Solving Multiple Square Jigsaw Puzzles with Missing Pieces

Solving Multiple Square Jigsaw Puzzles with Missing Pieces
Genady Paikin
Technion
genadyp28@gmail.com
Ayellet Tal
Technion
ayellet@ee.technion.ac.il
Abstract
Jigsaw-puzzle solving is necessary in many applications,
including biology, archaeology, and every-day life. In
this paper we consider the square jigsaw puzzle problem,
where the goal is to reconstruct the image from a set of
non-overlapping, unordered, square puzzle parts. Our key
contribution is a fast, fully-automatic, and general solver,
which assumes no prior knowledge about the original im-
age. It is general in the sense that it can handle puzzles of
unknown size, with pieces of unknown orientation, and even
puzzles with missing pieces. Moreover, it can handle all the
above, given pieces from multiple puzzles. Through an ex-
tensive evaluation we show that our approach outperforms
state-of-the-art methods on commonly-used datasets.
1. Introduction
Puzzle solving is important in many applications, such as
image editing [4], biology [15] and archaeology [2, 10, 3],
to name a few. Though the problem is NP-complete [7, 1],
various solutions have been proposed. These are based on
shape matching [21, 20, 11, 9] or on a combination of shape
matching and color matching [12, 6, 23, 14, 18, 16]. When
the parts are square, only color matching is possible [5, 22,
17, 8, 19, 13]. The latter is the focus of our work.
The problem is introduced by [5], where a greedy algo-
rithm, as well as a benchmark, are proposed. The algorithm
discussed in [22] improves the results by using a particle fil-
ter. Pomeranz et al. [17] introduce the first fully-automatic
square jigsaw puzzle solver that is based on a greedy placer
and on a novel prediction-based dissimilarity. Gallagher [8]
generalizes the method to handle parts of unknown orien-
tation. Son et al. [13] demonstrate a considerable improve-
ment for the case of unknown orientation, by adding ”loop
constraints” to [8]. Rather than pursuing a greedy solver,
Sholomon et al. [19] present a genetic algorithm that is able
to solve large puzzles.
Our work is inspired by [17]. However, it takes the next
step and not only improves upon state-of-the-art results, but
4287 pieces(14.5% missing) 9379 pieces(9.6% missing)
Figure 1. Accurate reconstructions of our solver, given jigsaw puz-
zles with missing pieces (black squares).
also solves puzzles with additional challenges. In particular,
it handles puzzles with missing pieces, unknown size, and
unknown orientation of the parts (Figure 1). Moreover, it
concurrently solves multiple jigsaw puzzles whose pieces
are mixed together, where neither the size of the puzzles
nor a priori knowledge regarding possible missing pieces is
given. This is illustrated in Figure 2, where more than 5000
pieces that belong to five different puzzles, some with many
missing pieces, are given, and our algorithm reconstructs
them faultlessly.
Like [8, 17], our algorithm is greedy. However, it incor-
porates three key ideas, which not only prove beneficial for
solving traditional puzzles, but also support the additional
requirements mentioned above. First, similarly to previ-
ous methods, our placement is based on the compatibility
between pieces. We propose a more accurate and faster
compatibility function, which takes advantage both of the
similarity between the pieces, and of the reliability of this
similarity. Second, since greedy solvers are extremely vul-
nerable to the initial placement, we take special care when
choosing the first piece. We require it to have distinctive
borders and to be located in a distinctive region. Third,
rather than choosing the best piece for a specific location,
we select the piece that minimizes the likelihood of erring,
regardless of its location. As a result, our algorithm is de-
terministic, conversely to previous approaches, which make
random choices and are thus sensitive either to the selec-
tion of the first piece or to an initial random solution. This
means that it suffices to run our algorithm only once.
1

Input: 5367 pieces
Output:
728 pieces (9.6% missing) 466 pieces (13.7% missing)
3018 pieces (8.5% missing) 723 pieces (10.2% missing) 432 pieces (0% missing)
Figure 2. Solving multiple puzzles with mixed and missing pieces. The input contains parts from five puzzles. Our algorithm has no
knowledge regarding the size of the puzzles, the number of pieces per puzzle, or the number of missing pieces. Nonetheless, it reconstructs
the puzzles accurately in 30 seconds.
Our method outperforms the state-of-the-art methods,
both in accuracy and in efficiency. For instance, it achieves
accuracy (to be later defined) of 97.7% on the dataset
of [19], which consists of very large jigsaw puzzles,
whereas [19] achieves 96.41%. This is done while accel-
erating the running times by a factor of 60-120.
Our contribution is hence a novel and efficient puzzle
solver, which produces good results even when the input
consists of pieces from multiple puzzles, each has missing
pieces and unknown size. This may be the case not only in
everyday situations, but also in archaeology, when torn doc-
uments or broken artifacts are found mixed and lack parts.
2. Algorithm Outline
Our goal is to reconstruct images, given a set of non-
overlapping, unorganized, square pieces. We require that
the algorithm solves this problem even when:
1. The size of the puzzles is unknown.
2. The orientation of the pieces is unknown.
3. There may be missing pieces.
4. The input may contain pieces from several puzzles.
Our algorithm pursues a greedy strategy. As such, early
errors may lead to later failures. To avoid this case, we base
our algorithm on three principles: First, only highly reliable
matches should be utilized. This means not only that the
compatibility function must be good, but also that smooth
areas should not be handled at early stages. Second, the
first piece to use should be chosen thoughtfully, in order not
to diverge from the solution right away. Finally, rather than
choosing the best piece for a specific location, as often done,

we prefer to select the most distinctive piece, one that min-
imizes the likelihood of erring, regardless of its location.
Our algorithm proceeds in three steps: calculating the
compatibility between the pieces, finding the best piece to
start with, and assembly. The first step (Section 3) evaluates
the compatibility between every pair of pieces. We note that
the dissimilarity between the parts is not sufficiently infor-
mative. Hence, we also use the reliability of a match. This
leaves smooth area to the end of the assembly and assists in
handling puzzles with missing pieces.
The second step (Section 4) finds the best piece to start
with. The key idea is to select a piece that is not only dis-
tinctive, but also belongs to a distinctive region. This leads
to high confidence in the first assemblies.
The last step places the pieces (Section 5). This is done
by iteratively selecting a piece to be assembled, utilizing the
compatibility function from Step 1 and the concept of ”best
buddy” [17]. A guiding principle is that the absolute loca-
tion of a piece is determined only when all the pieces found
their neighbors. This is very similar to the way people solve
puzzles, constructing different portions of the puzzles sep-
arately, before composing them together. This not only lets
us solve puzzles of unknown size and unknown orientation,
but interestingly, also handles puzzles with missing pieces.
The only input of the algorithm is the number of puz-
zles that need to be solved out of the mixed pieces. During
the assembly, when all the pieces have low compatibility
values, the algorithm starts a new puzzle (up to the given
number of puzzles), by selecting a new initial piece.
3. The Compatibility Metric
A good compatibility metric is a fundamental compo-
nent of every jigsaw puzzle solver. Given two pieces pi, pj,
the compatibility function C(pi, pj, r) predicts the likeli-
hood that pi and pj are neighbors in the spatial relation r,
r ∈ {up, down, left, right}. We compute the compati-
bility function in two steps. First, a dissimilarity between
every pair of pieces is calculated. Then, we compute the
confidence in this dissimilarity.
3.1. The Dissimilarity Between Pieces
Various dissimilarity measures have been proposed in
the literature, including L2 [5, 19] and MGC (Mahalanobis
gradient compatibility) [8]. Our work is inspired by that of
Pomeranz et al. [17], which introduces a prediction-based
dissimilarity. We modify it, as explained below, to make
it not only simpler and faster to compute, but also perform
better, as demonstrated in Figure 3.
Briefly, in [17] backward-difference estimation is em-
ployed. It uses the last two pixels in each row (column)
near the boundary, from which prediction of the first pixel
in the adjacent piece is obtained. The dissimilarity measure
between pixels in two pieces uses the Lq
p norm, as follows,
(a) Our result (b) Using [17]’s dissimilarity
Figure 3. Our dissimilarity results in a perfect reconstruction.
When changing the dissimilarity to [17]’s, our algorithm errs.
Figure 4. Our dissimilarity is better than [8, 17, 5] in terms of the
correct (& incorrect) best buddies percentage and of accuracy.
where K is the piece size and d is the dimension in the LAB
color space:
D(pi, pj, right) =
K
k=1
3
d=1
[([2pi(k, K, d) − pi(k, K − 1, d)] − pj(k, 1, d))p
+
([2pj(k, 1, d) − pj(k, 2, d)] − pi(k, K, d))p
]q/p
. (1)
In practice, p = 3/10 and q = 1/16 were found to give the
best results.
We modify this in two manners. First, rather than using
the L
1/16
3/10 norm, we use the L1 norm. This clearly acceler-
ates the computation. In addition, we found it to improve
the results, as demonstrated in Figures 3–4. This can be
explained by the fact that when using L2, even a small dif-
ference between pieces will cause them to be considered
dissimilar. On the other hand, when using L
1/16
3/10, despite
large differences, the pieces may still be considered similar.
L1 is a compromise between the two.
Second, we use an asymmetric dissimilarity (that is,
D(p i, p j, right) = D(p j, p i, left)). We empirically
found it to be beneficial for calculating the compatibility
function that is performed next (Section 3.2). This is espe-
cially important when the puzzles have missing pieces, yet

the results improve regardless. Thus, our dissimilarity is:
D(pi, pj, right) =
K
k=1
3
d=1
([2pi(k, K, d) − pi(k, K − 1, d)] − pj(k, 1, d)) (2)
Figure 4 compares the results using our dissimilarity
measure to those obtained using the dissimilarities of [8, 17,
5], when run on the dataset of [17] (66 images). The number
of correct best buddies, as well as the accuracy of the algo-
rithm, improve, while the number of incorrect best buddies
decrease. We define these terms later, yet intuitively best
buddies are pieces that agree that they are the most compat-
ible neighbors of each other, and accuracy is measured by
the commonly-used neighbor count, as explained in Sec-
tion 6. It should be mentioned that using our dissimilarity
measure, 37 puzzles are reconstructed perfectly, compared
to only 25 when using [17]’s. Finally, our dissimilarity
computation is three times faster than [17]. This is impor-
tant since the dissimilarity calculation accounts for about
80% of the running time of our algorithm.
Acceleration: We found the following modification to be
very effective. The goal is to roughly estimate the likelihood
of two pieces to be adjacent, before calculating the full dis-
similarity. We calculate the average color (in LAB color
space) for each boundary of every piece in advance. Then,
only if the difference of the average color of two bound-
aries is sufficiently small in L2 (16 in practice), the exact
dissimilarity between these pieces is calculated. Otherwise,
the dissimilarity is set to a maximal value. This accelerates
the algorithm by a factor of 2, while hardly sacrificing the
quality of the results.
3.2. The Compatibility Function
The goal is to evaluate the reliability that a small dissim-
ilarity between two pieces indeed indicates adjacency in the
puzzle. Intuitively, in smooth areas every piece has a small
dissimilarity to every other piece in the region. Conversely,
in “interesting” regions, there are very few pairs that have a
small dissimilarity value. Hence, having a small dissimilar-
ity value by itself does not tell the full story. Therefore, we
seek out pieces that have unique dissimilarities.
To realize this, given a piece, if its closest neighbor and
its second closest neighbor have similar values, we do not
conclude that they are necessarily neighbors. However, if
the dissimilarity to the closest neighbor and that to second
closest neighbor differ, the closest pieces are more likely
to be neighbors. Let secondD(pi, r) be the value of the
second best dissimilarity of piece pi to all other pieces with
relation r. We define the compatibility function as:
C(pi, pj, r) = 1 −
D(pi, pj, r)
secondD(pi, r)
, (3)
The higher the value of the compatibility function C, the
more reliable the match is. We also experimented with other
options (e.g., the quartile neighbor, the median) and found
neither to be as indicative as the second best neighbor.
The ”Best Buddies” metric: This metric was first intro-
duced by [17]. Intuitively, two pieces are best buddies if
both agree that the other piece is their most likely neighbor
in a certain spatial relation. Two pieces pi, pj with relation
r1 and opposite relation r2 are best buddies if both hold:
∀pk = pj, C(pi, pj, r1) ≥ C(pi, pk, r1)
∀pk = pj, C(pj, pi, r2) ≥ C(pj, pk, r2). (4)
In practice, it is very rare that two pieces are best buddies
when they are not real neighbors. It is still rare even if the
input lacks many pieces, as it means that both pieces choose
each other as most similar.
4. Finding the First Piece
Selecting a good starting point is crucial in greedy al-
gorithms, as early errors lead to later failures. Nevertheless,
in [17, 19], a random piece is selected as the first one. These
algorithms are then run several times and present multiple
solutions. Conversely, we take special care in selecting the
first piece.
We look for a piece that is both distinctive and lies in
a distinctive region. The question is how to define distinc-
tiveness. Recall that our compatibility function reflects how
much a border of a piece is distinctive. The best buddies
condition also reflects how much a border is mutually dis-
tinctive. Moreover, best buddies of distinctive pieces are
usually correct. Therefore, we use the best buddies measure
to determine the first piece.
We look for a piece that has best buddies in all four spa-
tial relations, which indicates that it is a distinctive piece.
Moreover, we require that the four neighbors of this piece
have best buddies in all four spatial relations as well, i.e.,
the piece lies in a distinctive region.
Out of all the pieces that satisfy the above conditions,
we should choose one. We wish to select a piece that has
the “strongest” best buddies in all four direction. To do it,
we define mutual compatibility as follows (recall that our
dissimilarity function is asymmetric):
C(pi, pj, r1) = C(pj, pi, r2) =
C(pi, pj, r1) + C(pj, pi, r2)
2
, (5)
where the relation r2 is the opposite of relation r1. Now we
select as the first piece as the one that maximizes the sum of
the values of the mutual compatibility function (Equation 5)
of its four spatial relations.

5. Greedy Placement
Our algorithm’s final step assembles the pieces. This is
done without prior knowledge regarding missing pieces or
the size of the puzzle. Our placer (Algorithm 1) is greedy
and relies on our mutual compatibility function C (Equa-
tion 5). It maintains a pool of candidate pieces to be placed.
It iterates on placing a piece from this pool and then adding
candidates to it. We elaborate below.
Algorithm 1 Placer
1: While there are unplaced pieces
2: if the pool is not empty
3: Extract the best candidate from the pool
4: else
5: Recalculate the compatibility function
6: Find the best neighbors (not best buddies)
7: Place the above best piece
8: Add the best buddies of this piece to the pool
The selected piece is the one having the highest mutual
compatibility function among the pieces in the pool. If the
candidates’ pool is empty, but the jigsaw puzzle is not yet
solved, the compatibility function is recalculated. This is
performed by considering only the pieces that have not yet
been placed. Then, the next piece to be placed is selected
according to the mutual compatibility function, regardless
of whether it is a best buddy of an already-placed piece. It
is placed, as before, next to an already-placed piece. Then,
the best buddies (at the various directions) of the newly-
place piece are added to the pool.
The placement is performed without determining the ex-
act location. That is to say, a piece is placed in a location
that is relative to other pieces and not in a specific abso-
lute location. However, if the algorithm gets as input the
puzzle’s dimensions, the placement is limited to the given
height and width.
The algorithm performs well even if the puzzle lacks
pieces. This is because the algorithm never searches for the
best piece for a specific place, as previously done. Thus,
holes need not be filled. Moreover, only highly reliable
neighbors are selected—those having distinctive borders.
This makes the likelihood that a chosen piece is not a real
neighbor low.
Similarly, when the algorithm is required to solve mixed
puzzles, the only adjustment needed is the following. When
all the placed pieces have a relatively small value of mutual
compatibility to all the unplaced pieces (0.5 in our experi-
ments), a new puzzle is started. In this case, the candidates’
pool is cleared, and a new first piece is selected. This pro-
cess repeats until the given number of puzzles is reached.
The placement proceeds simultaneity on all the partial solu-
tions, until there are no remaining pieces.
6. Results
Our puzzle solver is applied to the datasets of [5, 17, 19].
These contain three sets, each of 20 images, with 432 [5],
540 and 805 pieces [17] respectively, two sets of medium-
size puzzles, each containing 3 images with 2360 and 3300
pieces [17], and three sets of large puzzles, each containing
20 images with 5015, 10375 and 22834 pieces [19] respec-
tively. We first quantitatively compare our results to those of
previous works when applied to the traditional jigsaw puz-
zle problem, where the size of puzzle is known and there are
no missing pieces. Then we test our algorithm qualitatively
(as there is no previous work) on additional cases, including
puzzles of unknown size, puzzles with missing pieces, and
mixed puzzles.
Evaluation when applied to the classical problem: Ta-
ble 1 reports the results, averaged per set, using the three
common measures. The average neighbor measure [5] is
the standard measure, which considers the fraction of the
correct pairwise adjacencies. It shows the average result
on [5, 17, 19] datasets and compares the results to those
of the state-of-the-art algorithms of [13, 17, 19]. It can be
seen that our algorithm is competitive with [13, 19] and
outperforms [17]. However, it should be noted that while
our result is deterministic, [17, 19] run their algorithms 10
times and report the results, in particular the best one.
The direct measure [5] considers the fraction of the
pieces in the assembled puzzle that are in their correct ab-
solute position. Generally, this measure is considered to be
less accurate and less meaningful due to its inability to cope
with slightly shifted puzzle. Our algorithm performs better
than [17, 19] and is comparable to [13].
Finally, the perfect columns indicate the number of puz-
zles for which the algorithms produced perfect reconstruc-
tions of the puzzle. Our algorithm has a clear advantage in
this regard.
Table 2 reports the results when applied to the dataset
of [19] that contains large puzzles. As only the algorithm
of [19] was applied to this dataset, we compare only to it. It
can be seen that our algorithm is beneficial for large puzzles
when considering each of the measures.
Our algorithm is advantageous also in terms of running
times. Table 3 compares our running times to those reported
in [19], which was applied to large puzzles. Our algorithm
is considerably faster for all puzzle sizes.
As another comparison, it is reported in [13] that the al-
gorithm solves a 9801-piece jigsaw puzzle with unknown
orientation in 25.6 hours. Our algorithm solves a 10375-
pieces puzzle with unknown orientation in 2 minutes.
Our algorithm was implemented in Java and ran on an
Intel i7 processor with 32GB RAM (all recent work ran on
similar PCs). It should be noted that as our algorithm need
not be adapted in order to solve puzzles with missing pieces,

neighbor direct perfect
# of pieces Our [13] [19] [17] Our [13] [19] [17] Our [13] [19] [17]
432 95.82% 95.5% 96.16% 94.25% 96.16% 95.6% 86.19% 90.95% 13 13 9 13
540 96.1% 95.2% 95.96% 90.9% 93.22% 92.2% 92.75% 83.45% 13 - 8 9
805 95.09% 94.9% 96.26% 89.7% 92.47% 93.1% 94.67% 80.25% 9 - 10 7
2360 96.26% 96.4% 88.86% 84.67% 94.01% 94.4% 85.73% 33.4% 1 - 1 1
3300 95.29% 96.4% 92.76% 85% 90.69% 92% 89.92% 80.67% 1 - 1 1
Overall 95.68% 95.31% 95.64% 91% 93.8% 93.59% 90.9% 82.35% 37 - 29 31
Table 1. Comparisons of our results to the state-of-the-art. Our algorithm outperforms state-of-the-art algorithms in all the commonly-
used measures. (’-’ mean that the results are not reported.) The comparison is made to the best-out-of-10 result of [17, 19]. When
considering the worst or the average case, their results are slightly worse.
neighbor direct perfect
best worst best worst best worst
# of pieces Our [19] Our [19] Our [19]
5015 96.43% 95.25% 94.87% 95.79% 94.78% 90.76% 12 11 7
10375 98.94% 98.47% 98.2 98.63% 97.69% 96.08% 8 6 5
22755 97.74% 96.28% 96.17 96.78% 92.02% 91.74% 3 4 4
Overall 97.7% 96.67% 96.41 97.07% 94.83% 92.77% 23 21 16
Table 2. Comparisons of our results to the state-of-the-art. Our algorithm outperforms the best solutions of [19] considering the common
measures on the dataset of [19], which contains large jigsaw puzzles. The other methods do not provide results on this dataset.
# of # of [19] Our Ratio
pieces images
432 20 48.73 [sec] 0.59 [sec] 82.59
540 20 64.06 [sec] 0.92 [sec] 69.63
805 20 116.18 [sec] 1.52 [sec] 76.43
2360 3 17.6 [min] 8.73 [sec] 120.96
3300 3 30.24 [min] 14.99 [sec] 121.04
5015 20 61.06 [min] 28.77 [sec] 127.34
10375 20 3.21 [hr] 2.06 [min] 93.5
22755 20 13.19 [hr] 14.75 [min] 60.7
Table 3. Comparison of the average running times. Our algo-
rithm is 60-127 times faster.
the running times are similar in this case as well.
Evaluation when applied to puzzles with missing pieces
and to mixed puzzles: Figures 1, 3, 5 show our results on
puzzles with missing pieces, where the missing pieces are
uniformly distributed. In particular, Figure 5 demonstrated
the quality of our results on a very large puzzle.
Figure 6 illustrates our results on three different patterns
of missing continuous regions, some of which are pretty
large. Patterns with many missing pieces in the center are
especially challenging, because in many cases the salient
region of an image lies in its center.
When the orientation of the pieces is unknown, our algo-
rithm achieves a 95.41% success rate with direct compari-
son and 95.4% with neighbors comparison, on the set with
puzzles of 432 pieces . In comparison, the algorithm of [13]
achieves a 94.7% and 94.9% success rate respectively.
Finally, Figures 2 and 7 demonstrate the results of our
algorithm on complex inputs of mixed puzzles with missing
pieces, where the size of every puzzle is unknown. Other
examples can be found in the supplementary material.
Importance of the first piece: To understand the contri-
bution of choosing the first piece carefully, we started from
a random piece, as done in earlier works, and ran our algo-
rithm 10 times on the dataset of Table 1. The average worst
result dropped from 93.8 to 82.09 (direct comparison) &
from 95.68 to 94.36 (neighbor comparison).
Limitations: Our algorithm may produce unsatisfactory
results when the images are smooth or noisy, as demon-
strated in Figure 8. It should be mentioned that other al-
gorithms fail in these cases as well.
7. Conclusions
In this paper we have introduced a novel algorithm for
square jigsaw puzzle assembly. We utilized a greedy ap-
proach that makes use of a compatibility function between
pieces. Our function is more accurate and faster to compute
than previous functions. Moreover, since greedy solvers
highly depend on the initial placement, we take special care
when choosing the first piece. As a result, our algorithm is

(a) input (b) our solution (c) the original image
Figure 5. Our solution to a puzzle with 22755 piece, where 1160 of the pieces are missing. The puzzle is solved accurately.
432 pieces 249 pieces 303 pieces 180 pieces
(a) the original image (b) our solution (circle) (c) our solution (circular hole) (d) our solution (ring)
Figure 6. Our solutions to puzzles having different patterns of missing pieces. Our algorithm solved the puzzles perfectly, even though the
missing regions are large.
deterministic and we need not involve the user in choosing
the best solution.
We have shown that our algorithm outperforms the state-
of-art results for regular puzzles, in terms of both accuracy
and efﬁciency. Moreover, our algorithm handles success-
fully puzzles with missing pieces and input that contains
pieces from multiple puzzles.
Acknowledgments: This research was funded in part by
the Binational Science Foundation (BSF), the Israel Sci-
ence Foundation (ISF) 1420/12, GRAVITATE under 2020-
REFLECTIVE-7-2014, and Ollendorff Foundation.
References
[1] T. Altman. Solving the jigsaw puzzle problem in linear time.
Applied Artiﬁcial Intelligence, 3(4):453–462, 1990. 1
[2] B. J. Brown, C. Toler-Franklin, D. Nehab, M. Burns,
D. Dobkin, A. Vlachopoulos, C. Doumas, S. Rusinkiewicz,
and T. Weyrich. A system for high-volume acquisition and
matching of fresco fragments: Reassembling theran wall
paintings. ACM Transactions on Graphics, 27(3), 2008. 1
[3] S. Cao, H. Liu, and S. Yan. Automated assembly of shredded
pieces from multiple photos. In IEEE Int. Conf. on Multime-
dia and Expo, pages 358–363, 2010. 1
[4] T. Cho, S. Avidan, and W. Freeman. The patch transform.

Input: 13584 pieces
Output:
4322 pieces(13.8% missing) 5015 pieces(0% missing) 4517 pieces(9.9% missing)
Figure 7. Solving multiple puzzles with mixed and missing pieces. The input contains parts from 3 puzzles. Our algorithm has no
knowledge regarding the size, the number of pieces per puzzle, or the number of missing pieces. Nonetheless, it reconstructs the puzzles
perfectly in 77 seconds.
PAMI, 32(8):1489–1501, 2010. 1
[5] T. Cho, S. Avidan, and W. Freeman. A probabilistic image
jigsaw puzzle solver. In CVPR, 2010. 1, 3, 4, 5
[6] M. G. Chung, M. M. Fleck, and D. A. Forsyth. Jigsaw puzzle
solver using shape and color. In ICSP, 1998. 1
[7] E. Demaine and M. Demaine. Jigsaw puzzles, edge match-
ing,and polyomino packing: Connections and complexity.
Graphs and Combinatorics, 23:195–208, 2007. 1
[8] A. Gallagher. Jigsaw puzzles with pieces of unknown orien-
tation. In CVPR, 2012. 1, 3, 4
[9] D. Goldberg, C. Malon, and M. Bern. A global approach to
solution of jigsaw puzzles. In Symposium on Computational
Geometry, 2002. 1
[10] D. Koller and M. Levoy. Computer-aided reconstruction and
new matches in the forma urbis romae. Bullettino Della
Commissione Archeologica Comunale di Roma, 15:103–
125, 2006. 1
[11] W. Kong and B. Kimia. On solving 2d and 3d puzzles using
curve matching. In CVPR, 2001. 1
[12] D. A. Kosiba, P. M. Devaux, S. Balasubramanian, T. L.
Gandhi, and K. Kasturi. An automatic jigsaw puzzle solver.
In ICPR, 1994. 1
[13] K.Son, J. Hays, and D. Cooper. Solving square jigsaw puz-
zles with loop constraints. In ECCV, 2014. 1, 5, 6
[14] M. Makridis and N. Papamarkos. A new technique for solv-
ing a jigsaw puzzle. In IEEE Image Processing, 2006. 1
[15] W. Marande and G. Burger. Mitochondrial dna as a genomic
jigsaw puzzle. Science, pages 318–415, 2007. 1
[16] T. R. Nielsen, P. Drewsen, and K. Hansen. Solving jigsaw
puzzles using image features. Pattern Recogn. Lett, 2008. 1
Figure 8. Limitations: Our algorithm may fail on smooth images.
[17] D. Pomeranz, M. Shemesh, and O. Ben-Shahar. A fully au-
tomated greedy square jigsaw puzzle solver. In CVPR, 2011.
1, 3, 4, 5, 6
[18] M. Sagiroglu and A. Ercil. A texture based matching ap-
proach for automated assembly of puzzles. In ICPR, 2006.
1
[19] D. Sholomon, O. David, and N. Netanyahu. A genetic
algorithm-based solver for very large jigsaw puzzles. In
CVPR, 2013. 1, 2, 3, 4, 5, 6
[20] R. Webster, P. LaFollette, and R. Stafford. Isthmus criti-
cal points for solving jigsaw puzzles in computer vision. In
IEEE Trans. Systems, Man and Cybernetics, 1991. 1
[21] H. Wolfson, E. Schonberg, A. Kalvin, and Y. Lamdan. Solv-
ing jigsaw puzzles by computer. Annals of Operations Re-
search, 12:51–64, 1988. 1
[22] X. Yang, N. Adluru, and L. Latecki. Particle ﬁlter with state
permutations for solving image jigsaw puzzles. In CVPR,
2011. 1
[23] F. Yao and G. Shao. A shape and image merging technique
to solve jigsaw puzzles. Pattern Recognition Let., 2003. 1

Solving Multiple Square Jigsaw Puzzles with Missing Pieces

More Related Content

What's hot (13)

Viewers also liked (12)

Similar to Solving Multiple Square Jigsaw Puzzles with Missing Pieces (20)

Recently uploaded (20)

Solving Multiple Square Jigsaw Puzzles with Missing Pieces