Graphics Reference
In-Depth Information
I
I'
(a)
I
I'
(b)
Figure 3.35.
An original image
I
and a proposed retargeted image
I
. We consider the contri-
bution of the black pixel
j
in
I
to the bidirectional similarity cost function. (a) For the coherence
term, we consider the
W
2
patches in
I
that contain the black pixel
j
, and find their best matches
in
I
. (b) For the completeness term, we consider the
N
j
patches in
I
whose best matches in
I
contain the black pixel
j
, where
N
j
changes from pixel to pixel. The pixel color at
j
in the retar-
geted image is iteratively updated as the weighted sum of colors at the gray pixels in
I
involved
in both terms.
change in dimension is not too large. Applying the algorithm with a coarse-to-fine
approach using a Gaussian pyramid effectively creates multi-scale patches if
W
is the
same at every level, and produces good approximations for successively finer scales.
Weights can also be applied at each pixel, for example using the saliency measure
from Section
3.5.1
, to bias the algorithm to preserve important regions (e.g., faces),
or to place zero weights on pixels to be removed from the image for inpainting.
Bidirectional similarity seems to do a better job at creating extremely small ver-
sions of images compared to seam carving (which is not really of interest for visual
effects); for example, it allows repetitive textures like windows on a building to auto-
matically be condensed. However, this approach is exciting for its ability to easily
recompose or reshuffle images. For example, the user can roughly cut image features
out of an image, rearrange them on the target image
I
, and fix these pixels' intensi-
ties. Then the iterative algorithm is applied to find the remaining pixels' intensities so
that the resulting image is as complete and coherent as possible. An example of this
approach is illustrated in Figure
3.36
. Similarly, a realistic expansion of an original
image can be created by fixing the position of the original on a larger canvas and opti-
mizing the coherence term in Equation (
3.41
) only over the unknown border regions.
