The anthracotheriid cranium used in this study was found in South Dakota and
is housed at the Museum für Naturkunde Berlin, Germany (collection number
MB.Ma. 51832; Figure 1). The specimen is exceptionally well-preserved in
that the cranium is virtually complete, except for missing front teeth and
some damage to the basisphenoid. Post-mortem deformation is restricted to a
slight leftward shear that includes some dorso-ventral compression of the
right side of the cranium. Condylo-basal length is 37.5 cm.
MB.Ma. 51832 has not been formally described yet but is labelled as "Bothriodon
americanus". The presently valid species name would be Aepinacodon
americanus (Kron and Manning, 1998; Lihoreau and Ducrocq, 2007). However, A.
americanus is late Eocene in age. If the age information on our specimen,
early Oligocene, is correct, then it is possible we are dealing with
Bothriodon rostratus instead. But since the present study is purely
technical, description and assignment of the specimen on the species level
are beyond the scope of this paper. Aepinacodon and Bothriodon are likely
sister taxa within the Bothriodontinae (Lihoreau and Ducrocq, 2007).
Therefore, we decided to refer to our specimen as Bothriodontinae indet.
Bothriodontine anthracotheriids possibly lived along river banks in
swamp-like environments, a potential semi-aquatic lifestyle being discussed
controversially (Clark et al., 1967; Kron and Manning, 1998). They may have
been forest-dwelling browsers, or possibly generalist feeders (Kron and
Manning, 1998; Lihoreau and Ducrocq, 2007; cf. Janis, 1995).
Figure 1 -
The cranium of Bothriodontinae indet. (MB.Ma. 51832). A) Dorsolateral view.
B) Ventrolateral view. Scale bar is 10 cm.
In this section, settings and programs used on each dataset as well as the
conduction of the comparisons are described. Documentation of the learning
process and practicing times before successful application by MA was part of
the evaluation of each method. Surface names, methods and programs used, and
duration of 3D surface generation and processing are summarized in Table 1
for all models.
Photographs and surface files are available for download at https://figshare.com/s/73c651841
e9b3271d9ee. Further files related to this study are stored at the Museum
für Naturkunde Berlin and can be made available by the corres-ponding author
The cranium MB.Ma. 51832 was scanned in two segments at the Steinmann-Institut,
University of Bonn, Germany, using a micro-CT scanner (GEphoenix|X-ray
v|tome|x s). Isometric voxel length was 0.246370 mm. Image stacks of two
scans were registered based on gray values and transformed to a single file
using Amira 5.0 software.
Table 1 -
Users, methods and programs used, and duration of individual steps of 3D
model generation. *CT surface generation times reflect the complete duration
from opening image stacks to finished surface extraction, including
registration of CT data sets. **Surface generation settings in PhotoScan
were chosen to limit face count.
reconstruction was done twice by different users via isosurface rendering.
JMF extracted one 3D surface of each the anterior and posterior half of the
cranium after registration but before transformation of the image stacks
using Amira 5.0 (thresholds were 13,500 and 15,200, respectively). Both
anterior and posterior surfaces were cleaned of internal surfaces that are
naturally generated by threshold-based surface extraction from CT data, and
merged into a single 3D surface in GOM Inspect V7.5 software (http://www.
gom.com/de/3d-software/gom-inspect.html). MA extracted one complete 3D
surface from the transformed image stacks in ZIB-Amira 2015.24 (developed by
the Zuse Institute Berlin; threshold was 15,177.2) and used Geomagic Studio
10 software for post-processing. During post-processing, internal surfaces
were removed by cutting the model in half, clearing out the interior, and
stitching the model back together. Finally, in order to ensure
comparability, polygon counts of both resulting 3D surface models were
reduced to 750,000 in Geomagic Studio, and meshes were saved as .ply file
Photogrammetry is a method that uses photos of an object to calculate a 3D
model, originally employed in mapping and architecture, and successively
adapted for other fields of science. Camera setup and photo alignment used
to be rather complicated procedures, relying on the principle of
triangulation (e.g., Wiedemann et al., 1999), but with steady progresses in
computer technology, photogrammetry has become a simple, quick and
inexpensive method for 3D reconstruction. For a review of the method see
Photogrammetry is now widely used in paleontology, and reviews addressing
its application in paleontological contexts conclude that overlap and
consistent lighting and focus are the most crucial points for taking
suitable photos (Fahlke et al., 2013; Fahlke, 2014; Mallison and Wings,
For this study, overlapping high-resolution photos of the dorsal and ventral
sides of the cranium were taken with a hand-held Canon EOS 650D digital SLR
camera (5184 x 3456 pixels). Photos were taken while walking around the
specimen, and no turntable was used. In order to distinguish this method
from automated photography, we are referring to it here as “manual
Three sets of photos were taken by different people but processed by the
same user. The set taken by MA comprised of 484 photos (198 from dorsal
perspectives, 286 from ventral perspectives) of which 378 were used in
alignment, with white balance (WB) set to tungsten light, using a 50 mm
fixed object lens and an external RF550 Macro LED ring flash. The set taken
by Heinrich Mallison (HM, Museum für Naturkunde Berlin) comprised 150 photos
(46 from dorsal perspectives, 104 from ventral perspectives), using an
object lens with 18-135 mm focal length, automatic WB, and the built-in
flash. For this set, the camera was held upside-down to light the naturally
shaded areas and thus to avoid hard shadow. The set taken by JMF comprised
83 photos (41 from dorsal perspectives, 42 from ventral perspectives), with
WB set to fluorescent light, using a fixed 50 mm lens and no flash light.
Sensitivity (ISO) was constant at 200 and depth of field was automatic in
all three sets.
Individual 3D surface models were generated from all sets of photos by MA in
Agisoft PhotoScan Professional 1.2.0, using high accuracy. Models were also
scaled in PhotoScan based on a scale placed next to the cranium and included
in the photos. Scaling was subsequently verified by measuring the condylo-basal
length of the model in MeshLab v1.3.1 (http://meshlab.sourceforge.net/) and
comparing it to the known condylo-basal length of the original cranium.
Post-processing was done in Geomagic Studio for all models. During
post-processing, holes were fixed, and redundant polygons were removed. In a
last step, polygon counts were reduced to 750,000, and meshes were saved in
.ply file format.
For automated photogrammetry we used the MDS Witikon technology provided by
EDICO SK, Bratislava, Slovakia, that is usually used to create so-called
object panoramas. Witikon is a system that makes use of four suspended
high-resolution cameras that are automatically positioned and calibrated.
For the particular photos of our specimen, four PHASE ONE IQ1 40 megapixel
cameras with PHASE ONE AF Macro 120mm/4 lenses were used, and light provided
by four external stationary FOMEI DIGITAL PRO 1200X flashes. The cranium was
placed on a turntable which was rotated in six-degree increments so that 288
photographs were taken from each the dorsal and ventral side. Color
calibration, alignment of the photographs and scaling were performed by
EDICO staff using their generic company software. The reconstructed surface
that was then provided for download consisted of 39.6 million polygons and
was reduced by MA to 750,000 polygons using Geomagic Software.
The object panorama of MB.Ma. 51832 created and hosted by EDICO SK can be
accessed at http://short.edico.sk/short/bothriodon/58386f (last accessed
October 20, 2015).
Table 2 - Surface models used in comparisons,
and motivation for comparisons.
We conducted five surface comparisons using CloudCompare v.2.6.1 and v.2.6.2
beta (http: //www.danielgm.net/cc/; software downloaded July 27, 2015).
These comparisons are listed in Table 2. CloudCompare was chosen for all
comparisons because it provides excellent registration and visualization
features, as well as simple statistical tools for evaluation. Furthermore,
researchers can address their questions directly to the developer, Daniel
Girardeau-Montaut (Grenoble, France), at http: //www.danielgm.net/cc/forum/
(last acces-sed October 28, 2015).
In CloudCompare, two point clouds or 3D surfaces are aligned, one
representing the reference, and the other one being registered to it. The
distance of these two point clouds or surfaces is measured at an arbitrary
number of points (the size of the cells in which the nearest point is
determined is usually pre-set by CloudCompare but can be changed by the user
by manipulating the so-called octree level). For all comparisons we used 20
iterations with a random sampling limit of 50,000, with rotation and
translation but no re-scaling allowed during alignment, and recommended
octree level 8.
The five comparisons address various aspects in regard to the methods used:
In order to test the feasibility and accuracy of Micro-CT scan-ning as
opposed to manual photogrammetry, two surfaces generated from respective
data sets produced by the same person (MA) were compared. Reproducibility
(inter-user error) was tested using surfaces generated from the same
micro-CT dataset by different users (MA and JMF), each choosing surface
extraction settings intuitively. Furthermore, reproducibility from datasets
that were created by different people (MA, HM, JMF) using individual
settings of the same method (manual photogrammetry), and processed by the
same user (MA), was addressed. Last, a surface generated with manual
photogrammetry was compared with one generated using an automated setup
(Table 2). Results of all surface comparisons are listed in Table 3 and
shown in Figures 2-6.
Table 3 - Comparison results: mean, standard deviation, and largest
distances between surfaces.
In order to compare the efficiency of the different methods, durations of
data collection, surface generation, and post-processing as well as polygon
counts of the output surfaces were documented and are listed in Table 1.
Quantitative and visual comparisons were conducted between two surfaces at a
time. Below, distances of the respective registered surface from the
reference surface are shown and discussed. Positive distances mean that
points of the registered surface lie above, i.e., outside of the reference
surface, whereas negative distances refer to points below, i.e., inside of
the reference surface. In the illustrations (Figures 2-6), positive
deviations are shown in yellow to red colors, negative deviations are shown
in blue to purple colors, and areas with virtually complete consistency of
the two surfaces are shown in green. The same color scale applies for all
In the comparison of surfaces generated with different methods, namely
micro-CT scanning and manual photogrammetry (Figure 2), CT_MA deviated from
P_MA by -0.31 mm on average, with 95% of the distance measurements falling
between -2.63 and 2.00 mm. Largest deviations were recorded at -11.48 and
6.68 mm (Table 3, Figure 2C). Largest negative distances occurred on the
medial surface of the squamosals and inside a crack across the nasals, as
well as on the medial side of both zygomatic bones, in some depressions
along the ventral midline suture of the maxillae, and in the jugular notch
posteromedial to the tympanic bulla. Largest positive deviations were seen
on the dorsal side of the frontal and nasal bones, on the zygomatic
processes of the squamosals, and on what appears to be artificial bridges
between the tympanic bullae and the basioccipital, and between the right
maxillary and the palatine (Figure 2A,B). Parts of the nuchal crest and the
supraorbital process of the frontal were missing in this comparison.
Figure 2 -
Comparison of the surfaces P_MA generated by means of manual photogrammetry
and CT_MA generated from micro-CT scans. A) Anterodorsal view. B)
Posteroventral view. C) Distribution of distances between surfaces based on
612 classes. Gauss curve superimposed on histogram.
the results of surface extraction from the same CT dataset by different
users (Figure 3), mean distance and largest distances between CT_MA and
CT_JMF, as well as the standard deviation, were the largest recorded in all
comparisons. Average deviation of CT_JMF from CT_MA was -0.33 mm. 95% of the
distance measurements fell between -3.60 and 2.94 mm, with the largest
distances at -14.38 and 11.78 mm, respectively (Table 3, Figure 3C). These
extreme values are surprising in the light of the almost all-green
comparison indicating a high degree of similarity of both surfaces (Figure
3A,B), a matter that will be further elaborated on in the Discussion.
Greatest negative distances on the surface occurred on the ventral side of
the maxillae and in crevasses on the ventral side of the supraorbital
process of the left frontal, whereas the largest positive distances could be
found in small, possibly artificial mounds emerging from the surfaces of the
pterygoid and sphenoid bones. The artificial bridges between the tympanic
bullae and the basioccipital, and between the maxillary and the palatine
seen in the previous comparison were present in this comparison as well,
however, they were mostly green here, indicating no differences between the
two surfaces. Again, parts of the nuchal crest and the supraorbital process
of the frontal were missing in this comparison.
Figure 3 -
Comparison of the surfaces CT_MA and CT_JMF generated from micro-CT scans.
A) Anterodorsal view. B) Posteroventral view. C) Distribution of distances
between surfaces based on 612 classes. Gauss curve superimposed on
comparisons of surfaces generated from different sets of photographs, the
surface based on the largest set of photographs (P_MA) served as a reference
surface to which the surfaces derived from an intermediate number of photos
(P_HM) and from a small number of photos (P_JMF) were compared.
Figure 4 -
Comparison of the surfaces P_MA and P_HM generated by means of manual
photogrammetry. A) Anterodorsal view. B) Posteroventral view. C)
Distribution of distances between surfaces based on 612 classes. Gauss curve
superimposed on histogram.
first of these comparisons (Figure 4), P_HM deviated from P_MA by -0.19 mm
on average. 95% of the distance measurements lay between -1.51 and 1.13 mm,
reflecting the smallest standard deviation of all comparisons. The greatest
distances between these surfaces, -9.82 and 5.59 mm, were also the smallest
recorded throughout the analysis (Table 3, Figure 4C). Largest negative
deviations of P-HM from P_MA could be found on medial surface of the
squamosals and of the zygomatic bones, and in the jugular notch, and largest
positive deviation appeared in the dorsal region of the frontal and nasal
bones and on the zygomatic processes of the squamosals, resembling the
comparison between P_MA and CT_MA. Additional large negative distances
occurred dorsally towards the nuchal crest, and large positive distances on
the tips of some teeth, on the ventral side of the right supraorbital
process, and on the lateral surfaces of the sphenoid and frontal bones
(Figure 4 A,B). Note that the nuchal crest, the nasal and the supraorbital
processes were as complete as in the original cranium (Figure 1), and there
were no depressions on the ventral side of the maxillae and no bridges
connecting the petrosal and the basioccipital, and the maxillary and the
palatine in this comparison.
Figure 5 -
Comparison of the surfaces P_MA and P_JMF generated by means of manual
photogrammetry. A) Anterodorsal view. B) Posteroventral view. C)
Distribution of distances between surfaces based on 612 classes. Gauss curve
superimposed on histogram.
second manual photogrammetry com-parison (Figure 5), P_JMF deviated from
P_MA by -0.31 mm on average. Medium standard deviation yielded 95% of the
distance measurements between -2.27 and 1.66 mm, with largest distances at
-11.34 and 7.45 mm, respectively (Table 3, Figure 5C). Visually, this
comparison revealed strong positive and negative deviations close to one
another (Figure 5 A,B). As in the other comparisons involving P_MA as the
reference surface, largest negative deviations of were recorded on medial
side of the squamosal and of the zygomatic bone, and in the jugular notch
adjacent to the tympanic bulla. Additional notable negative deviations of
P_JMF from P_MA were on the dorsal side of the frontal and nasal bones, in
several places across the basicranium, on the tooth cusps of the cheek
teeth, on the ventral side of the left premaxilla, in the alveoli of the
right incisors and canine, and on the labial side of the only present
incisor in the right premaxilla. Largest positive distances could be found
on the anterodorsal portion of the rostrum, particularly on the anterior rim
of the premaxillae, on the squamosals, on the occipital, especially on the
posterior surface of the occipital condyles, and on the ventral (orbital)
side of the frontals. Again, the cranium was complete in this comparison,
lacking artificial grooves or bridges.
Figure 6 -
Comparison of the surfaces P_MA generated by means of manual photogrammetry
and WITI generated using automated photogrammetry. A) Anterodorsal view. B)
Posteroventral view. C) Distribution of distances between surfaces based on
612 classes. Gauss curve superimposed on histogram.
differences of manual and automated photogrammetry were analyzed comparing
P_MA and WITI (Figure 6). Mean distance of WITI from P_MA was -0.13 mm, the
smallest mean distance of the entire analysis. 95% of all distance
measurements fell between -1.56 and 1.29 mm, and the largest distances
between the two surfaces were recorded at -11.34 and 4.09 mm (Table 3,
Figure 6C). Again, the largest negative distances were on the medial
surfaces of the squamosals and zygomatics, and in the jugular notch, while
the largest positive distances occurred on the dorsal surfaces of the
frontals and nasals, and on the zygomatic processes of the squamosals.
Further positive deviations of WITI from P_MA were seen on some tooth cusps,
near the medial suture and inside a crack on the ventral side of the
premaxillae, in the alveolus of the right canine, and, notably, on some
crest-like elevations on the ventral side of the occipital condyles (Figure
6A,B). Besides these elevations, no artificial surface structures were
noticed on the complete cranium in this comparison.
Here, we will discuss accuracy, efficiency and reproducibility of the
methods used in this study. Strengths and weaknesses of the individual
methods with regard to their application in 3D surface generation are summed
up in Table 4.
Table 4 -
Strengths and weaknesses of the tested methods with regard to 3D surface
Accuracy of a 3D surface basically refers to the highest possible degree of
similarity between the 3D surface model and the original specimen. One
aspect in terms of accuracy is certainly the maximum possible resolution.
"Resolution" should not be used synonymously with "polygon count", because
there may be redundant polygons in a mesh. However, if we are looking at
cleaned meshes, minute anatomical features are usually resolved more
accurately the higher the polygon count in the respective area of the
surface. So, overall polygon count potentially influences resolution. Since
computing technology is constantly improving, it is certainly advisable to
collect data at the highest possible quality, e.g., use the best available
camera resolution, and to store the original large files, such as surfaces
with high polygon counts, for later use. However, for transport (email,
upload) and manipulation with morphometric or modeling software, usually
decimated meshes with smaller polygon counts are used.
Maximum polygon count of the generated surface prior to decimation (Table 1)
was highest in automated photogrammetry where the cleaned surface provided
for download was 39.6 million polygons. Witikon initially produced a model
with an even higher count at 185 million polygons which we did not use in
this study due to the very large file size of 3.72 GB. Using regular office
computers (4 GB and 16 GB RAM, 3.3 GHz and 3.1 GHz processor, respectively)
and common post-processing software (MeshLab and GOM Inspect), it was hardly
possible to open this file and impossible to manipulate it. And not every
researcher or working group is equipped with computers with suitable
processors and memory for dealing with files this size. So while high
polygon counts might impact resolution, and high resolution is definitely
wanted in a 3D model that is supposed to be used for morphological analyses,
file size and computing capacity are limiting factors. Additionally, when
using any kind of photogrammetry, it does not make sense to produce and use
3D surfaces that have a higher polygon density than the pixel density of the
photographs used for model generation. This kind of artificially high
resolution would be misleading to the user and would not add to the true
quality of the surface in question.
In our experience, 3D surfaces out of PhotoScan based on manual
photogrammetry, provide by far enough detail for subsequent morphological
analysis, regardless of whether the output is the maximum polygon count (P_HM,
3.2 million polygons) or limited to 1 million (P_MA, P_JMF). Even in the
surfaces downsized to 750,000 polygons, morphological details are clear and
crisp, and file size is very handy at approximately 14 MB. Of course, the
desired accuracy of morphological detail depends on the purpose of the 3D
surface. If small features need to be resolved in great detail, image
resolution should be high to begin with, and additional close-up photos of
the area in question will be more useful than a high overall polygon count.
One should also keep in mind that resolution is determined by number of
polygons (or vertices) per surface unit, which means that object size should
be considered when choosing the desired total number of polygons in its 3D
Scaling is another factor that influences the similarity of a 3D surface and
the original object. False scaling can lead to invalid measurements, and
thus, depending on the kind of study, flawed morphometric statistics or even
wrong species assignment. The only method in our study that provided
automatic scaling was micro-CT scanning. Varying between common file formats
usually provided by scanning facilities, voxel size is usually either
included in the file information or automatically recorded so that it can be
entered manually when opening the data. Thus, exact scaling of the resulting
surface model is ensured. In manual as well as automated photogrammetry,
usually a physical scale is included in the photographs. For scaling, two
points on this scale are picked manually in several photographs (the more
the higher the accuracy of the scale), and the known distance between these
points is typed in. Since it is not simple to mark the exact same points
every time, if, e.g., the area is blurred in one photograph and cast in
shadow in another, this procedure bears quite some potential for errors (in
addition to potential mistakes typing in the distance) that may lead to
imprecise or incorrect scaling of the surface model.
In search of differences between the digital surfaces and the original
cranium, and between individual digital surfaces, some artificial features
were readily identifiable in our study. These include the occurrence of
bridges between the tympanic bullae and the basioccipital, and between the
right maxillary and the palatine in both CT-generated surface models (CT_MA
and CT_JMF; Figures 2 and 3). These features were not visible in any other
surface model, and they are not part of the original cranium (cf. Figure
1B). We interpret them to be the results of scanning artefacts: there appear
to be extremely densely mineralized areas in these regions of the tympanic
bullae, appearing very bright in the images of the image stack provided by
the scanning facility. There are notable halos around these bright spots,
extending toward the basioccipital.
In both CT-MA and CT_JMF, small parts of the cranium are missing, including
parts of the nuchal crest and the supraorbital process of the frontal, as
well as the filling of a crack across the nasals. These parts were lost
during threshold-based surface generation in Amira. Apparently, these areas
had been repaired during fossil preparation using adhesives and fillings
with a lower density than the fossil bone. This low-density material appears
darker in the CT scan and is deleted when the threshold for the gray value
of the fossil bone is applied. If a bone fragment is glued back in place and
is surrounded by lower-density material, it will be disconnected from the
cranium in the surface reconstruction and accidentally be deleted during
post-processing, which is what happened at the nuchal crest of our specimen.
Further artificial structures that are not part of the original cranium are
ridges on the ventral side of the occipital condyles in the 3D model
generated by automated photogrammetry (WITI) and stand out in its comparison
with P_MA (Figure 6B). These ridges mark the places in which the cranium was
touching the turntable when the photos of the dorsal side were taken. In
photogrammetry, these contact points between an object and the surface
underneath (that is usually deleted from the point cloud or model), are
sometimes reconstructed as transitions. The results of this kind of
reconstruction error are bumps and ridges on the 3D surface parallel to the
area of contact. Correct reconstruction of these contacts is especially
difficult, if they lie in the shade in the photographs, or if the object and
the supporting surface have similar colors. Possibly, the latter was the
case when automated photogrammetry was applied to MB.Ma. 51832, as the
bright beige color of the fossil bone (Figure 1) was not a strong contrast
to the white of the turntable that was used to rotate the specimen.
Reproducibility and efficiency
By reproducibility we are referring to the ability to produce consistent 3D
surfaces, either by using different methods or by different people applying
the same method in different ways. Reproducibility was generally very good:
mean distance between two surfaces was 0.33 mm or less in all comparisons,
which is less than 1‰ of the length of the cranium (37.5 cm). Furthermore,
in all comparisons, 95% of all distance measurements were within 1% of the
cranium length (Table 3).
As discussed above, parts of the surface of the original cranium that were
made of less dense material (glue, plaster filling) were not reconstructed
in the 3D surfaces based on micro-CT scanning. Therefore, although mean
distance and standard deviation were comparable, reproducibility between
Micro-CT scanning and manual photogrammetry could be regarded as lower than
reproducibility between manual and automated photogrammetry, or between
various data sets when manual photogrammetry was applied. Reproducibility
within methods (i.e., inter-user differences or differences based on the
settings used in data collection) is addressed in the discussion of the
individual methods below.
Efficiency is defined here as the balance between the time and money
invested in surface generation and the amount of usable data produced when
using a certain method.
One general aspect in terms of efficiency of a method is the time required
for learning and practice before a method can be applied by the researcher
at all. Part of the purpose of this study was the evaluation of this
learning process. Therefore, one of the authors (MA), who had no previous
experience with 3D surface generation, documented the time and effort
needed. Training and practice are not an issue when using the Witikon
automated photogrammetry setup, because data generation and processing are
completely outsourced to EDICO SK. However, training time plays a major part
when it comes to making a decision on whether to use (micro-)CT scanning or
manual photogrammetry for surface reconstruction.
While with CT scanning, the fundamental data are almost always produced by
specialized staff at scanning facilities, researchers using photogrammetry
mostly produce these data, i.e., the photographs, themselves. Therefore,
photogrammetry requires more learning effort and practice for this very
first step toward a reconstructed 3D surface. Overlap, lighting, and focus
play a major role in basic photogram-metry (i.e. Fahlke et al., 2013; Fahlke,
2014; Mallison and Wings, 2014). Taking photos from the wrong angles with
insufficient overlap or bad lighting (i.e. too bright, hard shadows, etc.)
can lead to difficulties in the subsequent alignment process. Therefore,
practice is needed, and it takes longer to take a suitable set of photos for
a beginner than it does for a professional. Going back and taking another
set of photos because of failed alignment can lead to frustration and
immensely increased data acquisition time, if it is possible at all. Data
acquisition time for P_MA listed in Table 1 is actually the time needed for
the fourth set only. Three previous sets led to failed alignment, so it
would be adequate to estimate the learning and practicing time to be another
three hours of attempted data collection, as well as approximately another
12 hours of attempted alignment before successful application of
photogrammetry. This does not include the time needed to get acquainted with
new camera equipment or settings, since MA, at that point, had already been
using the camera for several weeks.
All programs used for data processing, photo alignment, surface generation,
post-processing, and surface comparisons require thorough introduction.
However, some are easier to use than others. While Amira (for surface
extraction from CT data) and CloudCompare (for surface comparisons) were
perceived as being complex and not particularly intuitive at first, they
provide detailed tutorials or manuals, and/or online support, which actually
made them reasonably easy to use after about a day of training. PhotoScan,
the only photogrammetry software used in this study, proved to be quite easy
to use, once the possible settings are understood. This program makes it
possible for the beginner to find the best settings for their purpose via
trial and error, but to save time, an introduction by someone who is
experienced with this program is recommended. Among the programs used for
post-processing, Geomagic Studio was perceived as the most intuitive.
However, it was the only post-processing program in this study that is not
free. So in case of a tight budget, researchers might want to resort to
either MeshLab or GOM Inspect that might require a slightly longer training
period but will serve the purpose as well.
Other aspects of efficiency such as processing times, costs, and usability
of the data produced shall be discussed in the respective sections on the
individual methods below.
In this study, micro-CT scanning yielded very detailed 3D surfaces, except
for the deletion of less dense material described above. This deletion is
certainly a disadvantage, if completeness is crucial for whatever analysis
the researcher has planned. However, it could be viewed as an advantage, if
one only wanted to reconstruct the original fossil surface without plaster
fillings or glue.
While manual and automated photogrammetry produce 3D surfaces featuring the
photo-realistic surface color of the original cranium, 3D surfaces derived
from (micro-)CT scans are generally not colored. Again, it depends on the
purpose of the study if one perceives missing surface color as a
disadvantage. For mere morphometric measurements or landmark analyses,
uncolored surfaces certainly suffice, but for taphonomic analyses or display
purposes, surface color might be needed, which excludes CT scanning as a
means for surface reconstruction.
In addition to the desired external surface, micro-CT scanning yielded a lot
of surface data that may be unwanted in 3D surface reconstruction: the
surfaces of internal bony structures of the cranium. These additional
surfaces increase file size and are often connected to the external surface,
resulting in holes and intersecting polygons that need to be fixed for
further manipulation of the model, especially for purposes like 3D printing.
Therefore these surfaces of internal structures should be removed, if the
focus is on the reconstruction of the external surface. This removal,
however, is extremely time-consuming, as connections within the outer shell
need to be cropped manually so that partial surfaces can be deleted,
resulting in long post-processing times of 6.3 and 17 hours, respectively
(Table 1). The easiest and fastest way to remove the internal structures
from both of our CT-generated 3D models (CT_MA and CT_JMF) was to clear out
the interior of two halves of the surface and subsequently merge these
halves to form a single surface. This approach was used right away with
CT_JMF. With CT_MA, a hole was cut into the palate first to remove internal
surfaces in the anterior part of the cranium before cutting the 3D model in
half along a frontal plane, and clearing out the partial surfaces as
described above. Surprisingly, the sutures connecting the previously
disconnected anterior and posterior portions of the CT-based models can
hardly be traced on the models themselves and show no deviations from other
3D surfaces in the comparisons (Figures 2 and 3). However, deviations seen
in grooves on the palate in these comparisons are a result of the hole that
was cut into CT-MA in this place and later filled using a generic filling
function in Geomagic Studio.
Of course, small bits of internal surfaces of both models remained, because
these tended to be folded and connected to the external surface in a way
that made complete removal impossible without destroying the external
surface. Furthermore, polygons of CT_MA and CT_JMF were not the same due to
different thresholds used in surface extraction. Therefore, manual cutting
and filling of the holes was not quite identical in both models. The
difference of these internal remainders and filled holes caused the large
standard deviation and greatest distances between these two models (i.e.,
lowered reproducibility between users; Table 3), whereas the visual
comparison of both CT-based 3D surfaces indicates a high degree of
similarity based on the green color of the external surface (Figure 3).
Costs for the micro-CT scan of our particular specimen were 714 Euros at the
time (2012), including tax, but excluding travel costs to the scanning
facility. This makes Micro-CT scanning the most expensive digitization
method in our study.
Reproducibility from different sets of photographs using manual
photogrammetry was very good. Particularly the two surfaces based on a
higher number of photos (P_MA and P_HM with 378 and 150 photos,
respectively) are very similar, having the smallest standard deviation and
smallest extreme distances from one another of all comparisons (Table 3).
P_JMF was less similar to P_MA because of strong positive deviations in the
front and back of the cranium and negative deviations on the dorsal and
ventral sides (Figure 5), meaning that P_JMF was dorsoventrally flattened
compared to P_MA (and, consequently, the other 3D surfaces). This might be a
result of different photo alignment settings: while generic pair selection
was selected in PhotoScan during alignment for P_MA and P_HM, pair selection
was disabled for P_JMF, because photo alignment had failed when using the
generic setting. One reason for alignment difficulties may be the lower
number of images (83 photos) used for the generation of P_JMF, leading to
less overlap. However, in the majority of these photos, the whole specimen,
or at least a large portion of it was captured, so overlap should not be an
issue. Another possible reason may be that the area of overlap of the photos
from the ventral and dorsal perspectives was poorly lit, because it was in
the shade either way, with the cranium dorsal side up and ventral side up,
and no flash light was used while taking these photographs.
The importance of proper lighting, camera settings and overlap for 3D model
generation becomes all the more obvious when details of the surface
comparisons are evaluated. For example, all 3D surfaces that were compared
to P_MA deviated to the negative in the same places on the squamosal bones
(Figures 2, 4-6). We interpret this as P_MA lying above the other surfaces
in these areas due to automatic hole filling in PhotoScan, and attribute
this effect to the mode of lighting used during photography, and possibly
problems with image overlap. The photographs used for the generation of P_MA
were taken with a ring flash and WB set to tungsten light. Thus the specimen
was evenly lit, and hard shadows were eliminated. Additionally the images
were themselves were quite dark, eliminating highlights as well. Therefore,
naturally shaded depressions may have not have been picked up as such.
Furthermore, these areas on the cranium are only captured in a few photos,
making it possible that not enough information for correct surface
reconstruction was gathered due to insufficient overlap.
Taken together, our evaluations suggest that the light, settings and
technique used by HM were most effective: the internal flash with the camera
held upside down for better illumination of the areas where the photos from
the dorsal and ventral perspectives overlap, and automatic WB.
Data collection and post-processing times were much shorter with manual
photogrammetry than they were with micro-CT scanning, and slightly larger
than when automated photogrammetry was applied. Duration of surface
generation from photographs approximately equaled that of surface generation
from micro-CT data (Table 1). While regular office computers were used for
surface generation from CT scans and all post-processing, we used a more
powerful server computer for surface generation from photos in PhotoScan.
This computer provides partitioned RAM for several users at the same time,
providing us with up to 40 GB of RAM at a time. While photogrammetric
surface generation was tested and is well possible on computers with 4 GB
RAM, surface generation time would of course be even longer. It needs to be
kept in mind here though, that photogrammetric surface generation in
Photoscan, once started, does not require interaction until it is finished,
allowing the researcher to accomplish other tasks in the meantime, whereas
the researcher needs to be present and alert during the whole time of image
stack registration and surface extraction from micro-CT scans, making manual
photogrammetry the more efficient method in this regard. Data collection
times and surface generation (calculation) times naturally increase the more
photographs are taken per set (Table 1).
Another advantage of manual photogrammetry is its versatility: no specimen
transport is needed, and photographs can usually be taken inside the
collections that house the respective specimen(s), independent of the
equipment and technology available on-site. For CT-scanning or automated
photogrammetry, either the specimen or the setup (usually quite heavy
equipment) need to be transported, increasing the risk of damage to the
The actual costs of manual photogrammetry can only be roughly estimated. In
our specific case, the costs for the Canon EOS 650D camera including
equipment such as a camera bag, an additional 50 mm lens and extra storage
were 1006 Euros. Experience with photogrammetry at the Museum für Naturkunde
has shown that it is possible to take well over 100,000 photos with this
particular camera model without any signs of wear-out failure. This puts us
roughly at 0.01 Euros per photograph, i.e. at 1.50 Euros for the set
containing the intermediate number of photos (P_HM, 150 photographs), not
including the electricity needed for charging the batteries, etc. So if the
researcher's budget is an issue for data generation, as it is oftentimes the
case, especially for student researchers, manual photogrammetry would be the
means of choice.
The 3D surface resulting from automated photogrammetry (WITI) was very
similar to those generated with manual photogrammetry. In order to have the
same reference surface in most comparisons, WITI was compared with P_MA
here, resulting in the smallest mean distance recorded in all comparisons
(Figure 6, Table 3). Further comparisons (not shown here to avoid
redundancy) revealed, however, that WITI was actually most similar to P_HM
among all used 3D surfaces. This result suggests that manual and automated
photogrammetry produce similarly accurate 3D surfaces.
Surface resolution is extremely high when using MDS Witikon for surface
reconstruction (Table 1), which is a result of the high-resolution cameras
used in the setup. On the one hand, this high resolution makes for
stunningly detailed object panoramas, the original purpose of this method.
On the other hand, as stated above, 3D surface files generated from these
photos are much too large to be handled with regular office equipment, so
that polygon count needs to be reduced anyway, making this method actually
slightly less efficient for 3D surface generation than manual photogram-metry.
Data collection time is lowest in automated photogrammetry, because the
automatic setup ideally does not require human interaction or correction
during photography. Also, surface generation time is very low (Table 1),
because computational requirements are greatly reduced when camera positions
are fixed compared to a handheld camera that changes positions between
The regular minimum costs for the generation of one object panorama with the
Witikon technology are 135 Euros. In case of a 3D model consisting of a
dorsal and a ventral set of photos, costs would amount to 270 Euros, which
is less than half the price of the micro-CT scan, but still almost two
hundred times as much as what is needed for manual photogram-metry. One
needs to bear in mind, however, that the costs for using MDS Witikon include
complete data processing, saving the researcher valuable time, and thus
making this the most convenient method for the researcher.