Appeared in ICCP 2012 (IEEE Int. Conf.

on Computational Photography)

Alignment and Mosaicing of Non-Overlapping Images

Yair Poleg
Shmuel Peleg
School of Computer Science and Engineering
The Hebrew University of Jerusalem
Jerusalem, Israel

Abstract
Image alignment and mosaicing are usually performed
on a set of overlapping images, using features in the area
of overlap for alignment and for seamless stitching. Without image overlap current methods are helpless, and this is
the case we address in this paper. So if a traveler wants
to create a panoramic mosaic of a scene from pictures he
has taken, but realizes back home that his pictures do not
overlap, there is still hope.
The proposed process has three stages: (i) Images are
extrapolated beyond their original boundaries, hoping that
the extrapolated areas will cover the gaps between them.
This extrapolation becomes more blurred as we move away
from the original image. (ii) The extrapolated images are
aligned and their relative positions recovered. (iii) The gaps
between the images are inpainted to create a seamless mosaic image.

1. Introduction
Mosaicing several images to give a panoramic view has
been extensively studied, and the technology has matured
enough to be performed in real time inside many cameras.
The area of overlap between the pictures is critical for these
methods. Features in the overlap area are used for image
alignment, and once aligned the overlap area enables seamless stitching. We will skip a survey of mosaicing since it is
very common, and interested readers can look at [15].
But what happens when no overlapping areas exist, as
illustrated in Fig. 1? A simple case of non-overlap is addressed by square jigsaw puzzle solvers, such as [12, 8],
where an image is to be reconstructed from non-overlapping
square patches. In [12] a single pixel extrapolation is used
to determine the positions of the patches. Placement and
stitching are simple since the finite canvas size is known
in advance, and the patches are contiguous and have a relatively small number of possible positions. In contrast to
[12], in our case the relative placements, the gap widths,
and the canvas size are all unknown.

People are very good at alignment of non-overlapping

images. Linear features such as skyline and roads provide
important clues. Prior knowledge about the world, such as
the knowledge that sky and clouds are normally above the
ground, is also useful. It is conceivable that detecting linear
features and aligning images such that these are continued
smoothly could align successfully non-overlapping images.
Instead of taking this non-trivial task we want to examine
the power of much simpler, intensity based methods, involving no feature detection or scene understanding.
Our simple attempt includes three stages. The first stage
is image extrapolation, hoping that such extrapolation will
cover the gaps between the images. Image extrapolation is
covered in Sec.2. The second stage is the alignment of the
extrapolated images, presented in Sec. 3. The third stage
described in Sec. 4 is the filling of the gaps between the
images. Examples are shown in Sec. 5 and final remarks
are made in Sec. 6.
The main contribution of this paper is presenting the
problem of aligning and mosaicing non-overlapping images, and showing that in some cases it can have a simple
solution. In addition, a new approach to alignment of multiple images is presented.

2. Image Extrapolation
While image inpainting is very common [9, 13, 1], image extrapolation is not as common. Early usage was to allow image filtering without reducing the image size [16]. In
[11] video extrapolation was used to avoid cropping when
performing video stabilization. In [5, 3] video extrapolation
is performed to improve viewing experience of video. Our
approach to image extrapolation is similar to [3] applied to
single images rather than to video. Note that traditional image inpainting is done to please the human viewer and inpainted regions should look indistinguishable from the original image. Image extrapolation is used in our case to enable alignment, and hallucinated high-frequency information such as produced by [10], should be avoided. Since
the Fourier transform of the autocorrelation function is the
power spectrum, it is clear that the autocorrelation of high

Figure 1. A set of non-overlapping images which we would like to align and to mosaic. While image locations have been computed using
our proposed alignment method, input images are given unordered and without any position information.

frequencies decays much faster than the autocorrelation of

lower frequencies. Therefore, high frequencies should be
reduced in extrapolated regions as we get further away from
the original image.
Extrapolation using a multi-resolution Gaussian Pyramid
[2] gives the appropriate reduction of high frequencies. Extrapolation starts at the coarsest level (smallest pyramid image), and continues to higher resolution levels until the full
image resolution is reached (See Fig. 3). As desired, the
obtained extrapolation has higher frequencies closer to the
given image, and reduced details further away from the image.

2.1. Single Scale Extrapolation

In this section we describe image extrapolation as performed in a single level of the Gaussian pyramid. This is
a patch based method, and is similar to [9, 3], whose basic step is shown in Fig. 2. Let A be a rectangular region
whose right half is inside the image and left half is outside
the image. To extrapolate the pixels in the left half of A that
are outside the image we search for a patch B in the image
whose right part is most similar to the right part of A. Once
such a patch is found, we copy to the pixels in the left part
of A the colors of the corresponding pixels in the left part
of B.
When we create a mosaic from a set of images we use
the knowledge that all images are from the same scene,
and search for the closest patch in all images, as well as
in all levels in their Gaussian pyramids. Copying from one
scale into another scale represents the assumption of a fractal world, and enables to use similar areas that are at different distances from the camera.

(a)

(b)

Figure 2. (a) The right side of the patch A is inside the original
image and its left side is outside the image. To extrapolate the left
part of A we search in the image for a patch B whose right side is
most similar to the right side of A. (b) The left side of patch B is
used to paint the left side of patch A.

To extrapolate an image by k pixels in each direction

we start by centering a patch of size 2k 2k around every
boundary pixel (we used k = 5). Half of such a patch is
inside the image and half is outside the image and needs
extrapolation. For each such patch on the boundary a most
similar patch is sought from all scales of all images as in
Fig. 2. Patch similarity is inverse to the sum of distances of
all corresponding pixels. The distance between two pixels p
and q is is computed in the LAB color space, and is based on
the difference in each of the L, A, and B color components:
p

(pL qL )2 + (pA qA )2 + (pB qB )2

(1)
Each pixel in the extrapolation area is covered by 2k
patches. In our case, where k = 5, each pixel is covered by
10 extrapolations, each centered around a different boundD(p, q) =

Figure 3. Multi scale extrapolation, using a multi-scale pyramid

built for an image. The coarsest scale (smallest pyramid level) is
extrapolated by k pixels in each direction using the closest patches
approach. The extrapolated region is magnified by 2, and is attached to the higher resolution level. At that higher level we extrapolate again the k closest pixels to the image, updating the extrapolated region coming from the lower resolution. The other
pixels that were magnified from lower resolution levels are not updated. This process continues to higher resolutions until the full
image resolution is reached.

ary pixel. The extrapolated value of a pixel is computed by

averaging all overlapping patches covering it, and this average is done in a feathered way, giving higher weight to
points closer to centers of patches. This patch averaging is
done to avoid sharp seams, and is different from [3], where
non-overlapping patches were used.

2.2. Multi-Scale Extrapolation

We perform the image extrapolation in a multi-scale approach as shown in Fig. 3. We first build a multi-scale Gaussian pyramid for each image. We start the extrapolation
from the smallest pyramid level (coarsest scale). At this
level we extrapolate a band of k pixels around the image
using the approach described in Sec. 2.1. Once the extrapolation of the smallest level is complete, we continue to the
higher resolution level.
The extrapolated strip is magnified by 2, and is placed
around the image at the higher resolution level. The magnification blurs the extrapolated strip. The k pixels closest to the current pyramid level are extrapolated again, this
time from the image at the current pyramid level. To preserve continuity and to avoid outliers we use also the extrapolated values of the target patch, as computed from the
coarser scale, when searching for the closest patch. The

Figure 4. An example of an extrapolated image generated by three

steps of iterative multi-scale extrapolation. The extrapolated region is blurred. We did not draw a line around the the original
image as it conceals the transition from real to extrapolated parts.

source patches, though, are only patches from given images

and their pyramids, and do not include areas generated by
extrapolation in previous steps.
We continue the extrapolation from each level to the
higher resolution level until the original image resolution
is reached. If the pyramid had l levels above the original
image, the extrapolated area is of width k 2l in each direction. The k pixels closest to the source image have the
highest resolution, and as we move away from the image the
extrapolated areas are more blurred due to multiple magnifications.

2.3. Iterated Multi-Scale Extrapolation

The width that a multi-scale extrapolation can add to
an image is limited to k 2l in each direction, where k is
the single-level extrapolation width and l is the number of
pyramid levels. Larger extrapolation can be done in several
ways. One option is to increase k (we currently use k = 5).
We found that using a large k brings undesirable artifacts as
shown in Fig. 5. Another approach to increase the extrapolated area is to repeat the entire extrapolation step again,
and extrapolate the already extrapolated images. Since the
already extrapolated areas are very smooth, we do not want
to include high frequency details in the extrapolated areas.
We therefore use blurred versions of the images for creating
the additional extrapolated areas. An example of an extrapolated images after three iterations is shown in Fig. 4.

function that increases the cost between pixels with no (or

low) gradients in RGB. W (x, y) is computed as follows:
W (x, y) = 1 kp (x, y)kkq (x, y)k,

(4)

Where was found to be best for values in the range

[0.1, 0.3].
(a)

(b)

(c)

3.2. Batch Alignment of All Images

Figure 5. An example of an extrapolated image generated with

different values of k. (a) Input image. (b) Extrapolation with k =
5. The horizon is extrapolated as expected. (c) Extrapolation with
k = 15. The extrapolation duplicates the peak of the mountain.

3. Alignment of Extrapolated Images

Given two extrapolated images we align them in a similar way to the alignment of regular images, but we need to
take into account that the areas of overlap will now be extrapolated areas rather than real image areas, and that the
overlap area will be relatively small. The cost function used
for the alignment will take this into account. The methodology for multiple image alignment initially places all images
at the same location on top of each other, and images are
iteratively shifted to reduce the global cost function.
The proposed simultaneous alignment of multiple images assumes that most of the time we have overlap areas
only between two images. This justifies the use of a simple cost function between two images as done in this paper.
Methods exist for alignment of multiple overlapping images
[14], but they are not applicable in our case.

3.1. Cost Function: Alignment of Two Images

Like the cost function between patches (Eq. 1), the cost
function comparing two extrapolated images is also based
on a weighted sum of pixel distances for each of their LAB
color components. The distance between two pixels p and
q is now
p
(pL qL )2 + (pA qA )2 + (pB qB )2 ,
(2)
where in our experiments = 0.5. The distance between two images P and Q, after the suggested shift has
been applied to them, is the weighted sum over their overlapping area divided by the area of overlap = P Q:
D (p, q) =

P
Cost(P, Q) =

(x,y)

D (P (x, y), Q(x, y))W (x, y)

||

(3)
where (i) The sum over (x, y) uses the global coordinate system after alignment, and (ii) W (x, y) is a weight

In batch alignment we try to find an arrangement of all

input images that will minimize the sum of all pairwise
image costs. This is performed in an iterative process: In
the initial arrangement all images are placed on top of each
other, i.e. all input images initially have the same location.
In each iteration step we examine all possible translations
for each image (within a limited range, say 20 pixels), and
select the image and its translation giving the largest cost reduction. Iterations repeat until the cost can not be reduced
any further.
Similar methods of multi resolution search of multiple possible translations were traditionaly used for alignment of two images. Why will it work for multiple image
alignment? The insight comes from [7], addressing image
alignment when each image has two different motions (e.g.
transparency). It was shown in that paper that accurate motion can be computed using an ordinary two-frame method,
even when multiple motions are present, since one motion
always becomes dominant, and the other motions do not
spoil the solution. In our case it means that if we compute
the error at several shifts of two unrelated images, the error
will not change much. But when the two images are related,
a shift towards the correct motion will reduce the error substantially. When adding all errors from all images, only the
related images will affect the location of the minimum.
A visualization of the convergence of batch alignment
is shown in Fig. 6. It should be mentioned that the greedy
minimization described above can converge to a local minimum, and no movement of a single image could reduce
the energy. This is the case where methods like simulated
annealing may overcome such local minima. All our experiments converged successfully without simulated annealing.

3.3. Handling Rotations

The method described so far assumed pure translations
between images. In reality, most images taken with handheld cameras have some rotations. We have found that such
small, unintentional rotations (around 2) do not hurt the
alignment and the quality of the final results. Unlike ordinary mosaicing where rotation errors are clearly visible, in
our case the inpainting stage (describe in Sec. 4) serves to
conceal discrepancies.
In order to cope with larger rotations, such as in Fig. 7,
adjustments need to be made to the batch alignment process

Figure 7. Extrapolated input images with 15 rotations. It should

be noted that the original images have a small area of overlap that
does not interfere with the alignment process.

(a)

(b)

(c)

Figure 8. Creating integrated images. Each row shows the process

for a different input image. (a) Input image. (b) one of 21 rotated
copies of the input image. (c) Integrated image created from 21
rotated copies of the input image.

Figure 6. Visualization of the convergence of batch alignment. The

top image shows the initial step where all three extrapolated images are superimposed at the same location. With every iteration
the extrapolated images slide to their correct relative locations, and
the bottom image shows the image locations after convergence.
The gaps are not visible since the images are extrapolated. Once
the extrapolated regions are removed the gaps are revealed as in
Fig. 1.

of itself, as shown in Fig. 8. We proceed to batch alignment

of the integrated images, as described in Sec. 3.2. Once we
found the approximate alignment, we perform a local search
for the best rotation for each pair of images. We will now
describe each step in detail.
The first step is building an integrated image Ii for each
input image Pi of the N input images. Each such image
is an integration of an input image and rotated copies of
the input image. The angles of the rotated copies we integrate depend on the maximal rotation angle max we expect
in the input. The rotated copies we create cover the range
[max , max ] in steps of degrees. Formally, let us denote the set of rotation angles by S:
e
d 2max

S=
described in Sec. 3.2. A simple solution to this problem
might be to add a few rotations to the search for optimal
alignment. While this solution might work well, it comes
with a substantial increase in the complexity of the search.
Instead, we propose a method for aligning multiple images
with rotations that takes the observations made in Sec. 3.2
one step further.
The proposed alignment approach has three steps. We
begin by integrating each input image with rotated copies

{max + j},

(5)

j=0

and S0 = S {0}. Then the integrated image Ii is defined

by the average of all rotated images:
1 X
Ii (x, y) =
Pi, (x, y),
|S0 |

(6)

S0

where Pi, is the input image Pi rotated by degrees. The

sum is done for each RGB channel separately. The values

of max and define the amount of rotated images we integrate, but they also affect the next steps, as we describe
next.
The second step is batch alignment of the integrated images as described in Sec. 3.2. We have found that the alignment process fails for high values max or low values of
. This is because high values of max or low values of
produce integrated images that do not contain any useful
data. The resulting integrated images are too blurred and
smeared to tell one from another. We have successfully experimented with max 21 and 3 . Fig. 9 shows an
alignment result of integrated images.
The third step is selecting a rotation angle S0 for
each input image. We scan the alignment result for pairs of
overlapping integrated images. For each overlapping pair Ii
and Ij , we search for i and j such that
(i , j )

arg min

Cost(Pi,i , Pj,j ),

Figure 9. The alignment recovered for the integrated images in

Fig. 8.c.

(7)

(i ,j )S0 S0

where the cost function is as defined in Eq. (3). We add

i as candidate rotation angle for Pi , and j as a candidate
rotation angle for Pj . When we are done, we have a list
of candidate rotation angles for each input image. In many
cases this list contains one candidate angle per input image.
This is because both the alignment process and the rotation
selection process are based on the same pairwise cost. In
these cases, we simply replace each integrated image Ii by
a rotated copy of the input Pi, . Fig. 10 shows the final result of the alignment process, after replacing the integrated
images with rotated versions of the input images. Fig. 11
shows a complete example over three images.
If there is more than one rotation angle candidate for
some input image, we can do one of the following: (i)
Choose the most popular candidate. (ii) Calculate a
weighted average of all the candidate angles and use the
result. (iii) Perform a global search for the best angle for
each image. A brute force approach would require evaluation of all |S0 |N permutations in the worst case. We can,
however, narrow down the search space by searching only
around candidate rotation angles found earlier.

Figure 10. Final result, after alignment using integrated images

and inpainting. The integrated images showed in Fig. 9 replaced
with rotated copies of the input images and the gaps are inpainted.

rotations.

4. Inpainting the Gaps

Once the images are arranged properly, the gap between
them needs to be inpainted in order to create a seamless
panoramic image. Any of the many existing inpainting tools
will work. We found the inpainting software in [13, 1] very
easy to use.

3.4. Multi Resolution Alignment

5. Experimental Results

As usual in direct image alignment methods [4, 6], alignment is performed in a multi-resolution approach. Gaussian pyramids are built for the extrapolated images, and
the alignment starts at the lowest resolution. At that level
all possible translations and rotations are examined. Once
an optimal alignment is found at a low resolution level, a
higher resolution level is examined. When moving from
a lower resolution to a higher resolution, all translations
found so far are scaled to the new level. In our case they are
multiplied by two. The search for the minimum is now performed only around the previously found translations and

The input images for the mountains scene in Fig.1 were

cut from a high resolution image. Fig. 6 visualizes the
convergence of the alignment process and Fig. 12 shows
the seamless inpainted panorama alongside the ground truth
image. The normalized RMS error of the image locations
from ground truth is 0.041 in the X-axis and 0.007 in the
Y-axis. The normalization is with respect to the geometric
mean of the ground truth image dimensions. It is reasonable
that the Y alignment is more accurate than the X alignment
in this panoramic case, and the main achievement in the X
alignment is the correct ordering of the images.

(a)

(b)

(c)

Figure 12. Top: A panoramic image created after inpainting the

gaps as shown in Fig. 1. Bottom: Ground truth image from which
the input images were taken.

(d)
Figure 11. (a) Input images. (b) Alignment result without the use
of integrated images. (c) Alignment result using integrated images. (d) Integrated images replaced by the best rotation angle
image. While there are some errors in the horizontal alignment,
the vertical alignment and the rotation angles are approximately
correct.

(a)

(b)

(c)

(d)

The experiment in Fig. 14 shows mosaicing in a 2 2

grid structure of the input images. As in the previous experiment, the input images here were cut out from a high
resolution image as well. The normalized RMS error of the
image locations from ground truth is 0.056 in the X-axis
and 0.074 in the Y-axis.
Our last example is presented in Fig. 13. The input images were shot with a hand-held camera. The ground truth
data for this experiment comes from a normal panoramic
mosaic image (with overlap...). The normalized RMS error
of the image location from ground truth is 0.053 in the Xaxis and 0.016 in the Y-axis. In this panoramic case it is also
reasonable that the Y alignment is more accurate than the X
alignment, and the main achievement in the X alignment is
the correct ordering of the images.

Figure 14. A 2 2 example: (a) Input images. (b) Input images

after alignment. (c) Final mosaic after inpainting. (d) Ground truth
image from which the input images were taken.

Figure 13. Panoramic input images shown after alignment, and the generated panoramic image created from them by inpainting.

6. Concluding Remarks
Automatic alignment and mosaicing of non-overlapping
images has been introduced. The general scheme includes
steps of image extrapolation, alignment of extrapolated images, and inpainting the gaps. We believe that the described
approach will work well even if one or more of the three
steps will be implemented in methods that are different from
those described in this paper. Some of our experiments were
performed assuming a pure translation between images, and
the inpainting process conceals small rotations. We have
also introduced an approach for handling relative rotations
between images.
We must admit that we were surprised by the success of
the simple batch alignment of multiple images, and from
the success of integrated images in aligning rotated images.
We hope that this simple approach can be used for different
tasks where alignment of feature points may not be possible.

7. Acknowledgement
This work was supported by the Israel Science Foundation and the Ministry of Science and Technology.

References
[1] Content Aware Fill, Adobe Photoshop CS5.
[2] E. Adelson, C. Anderson, J. Bergen, P. Burt, and J. Ogden. Pyramid methods in image processing. RCA Engineer,
29:3341, 1984.
[3] A. Aides, T. Avraham, and Y. Schechner. Multi-scale ultrawide foveated video extrapolation. In ICCP, 2011.
[4] P. Anandan. A computational framework and an algorithm
for the measurement of visual motion. International Journal
of Computer Vision, 2(3):283310, 1989.

[5] T. Avraham and Y. Schechner. Ultrawide foveated video extrapolation. IEEE Journal of Selected Topics in Signal Processing, 5:321334, 2011.
[6] J. Bergen, P. Anandan, K. Hanna, and R. Hingorani. Hierarchical model-based motion estimation. In ECCV, pages
237252, 1992.
[7] J. Bergen, P. Burt, R. Hingorani, and S. Peleg. A three
frame algorithm for estimating two-component image motion. IEEE Trans. PAMI, 14:886896, 1992.
[8] T. S. Cho, S. Avidan, and W. T. Freeman. A probabilistic image jigsaw puzzle solver. In IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), 2010.
[9] A. Criminisi, P. Perez, and K. Toyama. Region filling and
object removal by exemplar-based inpainting. IEEE Trans.
Image Processing, pages 12001212, 2004.
[10] A. Efros and T. Leung. Texture synthesis by non-parametric
sampling. In In International Conference on Computer Vision, pages 10331038, 1999.
[11] Y. Matsushita, E. Ofek, W. Ge, X. Tang, and H. Shum.
Fullframe video stabilization with motion inpainting. IEEE
Trans. PAMI, pages 11501163, 2006.
[12] D. Pomeranz, M. Shemesh, and O. Ben-Shahar. A fully automated greedy square jigsaw puzzle solver. In CVPR, 2011.
[13] Y. Pritch, E. Kav-Venaki, and S. Peleg. Shift-map image
editing. In ICCV09, pages 151158, Kyoto, Sept 2009.
[14] H. Sawhney and R. Kumar. True multi-image alignment and
its application to mosaicing and lens distortion correction.
IEEE Trans. PAMI, pages 235243, 1999.
[15] R. Szeliski. Image alignment and stitching: A tutorial. Foundations and Trends in Computer Graphics and Computer Vision, 2:1104, Dec 2006.
[16] B. White and D. Brzakovic. Two methods of image extension. Computer Vision, Graphics, and Image Processing,
50:342352, 1990.