isabella toschi, fabio remondino - eurosdr · 2015-10-28 · isabella toschi, fabio remondino 3d...
TRANSCRIPT
Isabella Toschi, Fabio Remondino
3D Optical Metrology (3DOM),
Bruno Kessler
Foundation (FBK),
Trento, Italy
EuroSDR/ISPRS Workshop on
Oblique Cameras and Dense Image Matching
19-20 October 2015
Southampton, UK
to demonstrate the behavior of different oblique imaging platforms in real-world
conditions and study the influence of the additional slanted cameras on:
The quality of 3D point triangulation is
evaluated under several imaging
configurations, by assessing
• accuracy/precision of 3D points;
• block deformations in object space.
• The completeness of the
reconstructed 3D scene is studied in
relation to the camera tilt angle and
the image overlap.
• Point cloud generation and filtering
experiences are reported.
• Two main approaches of 3D building
modelling are discussed taking into
account the contribution of oblique
imagery for façade reconstruction.
• 3D reconstruction experiences are
reported.
MIDAS 5 UltraCam Osprey I
Nadir Oblique 45° Nadir Oblique 45°
Sensor size (mm) 36 x 24
(Canon EOS 1Ds Mk3) 70 x 45
23.5 x 36 (L/R)
71.5 x 23.5 (F/B)
Focal length (mm) 80 100 51 80
FOV along/across (°) ~17/25 ~13.5/20.5 ~48/70 ~17/25.5 (L/R)
~51/17 (F/B)
MILAN (MIDAS 5) GRAZ (UltraCam Osprey I)
GSD (*) (m) ~0.10 ~0.12
Y/X Overlap (*) (%) 70/30 75/65
# Images Nadir/Oblique 125/375 20/160
(*) Computed on nadir images
MIDAS 5
Forward and backward oblique views
Side oblique views
along-track overlap
across-track overlap
UltraCam Osprey I
Forward and backward oblique views
Forward and backward oblique views
along-track overlap
across-track overlap
Side oblique views
MILAN (3 km by 5 km)
Camera network (0) (1) (2) (3) (4)
GRAZ (3 km by 1.5 km)
Camera network (1) (2) (3) (4)
GCP
CP
σGCP, CP
~ 0.05 m
Rupnik, E., Nex, F. and Remondino, F., 2013. Automatic orientation of large blocks of oblique
images. Int. Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,
40(1/W1), pp. 299-304.
ID
Constant/
Unknown/
Observ.
Input
images EOR IOR+AP
GCPs,
observed
in nadir
GCPs,
observed
in oblique
Tie
points
1FX
C
Nadir
U
O
2FX
C Nadir +
oblique
U
O
3FX
C Nadir +
oblique
U
O
1FR
C
Nadir
U
O
2FR
C Nadir +
oblique
U
O
3FR
C Nadir +
oblique
U
O
• FR – free camera parameters (IOR + AP, Fraser calibration model)
• FX – fixed camera parameters (IOR + AP)
• 1 to 3 – GCP configurations
• Each test is performed both with and without GNSS information
Rupnik, E., Nex, F., Toschi, I., Remondino, F., 2015. Aerial multi-camera systems: Accuracy and
block triangulation issues. ISPRS Journal of Photogrammetry and Remote Sensing, Vol. 101, pp.
233–246
GCPs config. 1FX 2FX 3FX 1FR 2FR 3FR Dataset
0 - 4 0.49 0.47 0.55 0.47 0.45 0.52 Milan
1 - 4 0.38 0.36 0.39 0.31 0.29 0.31 Graz
Internal quality (precision) of the adjustment process:
mean re-projection error [pixel]
Take-away messages:
• From scenario to scenario, σ0 remains near constant due to the applied
robust weighting function (Apero).
• Introducing oblique images into the network, but allowing GCP image
measurements only in nadir images (2FX, 2FR), brings about a slight
improvement in the re-projection error.
• When GCPs are measured both in nadir and oblique views (3FR, 3FX),
the re-projection error grows.
• σ0 decreases in self-calibrating BBA, but the improvement is minor.
Accuracy of the adjustment process:
RMS error on CPs [m]
Take-away message:
• For the Milan dataset, observations in oblique views improve object point
accuracy in height.
MILAN
GCP
config.
RMS
[m]
No.
GCP
No.
CP 1FR 2FR 3FR 1FX 2FX 3FX
0
x 0.12
(0.16)
0.12
(0.14)
0.15
(0.15)
0.10
(0.16)
0.07
(0.13)
0.13
(0.16)
y 7 12 0.12
(0.18)
0.10
(0.12)
0.15
(0.14)
0.18
(0.17)
0.28
(0.13)
0.30
(0.14)
z 0.42
(0.32)
0.30
(0.30)
0.13
(0.20)
0.45
(0.34)
0.55
(0.31)
0.20
(0.20)
1
x 0.23
(0.26)
0.22
(0.21)
0.18
(0.16)
0.36
(0.26)
0.16
(0.20)
0.21
(0.17)
y 4 15 0.22
(0.19)
0.14
(0.13)
0.16
(0.14)
0.27
(0.18)
0.48
(0.14)
0.45
(0.15)
z 6.85
(0.76)
1.39
(0.54)
0.26
(0.22)
10.68
(0.77)
3.68
(0.56)
0.64
(0.20)
Influence of oblique views
GRAZ
GCP
config.
RMS
[m]
No.
GCP
No.
CP 1FR 2FR 3FR 1FX 2FX 3FX
1
x 0.04
(0.04)
0.04
(0.04)
0.05
(0.05)
0.04
(0.05)
0.04
(0.06)
0.03
(0.05)
y 4 3 0.03
(0.03)
0.06
(0.04)
0.04
(0.03)
0.04
(0.03)
0.06
(0.04)
0.05
(0.04)
z 0.08
(0.04)
0.08
(0.10)
0.02
(0.04)
0.17
(0.20)
0.13
(0.18)
0.13
(0.16)
4
x 0.05
(0.04)
0.04
(0.06)
0.04
(0.06)
0.04
(0.06)
0.05
(0.15)
0.04
(0.13)
y 4 3 0.05
(0.04)
0.05
(0.04)
0.06
(0.03)
0.06
(0.03)
0.05
(0.03)
0.06
(0.06)
z 1.07
(0.07)
2.47
(0.13)
2.86
(0.18)
2.22
(0.21)
2.86
(0.28)
3.32
(0.42)
Accuracy of the adjustment process:
RMS error on CPs [m]
Take-away messages:
• For the Graz dataset the advantage of oblique views is less obvious.
• In the presence of unfavorable point distributions, nadir images (1FR)
deliver better Z-accuracy results if compared to 2FR or 3FR with the
oblique views included.
Influence of oblique views
Accuracy of the adjustment process:
RMS error on CPs [m]
Take-away message:
• For the Milan dataset, BBA without GNSS generally shows a gain in
accuracy if self-calibration is performed. In the GNSS-assisted solutions
this evidence is less pronounced.
MILAN
GCP
config.
RMS
[m]
No.
GCP
No.
CP 1FR 2FR 3FR 1FX 2FX 3FX
0
x 0.12
(0.16)
0.12
(0.14)
0.15
(0.15)
0.10
(0.16)
0.07
(0.13)
0.13
(0.16)
y 7 12 0.12
(0.18)
0.10
(0.12)
0.15
(0.14)
0.18
(0.17)
0.28
(0.13)
0.30
(0.14)
z 0.42
(0.32)
0.30
(0.30)
0.13
(0.20)
0.45
(0.34)
0.55
(0.31)
0.20
(0.20)
1
x 0.23
(0.26)
0.22
(0.21)
0.18
(0.16)
0.36
(0.26)
0.16
(0.20)
0.21
(0.17)
y 4 15 0.22
(0.19)
0.14
(0.13)
0.16
(0.14)
0.27
(0.18)
0.48
(0.14)
0.45
(0.15)
z 6.85
(0.76)
1.39
(0.54)
0.26
(0.22)
10.68
(0.77)
3.68
(0.56)
0.64
(0.20)
Influence of self-calibration
GRAZ
GCP
config.
RMS
[m]
No.
GCP
No.
CP 1FR 2FR 3FR 1FX 2FX 3FX
1
x 0.04
(0.04)
0.04
(0.04)
0.05
(0.05)
0.04
(0.05)
0.04
(0.06)
0.03
(0.05)
y 4 3 0.03
(0.03)
0.06
(0.04)
0.04
(0.03)
0.04
(0.03)
0.06
(0.04)
0.05
(0.04)
z 0.08
(0.04)
0.08
(0.10)
0.02
(0.04)
0.17
(0.20)
0.13
(0.18)
0.13
(0.16)
4
x 0.05
(0.04)
0.04
(0.06)
0.04
(0.06)
0.04
(0.06)
0.05
(0.15)
0.04
(0.13)
y 4 3 0.05
(0.04)
0.05
(0.04)
0.06
(0.03)
0.06
(0.03)
0.05
(0.03)
0.06
(0.06)
z 1.07
(0.07)
2.47
(0.13)
2.86
(0.18)
2.22
(0.21)
2.86
(0.28)
3.32
(0.42)
Accuracy of the adjustment process:
RMS error on CPs [m]
Take-away message:
• For the Graz dataset, freeing self-calibration parameters in the bundle
adjustment improves results in the Z coordinates for all GCPs
configurations.
Influence of self-calibration
Accuracy of the adjustment process:
RMS error on CPs [m]
Take-away messages:
• For the Milan dataset, in the presence of unfavorable point distributions
extrapolation effects emerge and are reflected in larger RMS error
values (the Z coordinate suffers the most).
• The inclusion of GNSS constraints largely mitigates the problem.
MILAN
GCP
config.
RMS
[m]
No.
GCP
No.
CP 1FR 2FR 3FR 1FX 2FX 3FX
0
x 0.12
(0.16)
0.12
(0.14)
0.15
(0.15)
0.10
(0.16)
0.07
(0.13)
0.13
(0.16)
y 7 12 0.12
(0.18)
0.10
(0.12)
0.15
(0.14)
0.18
(0.17)
0.28
(0.13)
0.30
(0.14)
z 0.42
(0.32)
0.30
(0.30)
0.13
(0.20)
0.45
(0.34)
0.55
(0.31)
0.20
(0.20)
4
x 0.28
(0.28)
0.20
(0.24)
0.19
(0.23)
0.63
(0.27)
1.16
(0.22)
1.46
(0.25)
y 3 16 0.21
(0.19)
0.23
(0.09)
0.28
(0.13)
0.72
(0.18)
2.02
(0.10)
2.50
(0.11)
z 4.30
(0.72)
2.24
(0.52)
1.59
(0.22)
7.62
(0.70)
4.75
(0.52)
2.30
(0.24)
Influence of GCP distribution
GRAZ
GCP
config.
RMS
[m]
No.
GCP
No.
CP 1FR 2FR 3FR 1FX 2FX 3FX
1
x 0.04
(0.04)
0.04
(0.04)
0.05
(0.05)
0.04
(0.05)
0.04
(0.06)
0.03
(0.05)
y 4 3 0.03
(0.03)
0.06
(0.04)
0.04
(0.03)
0.04
(0.03)
0.06
(0.04)
0.05
(0.04)
z 0.08
(0.04)
0.08
(0.10)
0.02
(0.04)
0.17
(0.20)
0.13
(0.18)
0.13
(0.16)
4
x 0.05
(0.04)
0.04
(0.06)
0.04
(0.06)
0.04
(0.06)
0.05
(0.15)
0.04
(0.13)
y 4 3 0.05
(0.04)
0.05
(0.04)
0.06
(0.03)
0.06
(0.03)
0.05
(0.03)
0.06
(0.06)
z 1.07
(0.07)
2.47
(0.13)
2.86
(0.18)
2.22
(0.21)
2.86
(0.28)
3.32
(0.42)
Accuracy of the adjustment process:
RMS error on CPs [m] Influence of GCP distribution
Take-away messages:
• The sensitivity of the Graz network is minor, especially when GNNS
constraints are included.
• The imaging geometry and extent of image overlap are the most
probable factors influencing this sensitivity.
Possible deformations and the level of noise in the derived 3D object
coordinates are evaluated. Two types of comparisons are performed:
• In order to establish the influence of the inclusion of oblique
images, scenarios 1FX and 1FR are compared to scenarios 3FX and
3FR. The influence of self-calibration on the triangulated 3D points
is expressed in the differences 1FX with 1FR and 3FX with 3FR.
Results obtained in GCP configuration 0 (Milan) and 1 (Graz) are
used.
• To determine the influence of the GCP distribution, the oblique
image blocks 3FR and 3FX, in all GCP configurations, are compared
against a reference result. The reference is unique for all comparisons
and was chosen to be the one reporting the optimal RMS error values
at CPs, namely 3FR in GCP configuration 0 (Milan) and 1 (Graz).
Mean differences, standard deviations and color coded maps of the
geometric differences (Euclidean distance) are produced.
Influence of GCP distribution
MILAN
Re
fere
nc
e:
3F
R in
GC
P c
on
fig
. 0
1 2 3 4
no-GNSS
μ = 0.18 m
σ = 0.09 m
μ = 0.50 m
σ = 0.27 m
μ = 1.10 m
σ = 0.51 m
μ = 0.98 m
σ = 0.63 m
GNSS
μ = 0.09 m
σ = 0.04 m
μ = 0.18 m
σ = 0.09 m
μ = 0.28 m
σ = 0.15 m
μ = 0.14 m
σ = 0.07 m
0
6σ [m]
Take-away messages:
• In both datasets, GCPs induce a ‘strain’ on the entire block, i.e. the discrepancies
are kept small within the general region enclosed by GCPs whereas extrapolation
errors increased away from this region.
• When GNSS constraints are introduced into the BBA, the magnitude of these
discrepancies is substantially reduced, as is the extrapolation error.
Influence of the tilt angle
𝛼 < tan−1𝑠
ℎ
Take-away messages:
• When the inequality holds, the point will not be occluded by the neighboring
building.
• The taller the city architecture the lower the camera incidence angle should be (s
and h play a major role in planning a successful urban survey campaign!)
• A compromise should be found between the camera tilt setting, given focal length,
sensor size, overlap and the geometry of the surveyed area.
View angle from the perspective centre to the foot of the building:
tilt angle ± the angular point position in the image
Street width
Building height
Influence of the overlap
Take-away messages:
• Because of the large dissimilarities between O/N views, points should be matched
across views of similar look direction (i) within the strip and (ii) from neighboring
strips.
• Both the along- and across-track fold calculated on the oblique images should
exceed 50% (but larger datasets and higher number of 3D points should be
managed!)
Reconstruction using only forward/backward
images within one strip
Reconstruction using only forward/backward
images within two adjacent strips
Graz dataset
UltraCam Osprey I - GRAZ dataset
SURE -
GENERAL VIEW
ONLY NADIR IMAGES
NADIR AND OBLIQUE IMAGES
UltraCam Osprey I - GRAZ dataset
MICMAC – GENERAL VIEW (nadir + oblique images)
UltraCam Osprey I - GRAZ dataset
PRE-FILTERING
MICMAC – CLOSE-UP VIEWS
(nadir + oblique images) POST-FILTERING
Mixed pixels filtering
(considering the direction of
the perspective rays and the
local surface directions).
Rupnik, E., Nex, F. and Remondino, F., 2014. Oblique multi-camera systems – orientation and
dense matching issues. International Archives of Photogrammetry, Remote Sensing and
Spatial Information Sciences, XL-3/W1, pp. 107-114.
UltraCam Osprey I - GRAZ dataset
PRE-FILTERING
POST-FILTERING
Vegetation filtering
(considering the RGB values
associated to extracted point
clouds)
PRE-FILTERING
MICMAC -
CLOSE-UP VIEWS
(nadir/oblique images)
important for the successive
building reconstruction
From the dense point clouds, there are basically two successive
steps:
Huge and unique mesh/polygonal model generation
3D modeling of each single building
(*) According to the OGC standard CityGML
General pipeline (based on Tridicon BuildingFinder module and filtered
point clouds)
Point cloud splitting into
tiles by user specified size.
Segmentation: the result of the segmentation
process is a partition of the points, where all
points in one segment belong to the same
shape.
Building ground height definition: a terrain
model is generated from the point cloud in for
determining the building ground heights.
Building generation: the detection of
building objects is performed by searching
for selected types of internal- (or user-)
defined roof types.
UltraCam Osprey I system - GRAZ dataset
From filtered
point cloud
To building
models
(LOD2, no roof
overhang)
UltraCam Osprey I - GRAZ dataset
SINGLE BUILDING MODELS
UNIQUE MESH MODELS
IGI PentaCam camera - DORTMUND dataset (EuroSDR/ISPRS Benchmark)
Microsoft Ultracam camera – TRENTO dataset (only nadir, 10 cm GSD)
demonstrated the behavior of different oblique imaging platforms in real-world
conditions and studied the influence of the additional slanted cameras on:
The inclusion of oblique imagery into the BBA
brings about:
- better accuracy on Z-coordinate;
- improved image block stability when the
overlap is small, or the GCPs distribution is
poor, or there is not GNSS assistance.
-To ensure high quality façade reconstruction
from oblique views, it is desirable that points
are intersected from at least two neighboring
strips.
- A compromise should be found between the
camera tilt setting, given focal length, sensor
size and the overlap within the flight lines.
- New solutions to derive structured
information out of unstructured point clouds
are required.
- The use of oblique imagery to provide
detailed information on building facades
should be integrated with data collection from
terrestrial viewpoints.
Isabella Toschi, Fabio Remondino
3D Optical Metrology (3DOM),
Bruno Kessler
Foundation (FBK),
Trento, Italy
Acknowledgments:
Ewelina Rupnik (IGN, France), Francesco Nex (Univ Twente / ITC, The Netherlands)