Graphics Reference
In-Depth Information
what I'm going to do. You're trying to find out, am I going to have to blend across the
seam or am I going to use a hard line? Is there architecture in there that's going to
help me hide what I need to do?
If it's a still background, there's nomovement, and it's slightly out of focus, using a
fuzzy or a soft matte line will usually help. I can use the width of blur along a vertical
surface in the background as a rule of thumb for how much to blur the seam. If the
background's really sharp, then you try to follow strong edges and cut though empty
regions with the dividing line. In the sports bar example, I may follow the edge of
a table, arbitrarily go through the blank wall since it won't matter, then follow the
edge of a TV since I don't want to split the TV in half, and so on. You'll get a kind of
jigsaw-puzzle pattern that will be less detectable than a straight line; you purposely
want to go back and forth, grabbing a table from the A side and a monitor from the
B side. But even when the background is sharp, you always want some amount of
blend, even if it's just a pixel and a half.
If the cameras are moving, then you need to track or keyframe that line. A lot of it
depends on howwell the two videos are registered. If they're bothmoving quickly and
at the same speed, you may be able to get away with a super-wide, soft line, because
something's always moving across the screen and the blend is twenty-five percent
of the image, you're not going to see where it's transferring from one side to the
other. Now, if there are very detailed objects in the background, say you see a specific
car go by, you may have to roto out that car frame by frame and use it only from
the B side.
This kind of thing happens in TV as well as movies. Sometimes it's easy — say an
actor misses his cue and walks into a room too soon. You can actually use the original
footage as its own clean plate, and basically “slip” him— instead of having him come
in on frame 1 you have him come in on frame 25, because the information's there
to create a clean back plate. You hear a lot of “we'll just fix it in post.” There are so
many of those things that people don't know about — things that were not meant to
obviously be an effects shot, but where effects were done.
RJR: How do you deal with inpainting problems, where you need to synthesize realistic
texture inside a hole?
Lambert: Themost common approach is to track and warp a piece of existing texture
into the hole. These days, every shot that comes into Digital Domain gets tracked, so
we know the camera motion in 3D space, and we often survey the set or environment
to obtain accurate 3D geometry (see Chapters
6
and
8
). Then the compositors have
tools where we can project a given image of the scene onto this 3D geometry. To fill
in a hole in one image, we can figure out which 3D surface lies behind the object
we want to replace, project that surface onto a region in a different image where the
background is visible, and fill in the pixels we couldn't see with a piece of texture
that looks correct and moves in the right way. We do that kind of thing all the time
for set extensions. It's like a 2D compositing tool that actually uses 3D under the
hood. Achieving spatial and temporal consistency is still difficult, though; it's very
noticeable if there's a slight camera shake and your texture all of a sudden swims in
the opposite direction of what it's supposed to be doing.
