Graphics Reference
In-Depth Information
human walking, climbing, dancing, and so on, even when less than tenmarkers were
used. This remarkable human ability to recognize human motion from very little
information suggests that the sparse input of marker motion is sufficient to capture
a recognizable, individual performance.
The topic by Menache [
323
] gives an excellent first-hand account of the history
and filmmaking applications of motion capture, as well as practical advice for set-
ting up a motion capture session and processing the resulting data files. Moeslund
and collegues [
331
,
332
] and Poppe [
372
] surveyed and categorized the literature
on markerless motion capture. In particular, [
332
] gives an exhaustive taxonomy of
vision-based pose estimation through 2006.
Today, a large amount of motion capture data is freely available to researchers
and animators. In particular, the Carnegie Mellon University motion capture
database
(
http://mocap.cs.cmu.edu
)
is an excellent resource containing thousands
of sequences from more than 100 performers in motion capture suits. The captured
activities range widely, including walking, dancing, swordfighting, doing household
chores, and mimicking animals. The data is available in many formats, and useful
tools for interacting with it are provided. The newer HumanEva datasets, also made
available by CMU [
449
], have a smaller range of activities, but include multi-view
video sequences synchronized withmotion capture trajectories frommarkers placed
onnormal clothing. Thesedatasets are valuable for developingmarkerless techniques
and have become benchmarks for the computer vision community.
In this chapter, we focused on motion capture technology using infrared lighting
and retro-reflective markers, but the same algorithms for marker triangulation and
processing apply to any markers (e.g., table tennis balls or ARTags), provided that
they can easily be detected and tracked in video. For example, the Imocap system
designed by Industrial Light and Magic uses white balls and binary patterns fastened
to a gray bodysuit to allow performances to be captured in natural environments
instead of on a motion capture stage. This technology was notably used in the
Pirates
of the Caribbean
and
Iron Man
movies. Also, facial motion capture markers are often
directly drawn or painted on the skin. Prototype technologies for newmotion capture
sensors include tiny ultrasound microphones coupled with miniature gyroscopes
and accelerometers [
519
] and lightweight photosensitive tags that respond to coded
optical transmitters [
382
]. These new technologies carry the promise of accurate
motion capture in real-world outdoor environments instead of carefully controlled
indoor stages.
We assumed that the kinematic models for humans were known; however, it's
also possible to learn these models and their relationship to motion capture markers
(e.g., [
394
,
241
]). Ross et al. [
399
] even showed how the kinematic models for unusual
objects (e.g., giraffes or construction cranes) could be estimated from tracked 2D
features alone. Such methods might be useful in situations where we have no good
prior model for the kinematic skeleton (e.g., motion capture of an unusual animal
like a kangaroo).
Yan and Pollefeys [
560
] extended factorization techniques for structure from
motion to estimate an underlying articulated skeleton from tracked features. Indeed,
the calibration and triangulation problems in motion capture are closely related
to aspects of the matchmoving problem we discussed in Chapter
6
, and famil-
iar techniques for structure from motion can be extended to markerless motion
