Unposing for data-driven facial expression spaces is unresolved

Establish a well-posed and reliable unposing procedure that removes pose-induced deformation from data-driven facial expression spaces so that expression parameters are strictly decorrelated from body pose and allow accurate modeling of gestures such as blinking.

Background

In discussing limitations of data-driven facial expression models, the authors note that such models typically retain residual pose variation due to the lack of a robust unposing method. Unposing refers to removing pose-dependent deformation to recover neutral expression geometry, which is necessary to ensure expressions are independent of body pose.

Because unposing is described as ill-posed and unresolved, data-driven expression spaces remain entangled with pose, hindering the optimization of expressions and the accurate modeling of common gestures such as blinking. MHR addresses some practical issues by adopting a semantic, artist-authored FACS-based expression space, but the broader research challenge of unposing remains open.