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
ous publications comparing the two different acquisition systems.
A thorough comparison of airborne laser scanning (ALS) and tra-
ditional manual photogrammetry for the generation of DSM and
DTM models is presented in (Baltsavias, 1999). The author un-
derlines that laser scanning has a lot of advantages in terms of
density of measurements, automation and rapidity compared to
traditional manual photogrammetry but nevertheless the two tech-
nologies are complimentary to each other since the one can out-
perform the other in certain tasks. In (Beraldin, 2004) the author
is also examining this complementarity between laser scanning
and photogrammetry and addresses the problem of data fusion
and multi sensor imaging systems and their practical applications.
He also presents a short review of the basic theory, the limitations
and the theoretical accuracies acquired by both laser scanning
systems and close range photogrammetry. In (El-Hakim et al.,
2008) the integration of Image Based Modelling (IBM) and laser
scanning techniques are also examined. In this paper laser scan-
ning is mixed with photogrammetry and aerial photogrammetry
with photo acquired by a balloon in order to model the Erectheion
in Athens. The challenges presented to both laser scanning and
photogrammetry by the marble surfaces due to their reflectance
and texture problems are underlined in these publications.
TLS techniques have been widely adopted for cultural heritage
documentation and therefore many papers have tried to investi-
gate the advantages and disadvantages of TLS and manual close
range photogrammetry (Kadobayashi et al., 2004), (Alshawabkeh
and Haala, 2004) (Boehler and Marbs, 2004), and (Grussenmeyer
et al., 2008). The general conclusion is that due to the complexity
of the scenes and of the materials used the choice of the method is
heavily correlated to the scene that has to be modelled and hence
a combination of the two methods could be interesting in various
cases.
The advances in automation of photogrammetric procedures by
the adoption of the computer vision techniques and the introduc-
tion of commercial and open source automatic or semi-automatic
software have led to significant advances in the use of IBM in
different applications. (Seitz et al., 2006) present a quantitative
comparison of various multi-stereo view (MVS) image based re-
construction algorithms for modelling small objects in indoor en-
vironments. The image datasets were compared to datasets ac-
quired with the use of a laser stripe scanner which provided very
high accuracy. The authors conclude that all the algorithms per-
form well offering accuracies very close to the ones of the laser
scanner. (Strecha et al., 2008) also present an evaluation of image
based reconstruction algorithms with TLS systems. This evalu-
ation concerns datasets of large objects acquired in outdoor con-
ditions. The authors conclude that (Furukawa and Ponce, 2010)
provide the best results. Nevertheless they point out that for large
scale outdoor scenes the multi-view algorithms still lack in ac-
curacy compared to laser scanning. In (Remondino et al., July
2008) a photogrammetric software for image matching and IBM
is presented. An evaluation of the technique is also provided with
a comparison of the photogrammetric software results for various
scenes with those of a phase shift laser scanner and a triangulation
scanner. The authors conclude that the new techniques adopted
by the photogrammetric community can lead to the production of
image based models with geometrical properties compared to the
ones produced by laser scanners with the advantage of portability
and lower cost.
The comparison that is presented in this paper aims to evaluate
3D models of the interior of a complex, low textured modern
building produced by a totally automatic photogrammetric open
source software and a TLS. We evaluate not only the accuracy of
the two models but also the time spent on field and office in order
to acquire them.
24
3 METHODOLOGY DESCRIPTION
The camera used for capturing the images is a Canon EOS 5D
Mark II coupled with a 8mm Samyang fisheye lens. The cam-
era's sensor is a full frame 21 megapixel CMOS sensor which
corresponds to a sensor pixel size of 6.4 jum. The Samyang fish-
eye lens is a low cost fisheye lens made for APS-S sized sensors
which means that only 70% of the diagonal of the camera’s sensor
is used. The Samyang lens is one of the cheapest fisheye lenses
in the market and provides 180°circular fisheye images to full
frame sensors. It is made of 10 optical elements, one of which
is aspherical, grouped in 7 groups. For the laser scanner cam-
paign we have used a Leica Scanstation 2. The Scanstation 2 is
a Time Of Flight laser scanner which offers at 50m a single mea-
surement position accuracy of 6mm, distance accuracy of 4mm,
horizontal angle accuracy of 60urad and vertical angle accuracy
of 60urad. The modeled surface precision offered by the scanner
is 2mm which guarantees that the noise levels in the cloud points
are low. For the laser point clouds alignment and registration the
Cyclone ((c)Leica) software was used. For the scan registration
black and white Leica targets where used on the field. For the
image derived point cloud the APERO/MICMAC (©IGN) Open
Source software (APERO/MICMAC, 2012) was used. The soft-
ware functionalities will be further analysed in the next section.
The two cloud points where compared using the CloudCompare
Open Source software (CloudCompare, 2012) (©EDF R& D).
The software offers various tools for the direct comparison of
the distances between different cloud points or/and meshes of the
same scene.
3.1 Data acquisition
The stairway dataset was chosen for our experimentations be-
cause it has several distinct characteristics that interest us. Firstly
it is inside a modern building which means that is not very rich
in texture and thus making it a rather challenging environment
for the extraction of SIFT points and for the generation of dense
point clouds from images. Secondly being part of a modern build-
ing its surfaces are rather vertical and planar allowing us to detect
easily the deviations in the geometries of its surfaces. Finally
the rather complex geometry of a modern building stairway is an
added challenge for the stability of the camera network and the
bundle adjustment of the photos. Our stairway is a typical U-
shaped stairway with full landings in the middle and black metal
handrails. The stairwell spans in 4 floors and is 12 meters high.
In order to be able to capture our scene with the less photos pos-
sible we had to use a lens that would permit us to have photos
with a large field of view in the confined space of a stairway. The
8mm fishey lens that was used allowed us to acquire photos with
a global view of the scene (Figure 1) and thus with textured zones
in every image, which is essential for the SIFT algorithm. It also
permitted us to have a sufficient overlap between more than two
images in order to be able to effectively perform the multi-image
matching. One photo was taken on each of the steps and about 8
photos were taken on each of its landings. In total we have ac-
quired 111 photos in RAW format. The total acquisition time was
of about 1 hour. At the same time we have also acquired a dataset
of 15 images for the lens calibration.
For the TLS acquisition we have opted to acquire laser scans
with a 5mm resolution. 2 scanstations were set at each landing.
The total number of scanstations is 20 and the average duration
of each scanstation was about one hour. For the consolidation of
the different scans in one point cloud we have used a mix of Le-
ica Black and White targets and spheres of known diameter. The
total acquisition time was of about 20 hours.
32
The
soft
One
soft
soft
deli
initi
nee
The
velo
PM
201
tion
phis
whi
soft
justi
awa
ity c
user
orde
The
step
for
faul
of t
has
all [
in li
user
self
