Conference PaperPDF Available

Computer-Assisted Reconstruction of Virtual Fragmented Cuneiform Tablets


Abstract and Figures

Cuneiform is one of the earliest known systems of writing consisting of wedge-shaped strokes forming signs impressed on clay tablets. Excavated cuneiform tablets are typically fragmented and their reconstruction is, at best, tedious but more often intractable given that fragments can be distributed within and between different collections. Digital archives relevant to cultural heritage, such as the Cuneiform Digital Palaeography Project and the Cuneiform Digital Library Initiative, now make richly annotated artefact media available to wide populations. Similarly, developments in computer-aided three-dimensional reconstruction methods offer the potential to assist in the restoration of broken artefacts. In this paper, a system for providing computer assistance for identifying and orientating matching fragment pairs is described. Metrics to grade the quality and likely significance of potential matches are also described. Finally, the results from experiments with scans of laboratory fabricated tablet fragments and genuine fragmented cuneiform tablets are reported.
Content may be subject to copyright.
This paper is a preprint (IEEE “accepted” status).
IEEE copyright notice. c
2014 IEEE. Personal use of this material is permitted. Permission from IEEE must
be obtained for all other uses, in any current or future media, including reprinting/republishing this material for
advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists,
or reuse of any copyrighted component of this work in other works.
Computer-Assisted Reconstruction of Virtual
Fragmented Cuneiform Tablets
Tim Collins
Sandra I. Woolley
Luis Hernandez Munoz
Electronic, Electrical and Systems Engineering,
University of Birmingham, Edgbaston,
Birmingham, B15 2TT, UK
Andrew Lewis
Digital Humanities Hub,
University of Birmingham, Edgbaston,
Birmingham, B15 2TT, UK
Eugene Ch’ng
School of Computer Science
University of Nottingham Ningbo China
199 Taikang East Road, Zhejiang Ningbo,
315100 China
Centre for Creative Content and Digital Innovation,
University of Malaya,
50603 Kuala Lumpur, Malaysia
Erlend Gehlken
Institut f¨
ur Arch¨
aeologische Wissenschaften
at, Gr¨
uneburgplatz 1
60629 Frankfurt/Main, Germany
Abstract—Cuneiform is one of the earliest known systems
of writing consisting of wedge-shaped strokes forming signs im-
pressed on clay tablets. Excavated cuneiform tablets are typically
fragmented and their reconstruction is, at best, tedious but more
often intractable given that fragments can be distributed within
and between different collections. Digital archives relevant to
cultural heritage, such as the Cuneiform Digital Palaeography
Project [1]–[4] and the Cuneiform Digital Library Initiative [5],
[6], now make richly annotated artefact media available to wide
populations. Similarly, developments in computer-aided three-
dimensional reconstruction methods offer the potential to assist
in the restoration of broken artefacts. In this paper, a system
for providing computer assistance for identifying and orientating
matching fragment pairs is described. Metrics to grade the quality
and likely significance of potential matches are also described.
Finally, the results from experiments with scans of laboratory
fabricated tablet fragments and genuine fragmented cuneiform
tablets are reported.
KeywordsComputational geometry, Iterative methods, Cost
function, Search problems, Computer graphics
Unlike jigsaw puzzles of thousands of pieces, which com-
puters can now easily solve [7], artefact fragments such as
broken pottery (potsherds) or cuneiform tablet fragments are
more complex three-dimensional objects and may belong to
an unknown number of complete or incomplete “puzzles”
whose surviving pieces may be eroded and may no longer fit
well together. Computer-aided reconstruction of archaeological
fragments has been an active area of research in recent years
although most published work has been specific to the joining
of potsherds [8]. A survey of computational reconstruction
The authors gratefully acknowledge the support of The Leverhulme Trust
research grant (F000 94 BP), and the multidisciplinary support of the Univer-
sity of Birmingham’s Digital Humanities Hub.
methods [9] observes that while approaches are sophisticated
and computationally efficient, none claim to be ready for
deployment as archaeological tools. Impediments to automated
reassembly, aside from the practical difficulties associated with
obtaining three-dimensional scan sets, include the extremely
difficult search problems, the lack of surface information inclu-
sion with object geometry and, significantly, the resolution of
issues associated with large numbers of false-positive matches.
More recently, a hybrid human-computer approach has been
proposed to refine computer and archaeological knowledge
bases with geometry, texture and feature knowledge to resource
match performance [10]. Tested on sculptural fragments this
approach demonstrated good performance.
Adopting a hybrid human-computer approach, we have
designed a virtual collaborative environment for the recon-
struction of cuneiform tablets [11]. A complete description
of the environment is detailed in a separate publication [12].
The design involves computer assistance on a number of
levels from simple quantitative join-quality feedback and semi-
automated alignment of nearby fragments (a form of “snap to
best fit”) through to fully automated searching and joining of
fragments. Though, even in the latter case, the computer assists
human effort; human intervention is necessary to validate
candidate matches and eliminate false-positives.
Cuneiform fragment matching differs from problems such
as mosaic or potsherd matching in that tablet fragments have
significant free-form variation in shape in all three dimensions.
However, to reduce the search problem, some assumptions
can be made about the expected size, shape, texture and
surface curvature properties of the completed tablet. In making
such assumptions, we can create additional criteria for the
assessment of match quality beyond those generally applied
in free-form 3D reconstruction algorithms [13]–[15].
We propose several criteria for match quality assessment:
the distances between pairs of sampled points on the adjoining
surfaces, the interior area of the adjoining surfaces, the degree
of interlocking of the join, the continuity of the written surfaces
either side of the join, and the continuity of the slope of
the written surfaces either side of the join. These criteria are
used both for iteratively finding optimal matching fragment
orientations and also for ranking candidate matches; false
positives should be rejected and the most promising joins given
priority when presenting to users for validation. Typically, in
previously published reconstruction algorithms, match quality
is assessed by a more limited set of criteria such as measures of
inter-surface distances and size of overlapping area [14], [15]
but we have found, in this application, that the introduction
of additional criteria aids in the prioritisation process required
prior to human validation.
When a match is attempted between a pair of fragments,
an iterative process is used to find the optimal translations
and rotations needed to minimise a cost function designed to
quantify the match quality. This process is repeated across
a search-grid of initial conditions in order to guarantee the
globally optimal match is identified. During the iterations,
the translations and rotations are applied to the model-view
matrix used by the graphical processing unit (GPU) of the
computer. This efficiently yields uniformly sampled depth
maps of the fragment from the irregularly sampled model data
of the archived scan [16]. The cost function to be minimised is
calculated through analysis of the depth maps of the two frag-
ment surfaces. Although the use of GPU-generated projective
depth maps has been reported in previously published related
work [14], [17], the iterative algorithm and the associated cost
function presented here are novel and, while developed for
cuneiform fragments, could be applicable for other fragmented
A. Pre-processing
Upon attempting to match a pair of fragments, six degrees
of movement must be considered. If one of the fragments
is considered to be fixed in space, the other may require
translation in three dimensions as well as rotation by the three
Euler angles in order to form a join. One of the degrees
of freedom can be eliminated if a restriction is imposed
that in any potential solution, the fragments must touch one
another but must not intersect. With five degrees of freedom
remaining, the search space is still prohibitively large for an
exhaustive search and an iterative approach is required. Due
to the intricate geometry of fragment surfaces, convergence
to a globally optimal solution cannot be easily guaranteed
so multiple iterations must be performed using a search grid
of different initial conditions. The first part of our algorithm
involves some pre-processing applied to the fragment scans in
order to reduce the scale of this search.
When a tablet fragment is scanned, the first part of our
process is to calculate the minimum-volume oriented bounding
box [18] of the vertices and to reorientate the vertex coordi-
nates so that the bounding box is aligned parallel with the x,
yand zaxes. For most cuneiform fragments of interest, we
expect at least one of the surfaces to contain script and, after
Fig. 1. Two virtual cuneiform tablet fragments with oriented bounding boxes.
The fragments, W 18349 (above) and Wy 777 (below) [19], are part of a letter
written in the Neo-Babylonian empire (626-539 BC, around the time of King
Nebuchadnezzar) and were found in the ancient city of Uruk (near modern-day
Basra, Iraq).
reorientation, this surface, being relatively flat, will be roughly
parallel with one side of the bounding box. This side can be
identified using the statistical properties of the surface textures
or, more reliably, manually at the time that the fragment is
scanned. After identification the fragment is rotated so that
the inscribed surface faces forwards as illustrated by the two
fragment scans shown in figure 1.
After these rotations, only the non-script sides of the
fragment are considered for potential matches. Furthermore,
when considering a pair of fragments, only the orientations of
the boxes with the inscribed surfaces facing forwards will be
considered. This reduces the number of possible permutations
to consider from 144 (each of the six faces of one box
presented to each of the six faces of the other with four possible
rotations of 0,90,180and 270) to just 16 (four faces of
one box presented to four faces of the other with no rotations
allowed). Also, because it is usually possible to determine the
direction of writing on the fragments, the number of possible
combinations can be further reduced to just four.
B. Pair-wise Matching
For each pair of faces, one of the fragments must be
translated and rotated to find the best matching position. Three
rotations are allowed corresponding to the three Euler angles.
Translations are made in the left-right direction (xoffset) and
in the front-back direction (yoffset). An up-down translation (z
offset) is also made but this is calculated from the rotations and
the xand yoffsets such that the restriction that the fragments
must touch but not intersect is satisfied. There are, therefore,
only five independent positional parameters (three angles and
two offsets) that require optimisation.
Fig. 2. Example illustration of the depth map calculation performed by the GPU. (left) Depth map of the top face of the first fragment, (centre) depth map
of the bottom face of the second fragment, (right) joined fragments and their summed depth map, z, after min(z) has been subtracted. The uniform, near zero,
depth values of the summed depth map indicate a constant distance between point-pairs and, therefore, a good quality match.
The approach we have adopted to optimise these five pa-
rameters for a given pair of fragments is based on the analysis
of two-dimensional projective depth maps of the surfaces to
be joined. The calculation of depth maps (also known as
z-buffers) is a part of the hidden-surface removal algorithm
performed whenever the GPU of a computer renders a three-
dimensional object. GPUs contain specialised processing units
within their architecture to efficiently perform the geometric
calculations needed to generate a z-buffer and the results can
be accessed via OpenGL [16]. This allows us to generate
a uniformly sampled grid of depths of the surface of the
join from the irregularly sampled vertex data of the object
model. In operation, the vertex and face data of each fragment
are stored in array buffers in GPU memory. Depth maps are
then generated on each iteration by manipulating the model-
view matrix of one fragment before rendering the depth data
to an off-screen framebuffer object. Similar approaches have
been reported in previous work [14], [17] but the subsequent
analysis of the depth map data presented here is novel.
Due to the intricate geometry of the surfaces of the join,
convergence to the globally optimal solution cannot be guar-
anteed by iteratively optimising all five of the rotation and
translation parameters. However, we have found that if the
xand yoffsets are fixed, the three rotations can be reliably
estimated iteratively. This is due to the orientation applied
during the pre-processing which ensures that at least two of
the three angles are relatively close (within a few degrees) to
the optimal solution because of the approximate alignment of
the planes of the inscribed surfaces.
Although it is still necessary to systematically search
through a two-dimensional grid of possible xand yoffsets,
the reduction in computational complexity from the original
five-dimensional search space is significant. For each potential
x, y offset in the search, the iterative estimation of the optimal
rotations uses a Nelder-Mead optimisation [20]. The optimi-
sation aims to minimise a cost function which is calculated
at each iteration using the algorithm outlined in the following
function cost(x, y, φ, θ, ψ )
fragment0.rotate(φ, θ, ψ)
fragment0.translate(x, y)
zfragment0.depthmap() + fragment1.depthmap()
// Bring fragments into contact without intersecting
// Calculate average cost function of all sampled points
for n= 0 to N1do
end for
return c/N
xand yare the offsets which are fixed for each optimisation;
φ, θ and ψare the Euler rotation angles which are to be opti-
mised. The GPU calculates the depth maps for each fragment
by rendering an orthographic projection forming a uniform
grid of sampled distances from a fixed observation plane to
the surface of the fragment. The rotations and translations are
applied to the model-view matrix of one of the fragments prior
to generating its depth map. The two depth maps are summed
and the minimum summed depth value subtracted from all
depths, thus satisfying the condition that the fragments must
touch but not intersect. For each element of the summed depth
map, z, the residual offset, zn, is converted to a cost-per-
sample by the function f(zn)to assess the goodness of fit.
This function is averaged over the Ndepth samples to give
the overall cost function. Examples of the depth maps formed
from scanned laboratory fabricated tablet fragments are shown
in figure 2 and the geometry of this process is illustrated by the
side view in figure 3. Note that in both figures the fragments
form very good matches and the summed depth map shown in
figure 2 has a near-uniform value close to zero over the entire
surface of the join.
The reliability of the offset search and of the iterative
rotation optimisation are highly dependent on the function,
f(z). Examples of functions used in similar applications
include the Mean Square (MS) of the distances [21] and the
Largest Common Pointset (LCP) metric [14]. Our initial testing
Fig. 3. Side view of the depths used to calculate the cost function. Note that
for a perfect match, all depth-pairs sum to the same value.
Fig. 4. The piecewise linear cost function used in our experiments. The
non-intersecting constraint means that zcan never be negative, hence only
positive values are shown.
with laboratory fabricated tablet fragments gave unsatisfactory
results with these methods. The MS cost function did not
perform well when a significant proportion of the points in
the depth maps cannot meet due to chips and erosion of the
surfaces. Such features are over-emphasised by the squaring
process and lead to poor convergence and large numbers of
false-positive matches. The LCP approach avoids this problem
but requires a relatively fine depth map resolution to converge
reliably. A linear cost function was also employed to attempt
to suppress the problems encountered with the mean square
approach and gave improved results but still had a tendency to
produce false-positive matches. The cost function we propose
as an alternative is a Piecewise Linear (PL) function designed
as a compromise between a linear function and the LCP
approach. The function, f(z), for the PL function is of the
f(z) = zfor z
γz +(1 γ)for z >  (1)
where is a threshold distance equivalent to the one used in
the LCP metric [14] and γ1is a constant. The form of
the function is illustrated in figure 4. In our experiments, the
values used were = 2 units (1 mm) and γ= 0.025.
The advantage of the PL cost function over the MS and
linear cost functions is that the impact of point-pairs that are
greater than apart is reduced. The reasoning is that these
point-pairs are so far apart that they are not, actually, part of the
matching surface and should not count towards the matching
metric. The small gradient, γ, helps to ensure convergence
when large numbers of points begin the optimisation with a
distance greater than . The advantage of the PL approach
over LCP is that the linear portion of the function in the
region zhelps to ensure convergence and suppresses false-
positives, especially when the depth map is coarsely sampled.
Figure 5 compares the four cost functions under consideration
using the pair of fragments shown in figure 1. All except the
proposed PL function exhibit significant false-positive peaks.
All of the plots in figure 5 exhibit multiple local minima
making convergence to the optimal solution using an iterative
process difficult to guarantee. Our approach is to locate the
optimal offset using a multi-resolution search algorithm. The
first pass searches the entire range of possible offsets over a
uniform grid of sample points with a relatively coarse 0.5 mm
separation. Subsequent passes reduce the extent of the search
to one quarter of the previous area, centred on the best result
from the previous pass, and refine the resolution by a factor
of two so the total number of sample points remains the same.
This process continues down to a resolution of approximately
10 microns. Subsequent refinement has not demonstrated ob-
servable improvement in the join with any of the fragment pairs
tested to date. This search process is outlined in the following
function search(xrange , yrange )
// Set initial centre of search and search resolution, d
Nxxrange ÷2d
Nyyrange ÷2d
while d > dmin do
for n=Nxto Nxdo
for m=Nyto Nydo
costn,m =solve(x, y)
end for
end for
(nmin, mmin )arg min
// Set search centre for next pass
end while
return (x0, y0)
where xrange and yrang e define the extent of the search area,
dinit is the search resolution for the first pass (0.5 mm in our
case) and the function, solve(x, y), performs the Nelder-Mead
iterative rotational optimisation, using offsets xand y, and
returns the cost function for the optimal solution.
C. Match Ranking
The pair-wise matching technique described in the previous
section can be applied on a number of levels from a semi-
automated alignment of nearby fragments (a form of “snap to
best fit”) through to fully automated searching and joining of
fragments. Regardless of the application, human intervention
will always be required following matching in order to validate
the computer generated match and eliminate false-positives.
Efforts to aid this process have already been described in
section II-B in the shape of the novel proposed piecewise-linear
cost-function. This has been shown to significantly reduce the
incidence of false-positive matches.
Fig. 5. The four cost functions under consideration shown as normalised functions of xand yoffset. Piecewise Linear (PL), Linear (L), Largest Common
Pointset (LCP) and Mean Squares (MS) functions are shown.
Fig. 6. (i) Small overlapping join area (only four points are closer than ): a
low priority match. (ii) Larger join area: likely to be of greater significance.
When the fully automated matching system is used, it is
desirable to prioritise the potential matches so that the human
validator can examine the most promising results first. Ranking
different potential matches could be done simply on the basis
of the final cost function of the join but this does not take
into account other factors that could indicate that one matched
pair is a more significant candidate than another and is worth
investigating with a higher level of priority. In addition to the
measure of the distance between pairs of points on the depth
maps (section II-B), the quality or significance of a match can
also be graded based on the following criteria:
the overlapping area of the adjoining surfaces
the degree of interlocking of the join
the continuity of the written surfaces either side of the
the continuity of the slope of the written surfaces
either side of the join
The overlapping area of the adjoining surfaces is estimated by
the number of points in the summed depth maps that have a
separation less than the threshold, . This metric is the same
as the common pointset size as used in the LCP cost function.
The greater the number of points, the larger the overlapping
area and, therefore, the more significant the join is likely to
be. This estimation is illustrated in figure 6.
The degree of interlocking of a join is estimated by fixing
the rotation angles, φ, θ, ψ, at their optimal values but moving
the offsets, xand y, by small amounts in each direction, i.e.
xx±δx and yy±δy (see figure 7). The average
increase in the cost function caused by these small translations
Fig. 7. (i) A join with tight interlocking before and after the translation by
δx. (ii) A loosely interlocking join; the average depth after translation is much
smaller in this case.
corresponds to the degree of interlocking; the tighter the join is
interlocked, the larger the average increase in the cost function
will be.
The last two criteria estimate the degree of any discon-
tinuity of the newly joined inscribed surface in terms of the
surface depth (figure 8i) and in the surface gradient (figure 8ii).
These discontinuities are jointly estimated by taking a second
depth map, this time of the joined tablets together viewed from
the front, and calculating the average value of the second-
derivative of depth with respect to distance. For a perfectly
smooth tablet and a perfect join, this will be zero across the
entire exterior surface.
The metrics described in this section have been tested using
scans of the real tablet fragments shown in figure 1 as well as
scans of laboratory fabricated fragments such as those shown
in figure 2. Unsurprisingly, the most significant predictor of
match quality is the cost function used for the iterative rotation
optimisation. Even when a join is only partial due to damage to
the tablet fragments, the cost function still returns a low value
for correct joins by virtue of the piecewise-linear cost function.
The next most significant metric is the overlapping area (the
common pointset size). This does not give extra information
about the quality of the join but does give an indication of
how significant the match is likely to be.
The degree of interlocking does not provide information
about the quality or likely significance of a match but does
provide information about the likelihood that a match is a false-
positive. Smoother surfaces with low degrees of interlocking
such as the edges of a tablet are more likely to find relatively
well fitting false-partners than more detailed surfaces that
interlock tightly. As such, when a tightly interlocking match
is discovered it is more likely to be a true-match and should
be given a higher level of priority.
Surface continuity has proved to be the least significant of
the metrics under consideration. This is due, in part, to the
problems in reliably estimating this metric given the irregular
nature of the tablet surface. However, with further refinement
to the algorithm used to calculate this metric, we believe it
Fig. 8. (i) A join exhibiting a discontinuity in surface depth. (ii) A join
exhibiting a discontinuity in surface gradient despite being continuous in
surface depth. (iii) A continuous join in terms of both depth and gradient.
will be possible to form a more robust estimator of surface
continuity, helping to identify false positive matches that have
escaped detection using the other criteria.
The method and cost function described have been de-
veloped and tested using laboratory fabricated clay tablet
fragments. Figure 9 shows an example of a test tablet that had
been broken into four fragments before 3D scanning and pro-
cessing using the method described. In this case, the matches
were chosen based on the cost function alone and the tablet
completely reconstructed without human intervention. Note
that even though the join between the top two fragments is
quite badly chipped, the optimal orientation was still estimated
correctly. This is possible because of the use of the piecewise
linear cost function which places a reduced emphasis on the
join area suffering from the chipping when compared with the
linear or least squares functions.
Scans of the real tablet fragments shown in figure 1 have
also been tested. The resulting matched fragment pair is shown
in figure 10. This is the first example of a pair of 3D scanned
cuneiform tablet fragments to be successfully joined by an
automated algorithm with no human intervention beyond the
initial scanning and pre-processing stages.
This paper presented a proposed 3D matching algorithm
designed specifically for fragmented cuneiform tablets. The
novel piecewise linear cost function employed by the algorithm
has proven successful in matching scans of both laboratory
fabricated test fragments as well as fragments of real cuneiform
Fig. 9. A laboratory fabricated clay tablet, broken into four fragments,
scanned and completely reconstructed using the proposed matching process.
tablets. This is the first reported example of successful auto-
mated joining of scanned cuneiform tablet fragments.
Several metrics, in addition to the cost function, have
been proposed to aid in the ranking of potential matches.
Although the reconstruction shown in figure 9 was performed
using the cost function alone, this was only possible because
of the small number of fragments involved. With a larger
library of fragment scans to draw from, the likelihood of
false positives and the need for human validation becomes
increasingly important. This is when the additional match
quality criteria become valuable because they allow joins to
be ranked not only by the match quality, but also by the likely
significance of the match.
Presently, although the fragment rotations are calculated
by an iterative optimisation process, the fragment offsets are
established by searching over a uniform grid. We plan to
improve the efficiency of the process by further optimisation
of the xy offset search using a second layer of iterative
optimisation (possibly employing some of the additional match
Fig. 10. The scanned tablet fragments, Wy 777 and W 18349 [19],
successfully joined using the proposed piecewise linear cost function.
quality criteria). We also plan to acquire further scans of real
tablet fragments in order to further refine the algorithm and
validate the importance of the match quality metrics.
Although further refinements are planned, even in its cur-
rent form the proposed matching algorithm has the potential
to benefit the Assyriology research community. There are
huge numbers of unmatched fragments in museums around the
world and, without an automated approach, the vast majority
of these would probably never be joined.
For the data acquisition with the 3D scanner of the
Heidelberg Graduate School of Mathematical and Computa-
tional Methods for the Sciences at the Interdisciplinary Center
for Scientific Computing, Heidelberg University, we would like
to thank Sonja Speck, Hubert Mara and Susanne Kr¨
The scanned tablets are published with the kind permission
of the German Archaeological Institute, Berlin.
The pre-processing computations described in this paper
were performed using the University of Birmingham’s
BlueBEAR HPC service, which provides a High Performance
Computing service to the University’s research community.
See for more details.
For assistance and the firing of the laboratory fabricated test
tablets, we would like to thank The Potters Barn, Sandbach,
Cheshire, UK (
[1] The Cuneiform Digital Palaeography Project. [Online]. Available:
[2] S. I. Woolley, T. R. Davis, N. J. Flowers, J. Pinilla-Dutoit, A. Living-
stone, and T. N. Arvanitis, “Communicating cuneiform: The evolution
of a multimedia cuneiform database,” The Journal of Visible Language,
Special Edition on Research in Communication Design, vol. 36, no. 3,
pp. 308–324, 2002.
[3] S. I. Woolley, N. J. Flowers, T. N. Arvanitis, A. Livingstone, T. R.
Davis, and J. Ellison, “3D capture, representation and manipulation of
cuneiform tablets,” Proc SPIE (Three Dimensional Image Capture and
Applications IV), vol. 4298 (0277), pp. 103–110, 2001.
[4] T. N. Arvanitis, T. R. Davis, A. Livingstone, J. Pinilla-Dutoit, and S. I.
Woolley, “The digital classification of ancient near eastern cuneiform
data,” British Archaeological Review, BAR International Series, p. 1075,
[5] CDLI (Cuneiform Digital Library Initiative). [Online]. Available:
[6] A. Lewis and E. Ch’ng, “A photogrammetric analysis of cuneiform
tablets for the purpose of digital reconstruction,” International Journal
of Heritage in the Digital Era, vol. 1, no. 1, pp. 49–53, 2012.
[7] A. Jacob, “Jigsaw-solving computer kicks puny humans aside,” New
Scientist, vol. 214, no. 2869, p. 24, 2012.
[8] M. Kampel and R. Sablatnig, “3D puzzling of archeological fragments,”
in 9th Computer Vision Winter Workshop Piran, Slovenia, 2004, pp. 31–
[9] A. R. Willis and B. D. Cooper, “Computational reconstruction of ancient
artifacts,” IEEE Signal Process. Mag., vol. 25, no. 4, pp. 65–83, 2008.
[10] A. Ad ´
an, S. Salamanca, and P. Merch´
an, “A hybrid human-computer
approach for recovering incomplete cultural heritage pieces,Computers
and Graphics, vol. 36, no. 1, pp. 1–15, 2012.
[11] E. Ch’ng, A. Lewis, R. E. Gehlken, and S. I. Woolley, “A theoretical
framework for stigmergetic reconstruction of ancient text,” in Visual
Heritage in the Digital Age. Springer, 2013, pp. 43–65.
[12] E. Ch’ng, S. I. Woolley, E. Gehlken, A. Lewis, L. H. Munoz, and
T. Collins, “The development of a collaborative virtual environment
for 3D reconstruction of cuneiform tablets,” in 20th International
Conference on Virtual Systems and Multimedia (VSMM), 2014.
[13] Q.-X. Huang, S. Fl ¨
ory, N. Gelfand, M. Hofer, and H. Pottmann,
“Reassembling fractured objects by geometric matching,” ACM Trans-
actions on Graphics (TOG), vol. 25, no. 3, pp. 569–578, 2006.
[14] C. S. Belenguer and E. V. Vidal, “Archaeological fragment character-
ization and 3D reconstruction based on projective GPU depth maps,
in 18th International Conference on Virtual Systems and Multimedia
(VSMM), 2012, pp. 275–282.
[15] S. Winkelbach and F. M. Wahl, “Pairwise matching of 3D fragments
using cluster trees,” International Journal of Computer Vision, vol. 78,
no. 1, pp. 1–13, 2008.
[16] OpenGL software development kit. [Online]. Available:
[17] G. Papaioannou, E.-A. Karabassi, and T. Theoharis, “Virtual archaeolo-
gist: Assembling the past,” Computer Graphics and Applications, IEEE,
vol. 21, no. 2, pp. 53–59, 2001.
[18] J. O’Rourke, “Finding minimal enclosing boxes,” International journal
of computer & information sciences, vol. 14, no. 3, pp. 183–199, 1985.
[19] E. Gehlken, S. I. Woolley, T. Collins, and E. Ch’ng, “Automated joining
of cuneiform tablet fragments,” Nouvelles Assyriologiques Br`
eves et
Utilitaires [NABU], awaiting publication, 2014.
[20] J. A. Nelder and R. Mead, “A simplex method for function minimiza-
tion,” The Computer Journal, vol. 7, no. 4, pp. 308–313, 1965.
[21] G. Palmas, N. Pietroni, P. Cignoni, and R. Scopigno, “A computer-
assisted constraint-based system for assembling fragmented objects,” in
Digital Heritage International Congress. The Eurographics Associa-
tion, 2013, pp. 529–536.

Supplementary resource (1)

... As part of the Virtual Cuneiform Tablet Reconstruction Project, [6], [7] virtual reconstructions have been reported for, typically, small form factor tablet fragments collocated within individual collections [8], [9]. This paper provides an account of the computational aspects of model acquisition and join geometry for the virtual reconstruction of fragments of a different form factor, in different collections, located in different countries. ...
... The overall project aim of virtual cuneiform tablet reconstruction [6] has been investigated using an online collaborative crowd-sourced approach [17], [18] as well as an automated joining process [8]. The automated process was used for the Atrahasis fragments because it reliably ensures that the optimal matching orientation is found and also yields statistics indicating the goodness of the fit. ...
... A piecewise linear mapping function is used to calculate the cost of candidate orientations and displacements. This mapping function helps to ensure that the correct optimal orientation is found even when the two fragment faces are incomplete and eroded [8]. The best result from these matching attempts is saved and presented as the result for human-expert verification. ...
Conference Paper
Full-text available
The epic of Atrahasis is one of the most significant pieces of ancient Mesopotamian literature. The account has survived millennia on sets of clay tablets inscribed with cuneiform script; a sophisticated early writing system comprising signs formed from wedge-shaped impressions. The third tablet belonging to one of the most complete copies of the Atrahasis epic is broken. For over fifty years, one fragment, held in Geneva, was believed to join with another held in London. However, due to their 1000 km separation, the join had never been physically tested. This paper contributes a technological account of the successful virtual joining of the fragments, the first ever long-distance virtual join of its type.
... Virtual reconstruction of the tablets obviates issues such as geographical distance as well as practical issues concerning the necessarily limited accessibility and the physical fragility of the fragments [10]. It also makes possible the use of computer-automated matching tools [11]. ...
... The open source Clustering Views for Multi-View Stereo (CMVS) algorithm [34,35] is used to reconstruct a dense 3D point-cloud from each set of photographs. The point density (typically between 100 and 300 points per square millimetre) is sufficient to resolve the features of the impressed cuneiform script and has proven more than adequate for the purpose of matching fragmented tablets [11]. There are other dense point-cloud reconstruction applications available, but CMVS was chosen for our application because it is free, has a well-documented command line interface making it easily automated, and produced comparable results, in our tests, to commonly recommended alternatives such as Photoscan by Agisoft (St. ...
Full-text available
The photogrammetric acquisition of 3D object models can be achieved by Structure from Motion (SfM) computation of photographs taken from multiple viewpoints. All-around 3D models of small artefacts with complex geometry can be difficult to acquire photogrammetrically and the precision of the acquired models can be diminished by the generic application of automated photogrammetric workflows. In this paper, we present two versions of a complete rotary photogrammetric system and an automated workflow for all-around, precise, reliable and low-cost acquisitions of large numbers of small artefacts, together with consideration of the visual quality of the model textures. The acquisition systems comprise a turntable and (i) a computer and digital camera or (ii) a smartphone designed to be ultra-low cost (less than $150). Experimental results are presented which demonstrate an acquisition precision of less than 40 µm using a 12.2 Megapixel digital camera and less than 80 µm using an 8 Megapixel smartphone. The novel contribution of this work centres on the design of an automated solution that achieves high-precision, photographically textured 3D acquisitions at a fraction of the cost of currently available systems. This could significantly benefit the digitisation efforts of collectors, curators and archaeologists as well as the wider population.
... There is additional evidence to support the idea that GPU processing using CUDA technology may be an effective tool for reconstruction tasks at the desktop level (Boyer, El Baz, & Elkihel, 2012; Onsjo & Aono, 2009; Schatz & Trapnell, 2007). Most recently, (Collins et al., 2014) describe a novel method of automatic reconstruction, developed specifically for use with cuneiform fragment models. This method relies on several techniques to narrow down the list of candidate matches for a fragment pair. ...
... The research outlined in (Collins et al., 2014) indicates that further development into matching that considers presorting of fragments by different criteria could lead to a computationally lighter heuristic system for automatic matching. If this research is realised, then the potential for artificial agents to interact with the existing system using notes or stigmergic interaction could be realised, ...
Full-text available
Reducing the time spent by experts on the process of cuneiform fragment reconstruction means that more time can be spent on the translation and interpretation of the information that the cuneiform fragments contain. Modern computers and ancillary technologies such as 3D printing have the power to simplify the process of cuneiform reconstruction, and open up the field of reconstruction to non-experts through the use of virtual fragments and new reconstruction methods. In order for computers to be effective in this context, it is important to understand the current state of available technology, and to understand the behaviours and strategies of individuals attempting to reconstruct cuneiform fragments. This thesis presents the results of experiments to determine the behaviours and actions of participants reconstructing cuneiform tablets in the real and virtual world, and then assesses tools developed specifically to facilitate the virtual reconstruction process. The thesis also explores the contemporary and historical state of relevant technologies. The results of experiments show several interesting behaviours and strategies that participants use when reconstructing cuneiform fragments. The experiments include an analysis of the ratio between rotation and movement that show a significant difference between the actions of successful and unsuccessful participants, and an unexpected behaviour that the majority of participants adopted to work with the largest fragments first. It was also observed that the areas of the virtual workspace used by successful participants was different from the areas used by unsuccessful participants. The work further contributes to the field of reconstruction through the development of appropriate tools that have been experimentally proved to dramatically increase the number of potential joins that an individual is able to make over period of time.
... These can otherwise only be identified as matching by a handful of world experts in cuneiform. The Virtual Cuneiform Tablet Reconstruction Project (VCTR) joined 3D scanned cuneiform tablet fragments automatically using a novel matching algorithm with measure of fit metrics which dramatically reduce false positive match reports (23). The matching algorithm works by iteratively finding the optimal relative orientation of the two fragments under consideration in three-dimensional space. ...
The main source of information regarding ancient Mesopotamian history and culture are clay cuneiform tablets. Despite being an invaluable resource, many tablets are fragmented leading to missing information. Currently these missing parts are manually completed by experts. In this work we investigate the possibility of assisting scholars and even automatically completing the breaks in ancient Akkadian texts from Achaemenid period Babylonia by modelling the language using recurrent neural networks.
... However, apart from pockets of research delving into imaging and storage databases, there appears to be a lack of progress in the area. In our research, we attempted to solve the problem using two different approaches: 1) a computer assisted crowd-source approach, and 2) an automated approach which made use of algorithms that reduce the search space as well as minimise the required level of human intervention by reducing the probability of generating falsepositive matches (details are in two separate publications [1] and [2]). ...
Conference Paper
Full-text available
The efficient reconstruction of ancient cuneiform tablets has been a challenging problem over many decades, which if an approach is found to speed up the rate of potential fragment joins, will lead to a discovery of rich historical records of knowledge inscribed 5,000 years ago. This paper describes The Leverhulme Trust funded development of a crowd-sourcing collaborative virtual environment via the identification of cooperative personalities. We conducted experiments in which 12 groups of 3 participants collaborated to reconstruct fragments in a laboratory 3D puzzle virtual environment. Personality types were analysed together with NASA TLX workloads and task performance. We discovered that patterns exist for each of the virtual interaction modalities, demonstrating a potential for profiling systems to group personality types for maximising the effective reconstruction of cuneiform fragments.
The main sources of information regarding ancient Mesopotamian history and culture are clay cuneiform tablets. Many of these tablets are damaged, leading to missing information. Currently, the missing text is manually reconstructed by experts. We investigate the possibility of assisting scholars, by modeling the language using recurrent neural networks and automatically completing the breaks in ancient Akkadian texts from Achaemenid period Babylonia.
Digital Humanities nowadays gain more and more importance in analyzing texts in new and interesting ways. At the same time, Digital Humanities become recognized by more and more faculties. One example of the application of the Digital Humanities and of computer linguistics can be found in the field of Archaeology. For archaeologists, a decent way of language processing of ancient texts is essential to their work. Many artifacts from ancient civilizations are in need of a textual analysis. Yet many of them have not been analyzed because of a lack of time and human resources or a lack of an automated assisting process for textual analysis. Optimizing the natural language processing chain for an archaeologist can therefore benefit this whole area of expertise. However, in order to analyze texts using a computer, texts have to be broken down to smallest analyzable units, so-called tokens e.g. words. While this has proven to be an easy task to do in European languages like German, French or English (words are usually separated by whitespaces), in non-alphabetic languages like Chinese, Japanese or Korean, word segmentation is a big initial obstacle in analyzing text using a computer. Words in those lan- guages are not clearly separated by distinctive stop chars such as whitespaces. To deal with this obstacle, rulebased, dictionarybased and statistical approaches have been developed for the Chinese and Japanese language. However, to this date no such approaches are known for other nonalphabetic-languages like cuneiform oder Egyptian Hieroglyphs. This Master Thesis applies Chinese and Japanese word segmentation algorithms on the Akkadian language, an ancient language from Mesopotamia written in cuneiform and provides fundamental research in this area. At first, fundamen- tal differences and similarities between Chinese/Japanese and the Akkadian language will be discussed and a model of classification will be introduced. Subsequently, a selection of upto 20 suitable segmentation algorithms adapted from Chinese and Japanese is presented and applied using data from three different epochs of Akkadian history to guarantee representative results. The performance of the algorithms is evaluated afterwards using several evaluation metrics and interpreted to propose further improvements for the segmentation of Akkadian and other related languages.
Conference Paper
Full-text available
The efficient reconstruction of ancient cuneiform tablets has been a challenging problem over many decades, which if an approach is found to speed up the rate of potential fragment joins, will lead to a discovery of rich historical records of knowledge inscribed 5,000 years ago. This paper describes The Leverhulme Trust funded development of a crowd-sourcing collaborative virtual environment via the identification of cooperative personalities. We conducted experiments in which 12 groups of 3 participants collaborated to reconstruct fragments in a laboratory 3D puzzle virtual environment. Personality types were analysed together with NASA TLX workloads and task performance. We discovered that patterns exist for each of the virtual interaction modalities, demonstrating a potential for profiling systems to group personality types for maximising the effective reconstruction of cuneiform fragments.
Conference Paper
Full-text available
We propose a computer-assisted constraint-based methodology for virtual reassembly of Cultural Heritage (CH) artworks. Instead than focusing on automatic, unassisted reassembly, we targeted the scenarios where the reconstruction process is not be based on shape properties only but it is build over the experience and intuition of a CH expert. Our purpose is therefore to design a flexible interactive system, based on the selection of a set of constraints which relates different fragments, according to the understanding and experience of the CH operator. Once the user has defined those constraints, the system searches for a suitable solution, using a global energy minimization strategy that considers simultaneously all the pieces involved in the reconstruction process. Additionally, our framework provides the possibility to work in a hierarchical way, mimicking the traditional physical procedure that archaeologists use to reassemble tangible fractured objects. The frameworks is designed to work even with fragments that could have been severely damaged or eroded. On those datasets, automatic approaches may often fail, since the fractured regions do not contain enough geometric information to infer the correct matches. We present some successful uses of our framework on real application scenarios.
Full-text available
Despite the advances made in the recording and cataloguing of cuneiform tablets, there is still much work to be done in the field of cuneiform reconstruction. The processes employed to rebuild cuneiform fragments still rely on glue and putty, with manual matching of fragments from catalogues or individual collections. The reconstruction process is hindered by inadequate information about the size and shape of fragments, and the inaccessibility of the original fragments makes finding information difficult in some collections. Most catalogue data associated with cuneiform tablets concerns the content of the text, and not the physical appearance of complete or fragmented tablets. This paper shows how photogrammetric analysis of cuneiform tablets can be used to retrieve physical information directly from source materials without the risk of human error. An initial scan of 8000 images from the CDLI database has already revealed interesting new information about the tablets held in cuneiform archives, and offer...
Full-text available
Our paper presents the work of the Cuneiform Digital Forensic Project (CDFP), an interdisciplinary project at The University of Birmingham, concerned with the development of a multimedia database to support scholarly research into cuneiform, wedge-shaped writing imprinted onto clay tablets and indeed the earliest real form of writing. We describe the evolutionary design process and dynamic research and developmental cycles associated with the database. Unlike traditional publications, the electronic publication of resources offers the possibility of almost continuous revisions with the integration and support of new media and interfaces. However, if on-line resources are to win the favor and confidence of their respective communities there must to be a clear distinction between published and maintainable resources, and, developmental content. Published material should, ideally, be supported via standard web-browser interfaces with fully integrated tools so that users receive a reliable, homogenous and intuitive flow of information and media relevant to their needs. We discuss the inherent dynamics of the design and publication of our on-line resource, starting with the basic design and maintenance aspects of the electronic database, which includes photographic instances of cuneiform signs, and show how the continuous review process identifies areas for further research and development, for example, the "sign processor" graphical search tool and three-dimensional content, the results of which then feedback into the maintained resource.
Conference Paper
This paper introduces the ancient cuneiform sign system, the oldest known writing, and describes some of the the problems faced by scholars working in this field. It then shows how digital photography can help; specifically in assisting the construction of a Web-based descriptive list of the cuneiform sign set. Such a database has been designed, and is described. (See
Conference Paper
Reconstruction of ancient artifacts from fragments found at archaeological sites, is a tedious task that requires many hours of work from the archaeologist and restoration personnel. In this paper, we introduce an automatic technique for fragment reconstruction based on projective GPU depth maps that takes, as input data, a set of geometric representations of broken flat fragments, and finds the best rigid transformation that maximizes the contact area between them. In order to efficiently perform the calculations, instead of using input B-REP models, we focus on the characterization of the fragments. Thus, with some pre-processing entirely executed in the GPU, the representation of the 3D models proposed in this paper reduces significantly the complexity of the automatic reconstruction problem. Since the amount of data to process is huge, exhaustive search techniques cannot be used in order to provide fast results. On the other hand, solving the correspondence problem implies providing the best solution among all possible ones. To solve the stated duality between efficiency and correctness, the search process hierarchically analyzes the orthographic projections of the fragments and, recursively, refines the representation ensuring the validity of provided results. Global convergency and correctness is granted using an optimistic estimation of the potential match during the search process. This way, the proposed algorithm can be classified as a global registration technique, that efficiently converges to the optimal solution without stopping in local minima solutions.
Computers can now beat humans at the genteel pastime of jigsaw puzzles, thanks to an algorithm capable of solving a 10,000 piece puzzle in 24 hours
Cuneiform script, an intellectual breakthrough 5,000 years ago, made recording information possible. Cuneiform is mankind’s first ever script, recorded and communicated using clay tablets for thousands of years across the entirety of the Ancient Near East. Remnants of the medium are now stored worldwide in many of collections and time required for the joining of the fragments using traditional methodologies means that the information recorded within these fragments will not be known in our lifetime. The research narrated in this chapter opens up a novel method for reconstructing the fragments, using nature-inspired approaches and new mobile digitising technology. It covers groundwork done to date for supporting a fullscale stigmergy reconstruction of cuneiform tablets and provides hypothetical scenarios within a theoretical framework for testing ‘in the wild’.