Graphics Reference
In-Depth Information
to the sequence to introduce parallax and add stability to the solution. The match-
mover alsomust incorporate on-setmeasurements whenever they are available, such
as the height of the camera or 3D locations of surveyed points, and be able to assess
what information should be collected while a shot is being acquired to simplify the
camera tracking solution. The topic by Dobbert [
122
] is an excellent reference on the
practical aspects of matchmoving for the visual effects industry.
There are now several software packages for production-quality camera match-
moving based on the algorithms discussed in this chapter. These include
boujou
(sold
by Vicon),
PFTrack
(sold by The Pixel Farm),
Matchmover
(sold by Autodesk),
Syn-
thEyes
(sold by Andersson Technologies), and the freeware packages
Voodoo
(created
at theUniversity of Hannover) and
Bundler
(created at theUniversity ofWashington).
It is possible to relax the assumption that the scene observed by the cameras is
rigid (i.e., static) in all views. For example, Bregler et al. [
66
] extended a factorization
approach to allow the observed scene points to be an unknown linear combination
of unknown basis shapes. The camera pose, basis shapes, and linear coefficients are
obtained by successively factoring a measurement matrix similar to Equation (
6.38
),
assuming the camera is orthographic. Torresani et al. [
496
] proposed a probabilistic
approach to the same problem, assuming that shapes are drawn from a Gaussian
probability distribution with unknown parameters. These methods are only appro-
priate when a single deformable object (e.g., a face) dominates the scene. We will
address similar issues in more detail in the next chapter.
We mentioned the concepts of image-based video stabilization and re-
cinematography in Section
22
. If we apply matchmoving to an image sequence, we
can then smooth the camera path in 3D to remove translational and rotational jit-
ter [
296
], re-rendering the sequence to make it more pleasing. Camera localization
techniques can also help automate
rephotography
, the attempt to exactly duplicate
the vantage point of a historical photo in modern day [
24
], as well as algorithms to
automatically infer the order in which historical photos were taken [
428
].
While we focused exclusively on image-based methods for camera tracking in
this chapter, the same problem can also be solved with high precision by several
additional means. Welch and Foxlin [
542
] give a good survey of many different real-
time camera tracking systems. These systems can be based on mechanical sensing
(e.g., using potentiometers or shaft encoders), inertial sensing (e.g., gyroscopes and
accelerometers), acoustic sensing (e.g., ultrasonic ranging), magnetic sensing (com-
mon inhead-mounted displays), and optical sensing (e.g., active lighting using visible
or infrared LEDs). We will discuss several of these technologies in the context of
motion capture in the next chapter.
6.9
HOMEWORK PROBLEMS
6.1 Find the values of
d
x
and
d
y
in Equation (
6.2
) for a consumer digital camera.
6.2 Consider a 4096
2160 digital image taken using a camera with principal
point at the center (i.e.,
x
0
=
×
2048,
y
0
=
1080) and pixels that are physically
6.7
µ
m square. Suppose the lens distortion parameters for the camera are
10
−
3
,
κ
1
=
0. What is the observed (distorted) position of a pixel whose
ideal projection is at
κ
2
=
(
)
=
(
)
x
,
y
3000, 200
?
