DISCUSSION
The data
presented here makes it apparent that the Kinect™ can be a useful tool for
palaeontologists. Although the sensor, and thus resulting model, lacks the
resolution needed for smaller detailed scans, or for archival purposes (which
should always be carried out at the highest resolution available), it is able to
quickly capture the morphology of a specimen, and produce a digital model that
requires only minimal post-processing. One area set to particularly benefit from
using the Kinect™ sensor is in body mass and biomechanical simulations, where
models are required of individual bones or complete mounted skeletons, but the
sub-centimetre detail is unused, such as in convex-hulling specimens to estimate
mass (Sellers et al., 2012).
One real
advantage of using the Kinect™ sensor in this way is that the model is produced
on-the-fly, including meshing, so any errors (e.g. holes in the data collection)
can be corrected whilst scanning, contrary to photogrammetry where the quality
of the model is not known until after all processing is complete, which can take
some time and is algorithms used on photogrammetric point cloud data may fail,
or require time-consuming input from the user. Of the samples used here, the
pronghorn benefits from this immensely, as the Poisson Surface Reconstruction
used to mesh the point cloud struggles to differentiate between close features,
such as the ribs, as demonstrated in Figure 4. The meshed photogrammetric model
could be improved dramatically by manually segmenting parts of the point cloud
prior to surfacing, though obviously this would contribute considerably to the
total time of the workflow.
A significant
disadvantage to using ReconstructMe or the Point Cloud Library Kinfu is that the
scanning volume is limited, and must be set prior to scanning. Increasing that
volume either decreases the resolution of the scan, or increases the graphics
memory used. For the specimens used in this paper, this was not an issue, as the
scanning volume could be set relatively small (< 1 m3) and a high resolution and
low memory usage maintained. Fortunately, this limitation may be short-lived, as
the pre-release version 7 of the Point Cloud Library contains an application
‘kinfu_large_scale’ which is able to record larger, dynamically sized
environments (http://pointclouds.org/documentation/tutorials/using_kinfu_large_scale.php.
Another
disadvantage is that the models produced in this paper by the Kinect™ lack any
colour information. This is in contrast to photogrammetry, and some traditional
laser scanners which can photo-texture their resulting models. However, in
addition to the infrared depth sensor, the Kinect™ (and Asus Xtion Pro) also
possesses RGB cameras capable of recording colour information. Although in the
early stages of development, all of the relevant software packages discussed
here are actively developing features to incorporate texture generation.
Back
Next