondly, for the evaluation of the hypsometric
differences i.e. the vertical differences between the extracted 3D
building and the ground truth (Figure 4).
In order to assess the horizontal accuracy of the extracted build
ing footprints the measures of Horizontal True Positives (HTP),
Horizontal False Positives (HFP) and Horizontal False Negatives
(HFN), were calculated.
voxels with an hypsometric difference with the ground truth, con
taining all the corresponding voxels from the HFP and the corre
sponding ones from the HTP (those with a higher altitude than the
ground truth). Respectively, the Vertical False Negatives are the
voxels with an hypsometric difference with the ground truth, con
taining all the corresponding voxels from the HFN and the corre
sponding ones from the HTP (those with a lower altitude than the
ground truth). To this end, the 3D quantitative assessment was
based on the measures of the 3D Completeness (detection rate),
3D Correctness (under-detection rate) and 3D Quality (a normal
ization between the previous two), which were calculated in the
(Sar-poiiowing way:
2D Completeness =
2D Correctness
area of correctly detected segments
area of the ground truth
HTP
HTP + HFN
area of correctly detected segments
area of all detected segments
HTP
2D Quality
HTP + HFP
HTP
HTP + HFP + HFN
Moreover, for the evaluation of the hypsometric differences be
tween the extracted buildings and the ground truth the measures
of Vertical True Positives (VTP), Vertical False Positives (VFP)
and Vertical False Negatives (VFN) were, also, calculated. The
VTP are the voxels among, the corresponding Horizontal True
Positive pixels, that have the same altitude with the ground truth.
Note that Horizontal True Positives may correspond (i) to voxels
with the same altitude as in the ground truth (VTP) and (ii) to
voxels with a lower or higher altitude than the ground truth (VFN
and VFP, respectively). Thus, the Vertical False Positives are the
3D Completeness =
3D Correctness =
VTP
VTP + VFN
VTP
VTP + VFP
VTP
3D Quality =
VTP + VFP + VFN
The developed algorithm has been applied to a number of scenes
where remote sensing data was available. The algorithm man
aged in all cases to accurately recover their footprint and over
come low-level misleading information due to shadows, occlu
sions, etc. In addition, despite the conflicting height similar
ity between the desired buildings, the surrounding trees and the
other objects the developed algorithm managed to robustly re
cover their 3D geometry as the appropriate priors were chosen
(Figure 1). This complex landscape contains a big variety of tex
ture patterns, more than 80 buildings of different types (detached
single family houses, industrial buildings, etc) and multiple other
objects of various classes. Two aerial images (with a ground res
olution of appx. 0.5m) and a the coarser digital surface model
(of appx. 1,0m ground resolution) were available. The robust
ness and functionality of the proposed method is illustrated, also,
on Figures 3 and 4, where one can, clearly, observe the Horizon
tal and the Vertical True Positives, respectively. The proposed
generic variational framework managed to accurately extract the
3D geometry of scene’s buildings, searching among various foot
print shapes and various roof types. The performed quantitative
evaluation reported an overall horizontal detection correctness of
90% and an overall horizontal detection completeness of 84%
(Table 1).
In Figure 4c, the hypsometric/vertical difference between the ex
tracted buildings and the ground truth is shown. With a red color
are the VFN voxels and with a green color the VFP ones. Sim
ilarly, at Figure 4c where the -corresponding among the HTP
pixels- VFN and VFP voxels are shown. The performed quan
titative evaluation reported an overall 3D completeness and cor
rectness of appx. 86% (Table 1).