figure 12: seam insertion: finding and inserting the optimum seam on an enlarged image will most likely insert the same seam again and
again as in (b). inserting the seams in order of removal (c) achieves the desired 50% enlargement (d). using two steps of seam insertions of
50% in (f) achieves better results than scaling (e). in (g), a close view of the seams inserted to expand figure 10 is shown.
Duplicating all the seams in an image is equivalent
to standard scaling (see Figure 12(e)). To continue in
content-aware fashion for excessive image enlarging (for
instance, greater than 50%), we break the process into
several steps. Each step does not enlarge the size of the
image in more than a fraction of its size from the previous step, essentially guarding the important content from
being stretched. Nevertheless, extreme enlarging of an
image would most probably produce noticeable artifacts
(Figure 12(f) ).
4. 4. object removal
We use a simple user interface for object removal. The user
marks the object to be removed using negative weights,
in effect drawing the seams to pass through these pixels.
Consequently, seams are removed from the image until all
marked pixels are gone. The system can automatically calculate the smaller of the vertical or horizontal diameters
(in pixels) of the target removal region and perform vertical
or horizontal removals accordingly (Figure 13). Moreover,
to retain the original size of the image, seam insertion is
employed on the resulting (smaller) image (see Figure 14).
Note that this scheme alters the whole image (either its
size or its content if it is resized back). This is because both
the removed and inserted seams may pass anywhere in the
The reader is referred to more examples of retargeting
and resizing of video and images at
Most examples shown in this paper were computed automatically using the e error function. However, it is clear that
this scheme does not work well on all images. Other types
of importance functions either manual or automatic could
figure 13: simple object removal: the user marks a region for
removal (green), and possibly a region to protect (red), on the original
image (see inset in left image). on the right image, consecutive
vertical seam were removed until no “green” pixels were left.
be used in combination with higher level cues such as face
detectors to achieve better results (Figure 11).
Still, there are times when not even high level information
can solve the problem. We can characterize two major factors
that limit the seam carving approach. The first is the amount
of content in an image. If the image is too condensed, in the
sense that it does not contain “less important” areas, then
any type of content-aware resizing strategy will not succeed.
The second type of limitation is the layout of the image content. In certain types of images, albeit not being condensed,
the content is laid out in a manner that prevents the seams
from bypassing important parts (Figure 15).