Graphics Reference
In-Depth Information
under the truck to marry the two shots together. It's stuff like contact shadows that
really sell the effect. If the blacks in the underside of that truck aren't exactly the same
as the actor's, when the truck shadows and his shadows are in the same space, they
won't match, and it'll be a dead giveaway. It won't look like that truck is really rolling
over that guy. They even shot stills of the underside of that truck as it rolls over, since
they wanted to see more detail; we had to track those underside stills into the plate,
just by hand.
Since the guy actually hangs onto the bomb throughout thewhole B side, the bomb
itself was CG past a certain point and it also had to be painted out. This again is all
done essentially with roto shapes hand-drawn frame by frame, which is difficult since
there's a lot of motion blur as he's pulling the bomb back to his chest.
In the next shot you see the truck drive away and explode after a few moments.
That explosion was another A and B situation—putting them together and painting,
making truck parts fly apart. The A plate was basically the truck driving off safely, not
much happening, and the B platewas a practical explosion—the two shots are totally
different. The B shot looks like it's nighttime, but it's not; they just cranked down the
exposure so the explosion would be exposed right. Even though it's literally only a
few frames, there's just a lot of hand working to make it look good to the human eye.
There are very few things that can really be automated; software gets you so far, and
then it comes down to your ingenuity of actually connecting the A and B sides and
making them live in the same plate.
3.8
NOTES AND EXTENSIONS
Wang and Cohen [
533
] studied the problem of simultaneous matting and composit-
ing; that is, instead of treating the process as two separate steps, the matte of the
foreground region is optimized to minimize visual artifacts in the resulting compos-
ite. The benefit is that the foreground matte need not be highly accurate in regions
where the foreground and newbackground are similar. The computation of thematte
is similar to the methods in Section
2.6
.
Early work on image editing bymanipulating gradients was proposed by Elder and
colleagues (e.g., [
130
]). Later, McCann and Pollard [
318
] proposed tools for directly
“painting” in the gradient domain for image manipulation and art creation. Orzan
et al. [
357
] showed how smooth-shaded vector graphics artwork could be created
by specifying the locations of curved edges and the colors on either side. Gradient-
domain techniques can also be used to combine or remove artifacts in flash/no-
flash pairs [
10
]. Sunkavalli et al. [
483
] observed that the results of gradient-domain
compositing may exhibit a mismatch in
texture
between the source and target, even
if the colors are consistent. For example, the source patch may be sharp and clear
while the target image might have visible film grain and blurring. They proposed a
wavelet-pyramid-based compositing framework that matches the histograms of both
the image intensities and the estimated noise in the source and target images, to
propagate the target texture into the pasted source region. They called this approach
“image harmonization.”
Lo et al. [
298
] proposed a compositingmethod for 3D stereoscopicmovies, an area
of increasing interest to the visual effects community (see Chapter
5
). An unusual
