Motion Texture: A TwoLevel Statistical Model for Character Motion Synthesis

1
Motion Texture A Two-Level Statistical Model
for Character Motion Synthesis
CS780, Fall 2006

• Yan Li Tianshu Wang Heung-Yeung Shum
• Microsoft Research Asia Xian Jiaotong Univ.
• SIGGRAPH 2002

2
Outline
Introduction Related Work Motion Texture Learning
Motion Texture Synthesis with Motion
Texture Experimental Results Conclusion
3
Introduction
Motion editing Interactive motion
editing Adaptation to new characters and
environments Motion synthesis Retain the realism
of original captured data Allow the user to
control and direct the character Reordering the
captured data Chop into motion clips model
their transitions Kovar 02, Lee 02, Arikan 02,
Video Texture Schodl00
4
Two-level Statistical Model
Learning motion dynamics from captured
data Local dynamics a segment of frames linear
dynamic system (LDS) Global dynamics entire
sequence switching between LDS
5
Related Work
Motion Texture Bregler 02 Textons Julesz
81, Malik 99, Zhu 02, Guo 01, Liang 01 Linear
Dynamic System(LDS) Bregler 97, Soatto 01,
Fitzgibbon 01 Modeling nonlinear
dynamics Brand 00, Pavlovic 00
6
Representation of Observation and State variables
7
Motion Texture
Motion Texton LDS Distribution of Textons
first-order Markovian dynamincs
8
Learning Motion Texture
ML solution
By using L and H, and applying the first-order
Markovian,
9
Maximum likelihood problem
EM algorithm to solve the ML problem E-step
inference process for H and L M-step update
by fitting LDS
10
E-step
By a dynamic programming algorithm Gn(t) max
value of likelihood En(t) label of the last
segment Fn(t) beginning point of the last
segment
11
M-step
Second-order linear dynamic system
Closed form approximated estimation
12
Synthesis with Motion Texture
Two-step for synthesizing new motions 1. Texton
path 2. motion sequence for each texton
13
Texton Path Planning
1. Lowest Cost Path
2. Specifying the Path Length
14
Texton Synthesis by Sampling Noise
Given a texton and its key poses, a motion
sequence is synthesized frame by frame with the
learnt LDS and sampled noise. Synthesis error
accumulates !
15
Texton Synthesis with Constraint LDS
Incorporate the key poses of the following texton
as hard constraints.
16
Experimental Results
20 minutes of captured dance motion 4 hours for
learning the motion textures of 49800 frames 246
textons Length of a texton 60172 frames
17
Experimental Results
Synthesis with two adjacent textons
Noise-driven synthesis with different fine
details
18
Experimental Results
Extrapolating new dynamics
Editing a texton
19
Conclusions
Two-level statistical model to capture complex
motion dynamics Real-time synthesis Texton path
planning Constrained synthesis Generate fine
details Interactive editing Limitations training
data Physical correctness Similar original
motion Interaction with environment