Fuzzy Velocity-Based Temporal Dependency for SVM-Driven Realistic Facial Animation

Driving a realistic facial animation with Support Vector Machine(SVM) requires determining the shape-to-wrinkle correspondence, which includes not only spatial dependency, but also temporal dependency. A few available frameworks(e.g., Recurrent Neural Network and Long Short-Term Memory), represent temporal dependency as the dependency of output on position input series, which however may bring about spatial redundancy in some cases. We argue that temporal dependency should be represented as the dependency of output on velocity input series. Besides, due to the weak temporal dependency between shape change and wrinkle change, we put forward Fuzzy Embedding to convert velocity into fuzzy velocity. The shape-wrinkle synthesis demonstrates that, in determining the temporal dependency between wrinkle change and shape change, fuzzy velocity provides more valuable information than velocity and thus enhances the degree of the realism effectively.