XIX-B3, 2012
IPTION
is a Canon EOS 5D
eye lens. The cam-
"MOS sensor which
. The Samyang fish-
APS-S sized sensors
f the camera’s sensor
eapest fisheye lenses
sheye images to full
ments, one of which
> laser scanner cam-
The Scanstation 2 is
at 50m a single mea-
ce accuracy of 4mm,
rtical angle accuracy
ffered by the scanner
Is in the cloud points
t and registration the
the scan registration
on the field. For the
MAC (©IGN) Open
was used. The soft-
1 in the next section.
v the CloudCompare
12) ((OEDF R& D).
lirect comparison of
or/and meshes of the
>xperimentations be-
at interest us. Firstly
that is not very rich
lenging environment
> generation of dense
wt of a modern build-
allowing us to detect
its surfaces. Finally
lilding stairway is an
era network and the
rway is a typical U-
ddle and black metal
d is 12 meters high.
| the less photos pos-
üt us to have photos
ce of a stairway. The
) acquire photos with
s with textured zones
FT algorithm. It also
‘ween more than two
form the multi-image
the steps and about 8
In total we have ac-
acquisition time was
so acquired a dataset
) acquire laser scans
> set at each landing.
the average duration
the consolidation of
ve used a mix of Le-
«nown diameter. The
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
Figure 1: Stairway fish-eye images
3.2 APERO/MICMAC
The IGN has decided in 2007 to deliver as Open Source several
software that have been developed within the Matis laboratory.
One of these is the APERO/MICMAC suite. The APERO is a
software that computes orientation of images and MICMAC a
software that calculates depth maps of oriented images and can
deliver them as dense point clouds. The MICMAC software was
initially used in aerial images but nowadays is also adapted to the
needs of close range and terrestrial photogrammetry.
The main difference of the APERO/MICMAC from software de-
veloped within the Computer Vision community like Bundler-
PMVS (Furukawa and Ponce, 2010) or Samantha (Gherardi et al.,
2011) is the introduction of photogrammetric rigidity in the equa-
tions. Furthermore the camera calibration model used is more so-
phisticated and allows the calibration of fish-eye lenses, an option
which is not proposed to our knowledge by other Open Source
software. It also allows the self-calibration during the bundle ad-
justment steps. However the user of APERO/MICMAC must be
aware that this software does not trade precision to the flexibil-
ity of creating 3D models from unordered images. Therefore the
user should follow the rules of photogrammetric acquisition in
order to get the optimal results.
The 3D modelization process is done in three steps. In the first
step tie points are computed between the images. A modified
for large images version of sift++ (Vedaldi, 2010) is used by de-
fault but the user could use any other detector for the extraction
of tie points within the APERO/MICMAC pipeline. The user
has the possibillity to select between computing tie points for
all pair of images or define the number of images that overlap,
in linear datasets, thus accelerating the computation time. The
user may choose to provide calibration data for its camera or a
self-calibration may be performed during the bundle adjustment
25
procedure. Several calibration models are proposed by APERO:
e Distortion free model
e Radial distortion polynomial model
e Radial distortion with decentric (fraser)
e Ebner's and Brown's model
e Polynomial models from degree 3 to 7
e Fish-eye models for diagonal and spherical fish-eyes
The two models used for fish-eye lenses are made by a combi-
nation of theoretical equidistant model and a polynomial distor-
tion. The polynomial model has 14 degrees of freedom 1 for focal
length, 2 for principal point, 2 for distortion center, 3 coefficients
of radial distortion, 2 decentric parameters and 2 affine parame-
ters. The main difference is that for a diagonal fish-eye the model
considers that the useful area in an image is within the 95% of its
diagonal whereas for a spherical fish-eye this percentage is 52%.
The second step is the computation of the orientations by the AP-
ERO. The relative orientation is calculated from the tie points and
if it is needed the relative orientation can be converted to absolute
orientation with the use of control points or GPS/INS data.
In order to calculate an initial solution an image is selected ei-
ther by the user or by APERO which sets the coordinate system.
The next image for orientation is chosen based on certain criteria
such as the number of common tie points and their distribution
in the images. APERO uses the essential matrix coupled with
RANSAC and if there are enough tie points the space resection
with RANSAC. The best solution is chosen at the end of the pro-
cedure. A bundle adjustment of the oriented images is performed
in regular intervals in order to avoid the solution’s divergence.
The bundle adjustment follows the classical procedure presented
in (Triggs et al., 2000). An estimation of the ground point is cal-
culated by bundle intersection of all images it is seen and a mini-
mization term, which is the sum of the retroprojection in the im-
ages of the ground point, is then added. The term is linearised and
is added to a global quadratic form that has to be minimized. The
system is then solved using the Cholesky decomposition method.
The MICMAC software and the generation of 3D point clouds
through multi-scale and multi-resolution matching are extensively
described in (Pierrot-Deseilligny and Paparoditis, 2006)
3.3 Data Treatment
The TLS scans were treated using the Leica software of Cyclone.
The black and white targets and the spheres that were acquired
during the TLS acquisitions where used to register the different
scans in one global scan of the whole stairway. The windows and
everything that was outside the stairwell was excluded from the
global scan in order to be able to effectively compare the TLS
dataset to the IBM dataset. The use of targets and spheres for the
registration of the scans meant that the process was automated
and very fast compensating partially the long on-site acquisition
times(Figure 2).
The initial step for the orientation of our acquisition images is
the auto-calibration of the camera lens with the use of the APERO
software. The duration of the whole procedure was about 2hours
on a Intel 2.83GHz Core2 Quad machine with 4Gb of RAM. The
root mean square error (RMS) of the bundle adjustment for the
calibration dataset was 0.35 pixels. The auto-calibration values
where then used as initial values for the bundle adjustment of the
stairway dataset. In order to accelerate the process of tie points
generation we have considered that an image can overlap with
