- 最后登录
- 2017-9-18
- 注册时间
- 2011-1-12
- 阅读权限
- 90
- 积分
- 12276
  
- 纳金币
- 5568
- 精华
- 0
|
1. Introduction
In most of the techniques to extract keyframes from motion capture
(mocap) data, criteria to guarantee the quality of keyframes
are provided by manually adjusting evaluation parameters. In this
study, the authors try to apply the Bayesian information criterion
(BIC) [Schwartz 1978] to keyframe extraction. BIC is a model-
selection criterion; models are evaluated under the tradeoff
between the goodness of fit to observed data and the model complexity.
Applying BIC allows us to automatically asses the quality
of keyframes under the tradeoff between the accuracy of interpolated
motions and the reduction of the number of keyframes.
2. Curve-Simplification Algorithm with BIC
Here, the curve-simplification algorithm [Lim et al. 2001] is
adopted as a keyframe-extraction method. In this algorithm, a
high-dimensional mocap-data curve is divided into two segments
at the point most distant from the straight line connecting the
endpoints of the curve. The point becomes the endpoint of each of
the new two straight lines representing a simplified curve, and is
regarded as a keyframe; this procedure is recursively repeated. In
this study, a curve is obtained in the joint-angle space in which the
exponential map is used instead of the axis-angle representation;
motion saliency is thereby more effectively reflected on a curve.
At each of the stages in the recursive process, the error between
the original curve u(n) and the simplified one ( ) SC u n ( n : frame
number) in the P-dimensional joint-angle space is calculated as
( ) ( ) ( ) SC ε n = u n − u n , and the time axis is treated as being divided
into K segments at keyframes. Under the assumption that ε has a
normal distribution at every segment, the log-likelihood function
L representing the goodness of fit is given as follows: |
|