Graphics Reference
In-Depth Information
speed up the bundle adjustment process by selecting a skeletal set of images —
that is, a much smaller subset of the large image collection that excludes redundant
images but still results in a complete and accurate scene reconstruction. After bun-
dle adjusting over the skeletal set, the parameters of the cameras not in this set can
be obtained by resectioning. An optional full bundle adjustment over all the camera
parameters and scene points is then much faster than the method described in the
previous paragraph, since the initial estimate is already very good.
6.7
INDUSTRY PERSPECTIVES
Doug Roble, creative director of software, and SomShankar, integration supervisor of
Digital Domain, in Venice, California discuss the role of matchmoving on amovie set.
In 1999, Doug Roble won a Scientific and Technical Achievement Academy Award for
writing Digital Domain's in-house 3D tracking/scene reconstruction program, called
“Track.”
RJR: How do you estimate the lens distortion of a movie camera?
Shankar: To undistort real movie camera lenses on feature films, we shoot a known
square grid with each camera setup and lens, and process those grids to create the
mapping between distorted and ideal coordinates. Sometimes we don't even use a
parametricmodel for the lens; we just use the nonparametricmapping obtained from
the grid.
Part of the reason is that movie lenses do weird things. Even though two of the
same lenses from the same manufacturer are supposed to have exactly the same
distortion, there are tiny anomalies between those lenses because they're still pieces
of glass. Anamorphic lenses, which stretch a widescreen image vertically to cover
the entire recorded film frame, are even more complicated since they're oval. Even
the lenses on new high-end digital cameras behave very interestingly; the images are
huge but the quality falls off toward the edges since you're not really meant to see
image through those parts of the lenses.
For final movie frames, we never undistort the original images. In our 3D com-
positing pipeline, we create all of our effects over the “flat,” undistorted version of
the plate, and then at the very end use the estimated lens distortion for that shot to
re-distort our 3D elements to match the original plate.
RJR: What kind of surveying do you do on set to help with matchmoving?
Shankar: We're often invited to a built set that will be filmed from different angles
for a whole sequence of shots. Our team goes to the set with a Leica total station (see
Chapter 8 ) — it's the same kind of device you see surveyors using on roadsides and
construction sites. We survey a sparse set of important locations in the room, such as
corners of objects, and record their 3D coordinates with this very precise device. We
also take a lot of photographs of the set, and then using our in-house software, Track,
we line that survey up to those photographs. It's also becoming muchmore common
to scan entire sets with LiDAR, which gives us a much denser sampling of 3D points
(see Section 8.1 ).
Search WWH ::




Custom Search