Morphable Graph  is a generative, graph-based approach for data-driven motion modeling and synthesis. Motion capture data is represented by a directed graph and motion synthesis tasks are converted to graph searching problem.
Natural human motion appears to have infinite variations. Therefore, in order to represent large variations in previously-recorded motions, high-level structures need to be defined for each type of motion. For example, the normal walking can be decomposed as a combination of six kinds of atomic motion clips: leftStance, rightStance, startLeftStance, startRightStance, endLeftStance and endRightStance. These high-level structures which are named motion primitives offer us an efficient and compact way to describe different behaviors. In morphable graph, each node represents a motion primitive, which is statistical model describting the distribution of motion clips under this category. Each edge represents the possible transition from one node to the other. Figure 1 shows an example of using morphable graph to represent normal style walking. In addition, morphable graph provides a compact way to store the large size of motion capture data.
Any number of motions with different variations can be generated using morphable graph. For random motion synthesis, a random graph walk is generated from morphable graph. For the nodes in the graph walk, random motion clips are sampled from each motion primitive and concatenated. For controlled motion synthesis, the statistical models naturally support different kinds of constraints. The controlled motion synthesis problem is formulated in a maximum a posterior (MAP) framework. We apply optimization to find the optimal motion which tries to reach constraints as good as possible while staying in natural space learned from captured data.
 Min, J. and Chai, J (2012): Motion Graphs++: a Compact Generative Model for Semantic Motion Analysis and Synthesis. In: ACM SIGGRAPH Asia 2012.
Contact: Klaus Fischer