Face Tracking with Multilinear (Tensor) Active Appearance Models

Face tracking in an unconstrained environment must contend with images that vary with viewpoint, illumination, expression, identity, and other causal factors. In a statistical approach, the multifactor nature of the image data makes the aforementioned problem amenable to analysis in a multilinear framework. In this paper, we propose Multilinear (Tensor) Active Appearance Models (MAAMs). The MAAM is a multilinear statistical model of facial appearance and shape that generalizes the linear Active Appearnce Model (AAM). As models of data variability, the latter fail to distinguish and account for the different sources of variabilty. On the other hand, our MAAMs explicitly represent the underlying processes of image formation, thus preserving attributes that are relevant to the task of tracking the human face.