1. This paper presented a technique, called motion texture,for synthesizing complex human-figure motion that is statistically similar to the original motion captured data.

2. Motion texture is a two-level statistical model,capturing complex motion dynamics. Motion texture is represented by a set of motion textons and their distribution. Local dynamics are captured by motion textons using linear dynamic systems, while global dynamics are modeled by switching between the textons.

2.1 Motion texton 
Each motion texton is represented by an LDS.

2.2 Distribution of textons
The distribution of textons is modeled as a first-order Markovian dynamics, which could be represented by a transition matrix.

2.3 Learning motion texture
A maximum likelihood algorithm can be used to learn the motion textons and their relationship from the captured dance motion.

3. Motion synthesis with motion texture

3.1 Motion can be generated from motion texture through a two-step algorithm: 
First, generate a texton path from state space.
Second, synthesize motion from textons

3.2 Texton path planning
If given two key textons specifying the start and end textons, a texton path can be find either through Dijkstra algorithm or Dynamic Programming algorithm.

3.3 Texton sythesis
There are two different ways to synthesis motion from a texton.
(1) Texton synthesis by sampling noise
The limitation of this algorithm is that the resulting motion will deviate from the original a lot,and will arrive a steady state finally.

(2) Texton synthesis with Constraint LDS
Constrain the resulting motion with start and end poses, and solve a linear system to get the intermediate poses.
The advantage of this method is the dynamics of the original motion can be preserved. Also it is useful to make smooth transition between motion synthesized by different textons.

4. Limitations

4.1 Because the texton distribution is calculated by counting how many times a texton is switched to another, this approach is best suited for motions consisting of frequently repeated patterns such as disco dance. The synthesized motion may lack global variations when the training data is limited.

4.2 There is no guarantee that the synthesized motion is physically realistic in the absolute sense.

4.3 Although this algorithm allows users to edit the motion at the texton level, the edited pose can not deviate from the original one too much.

4.4 Interacting with environment objects is not taken into consideration.