XXXIX-B3, 2012
t cloud
d ery (1)
id the models proposed
1-Montaut et al., 2005)
int cloud and the TLS
m. The maximum dis-
the dense point cloud
1 a sigma of 6 cm. The
Figure 5).
d in order to better un-
he sparse IBM point
is purpose horizontal
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 5: Distance comparison between a part of the IBM dense
point cloud with the TLS point cloud
and vertical sections of the TLS and the IBM point clouds have
been made. As it can be seen in Figure 6 it is clear that the main
geometry problem of the sparse point cloud is on the Z axis of
our IBM mode where there is a systematic difference between
the TLS point cloud and the IBM point cloud. This difference
could be due to the fact that the IBM point cloud is very sparse
on the ceiling and therefore the point cloud is ill-referenced on
the Z axis. On the other hand the X and Y axis, as it can be seen
on Figure 7, don't suffer from significant geometry problems we
can therefore assume that the results would have been signifi-
cantly better if the ceiling was sufficiently textured.In Figure 5
we can also notice that during the dense point cloud generation
MICMAC was not able to sufficiently compensate the distortions
of the fish-eye lens and therefore the differences between the TLS
point cloud and the dense point cloud seem to follow, much less
extensively, the pattern of distortion of the fish-eye lens which
could be attributed to the poor quality of the optics used for the
construction of the low cost fish-eye lens that was used in our
acquisition.
M Proto
Figure 6: Horizontal section of IBM generated point cloud and
TLS point cloud)
de
d Photo
Figure 7: Vertical section of IBM generated point cloud and TLS
point cloud)
5 CONCLUSIONS AND FUTURE WORK
We have presented a comparison of TLS and of a fully automatic
photogrammetric work-flow IBM. We should note that we have
not compared the IBM results of our software with results of other
commercial or open source software such as Bundler-PMVS due
to the fact that to our knowledge these software don't offer the op-
tion of calibrating fish-eye lenses on which we have solely been
based for capturing our complex interior scene. However this
choice was essential to our purpose since it allowed us to capture
our scene using the fewer possible images with big overlaps and
thus to a)accelerate the photogrammetric process and b)establish
that the algorithm would be able to converge to a satisfactory re-
sult. The overall result of our approach may not reach the geo-
metric precision of a laser scanner, since it is heavily constrained
by the lack of texture that characterizes most of the modern build-
ings, however it still offers an interesting solution to TLS. In fact
the use of IBM in our case trades part of the TLS accuracy for
lower cost since the IBM dataset can be captured with any of the
shelf dSLR camera equipped with a low cost fish-eye lens and an
open source software instead of using an expensive laser scanner
and its proprietary software. Another advantage of the IBM mod-
elling solution compared to the TLS is the significantly faster ac-
quisition of the images for the model generation, which could be
essential for applications that demand quick and low cost mod-
elling of complex interior spaces with low texture instead of a
very accurate 3D model. Another advantage of the IBM mod-
elling is its portability since a dSLR camera -or any other digital
camera- and a tripod can be easily transported by a single per-
son whereas TLS tend to be heavy and bulky thus making their
transportation a difficult and sometimes complicated and expen-
sive task. Another advantage of the IBM method compared to
the TLS used in our experimentations is that it directly provides
a high quality textured by the R,G,B information of the images
whereas the texture in the TLS model is of inferior quality due to
the significantly lower resolution of its camera.
We believe that the results of our method can be improved with
the use of a better quality fish-eye lens. Another step that could
be added to our approach in order to achieve better results is the
use of lenses of longer focal length for capturing images, with or
without sufficient overlap, of higher resolution for certain areas of
interest or even for the whole scene. The highly overlapping fish-
eye images would therefore provide the skeleton of our model
allowing the connection of the longer focal-length images that
can be used for the generation of dense point clouds. Finally the
