Graphics Reference
In-Depth Information
If it's a locked-off, stationary camera we just need a couple of good markers that
stay out of people's hair. If we have a moving camera I have to think about where to
place these points so that we get good parallax with good visibility, where the actors
aren't going to walk.
There's a tradeoff between our data integration and roto/paint departments since
everymarker we put up on a set, they eventually have to paint out later in production.
It can be very difficult— for example if you have an actor's hair in front of amarker on
a blue screen, it's murder to get rid of it. However, especially with a moving camera
in front of a blue screen, we need to make sure we have enough markers to give us an
accurate 3D camera track — that's the priority.
We also have what we call our “tepee”s. These are simple assemblies made of
tubes and balls that look like an upside-down three-sided pyramid on top of another
three-sided pyramid. We carry the pieces around in a bag and can quickly assemble
and dissemble them on a set to quickly get objects in the scene that have known
dimensions and geometry. For smaller-scale scenes we have the “cubee,” which is
just a rigid metal frame with balls on the vertices.
Roble: A lot of times, people not in the filmmaking business will say, “Why don't you
try this cool new technique?” The environment on set is extraordinarily stressful. The
producers are therewatching howmuch time everything is taking; they basically have
this little stopwatch that counts in units of money. The guys in the data integration
group are really good about running into the scene, throwing up just enoughmarkers
to get what they need, and then running out before the director notices. A lot of
careful setup is a rarity. Most of it is get what you can and go.
RJR: Any other interesting applications of features?
Roble: For AI: Artificial Intelligence , Steven Spielberg wanted to be able to just walk
around the set, film it with a handheld camera, and see the result in real time on
a kind of virtual set. They mounted a couple of little cameras on top of the camera
pointed upward, and put a whole bunch of checkerboard patterns on the ceiling.
The upward-facing cameras could then find the checkerboards no matter where he
was on the set, and since they knew exactly where the checkerboards were in 3D
they could track the camera position pretty accurately in real time. Industrial Light
and Magic uses similar black-and-white patterns for their iMocap system, which lets
them track the heads and bodies of performers on set who will later be replaced by
CGI characters.
4.7
NOTES AND EXTENSIONS
Tuytelaars and Van Gool [ 503 ] proposed two affine-covariant-region-based feature
detectors not discussed here. The first is based on the edges in an image, under the
assumption that edges and their intersections can be reliably detected in views of the
same scene from different perspectives. The basic idea is to detect a Harris corner
and trace two roughly perpendicular Canny edges in its neighborhood until a pho-
tometric measure reaches its extremum, ultimately producing an affine-covariant
parallelogram. The second method is less reliant on edges. Instead, we begin at local
Search WWH ::




Custom Search