“…In other words, we seek to find two low-dimensional spaces: the identity space and the AU-intensity-specific space, which, together, generate the observed facial features. Various approaches that address this task have been proposed [19,4,20,21]. These are based on the tensor representation of different factors (i.e., identity, pose, illumination, and expression), decoupling of which is attained by means of multilinear generalizations of Singular Value Decomposition (SVD).…”