Statistical modelling for facial expression dynamics

Abstract

One of the most powerful and fastest means of relaying emotions between humans are facial expressions. The ability to capture, understand and mimic those emotions and their underlying dynamics in the synthetic counterpart is a challenging task because of the complexity of human emotions, different ways of conveying them, non-linearities caused by facial feature and head motion, and the ever critical eye of the viewer. This thesis sets out to address some of the limitations of existing techniques by investigating three components of expression modelling and parameterisation framework: (1) Feature and expression manifold representation, (2) Pose estimation, and (3) Expression dynamics modelling and their parameterisation for the purpose of driving a synthetic head avatar. First, we introduce a hierarchical representation based on the Point Distribution Model (PDM). Holistic representations imply that non-linearities caused by the motion of facial features, and intrafeature correlations are implicitly embedded and hence have to be accounted for in the resulting expression space. Also such representations require large training datasets to account for all possible variations. To address those shortcomings, and to provide a basis for learning more subtle, localised variations, our representation consists of tree-like structure where a holistic root component is decomposed into leaves containing the jaw outline, each of the eye and eyebrows and the mouth. Each of the hierarchical components is modelled according to its intrinsic functionality, rather than the final, holistic expression label. Secondly, we introduce a statistical approach for capturing an underlying low-dimension expression manifold by utilising components of the previously defined hierarchical representation. As Principal Component Analysis (PCA) based approaches cannot reliably capture variations caused by large facial feature changes because of its linear nature, the underlying dynamics manifold for each of the hierarchical components is modelled using a Hierarchical Latent Variable Model (HLVM) approach. Whilst retaining PCA properties, such a model introduces a probability density model which can deal with missing or incomplete data and allows discovery of internal within cluster structures. All of the model parameters and underlying density model are automatically estimated during the training stage. We investigate the usefulness of such a model to larger and unseen datasets. Thirdly, we extend the concept of HLVM model to pose estimation to address the non-linear shape deformations and definition of the plausible pose space caused by large head motion. Since our head rarely stays still, and its movements are intrinsically connected with the way we perceive and understand the expressions, pose information is an integral part of their dynamics. The proposed 3 approach integrates into our existing hierarchical representation model. It is learned using sparse and discreetly sampled training dataset, and generalises to a larger and continuous view-sphere. Finally, we introduce a framework that models and extracts expression dynamics. In existing frameworks, explicit definition of expression intensity and pose information, is often overlooked, although usually implicitly embedded in the underlying representation. We investigate modelling of the expression dynamics based on use of static information only, and focus on its sufficiency for the task at hand. We compare a rule-based method that utilises the existing latent structure and provides a fusion of different components with holistic and Bayesian Network (BN) approaches. An Active Appearance Model (AAM) based tracker is used to extract relevant information from input sequences. Such information is subsequently used to define the parametric structure of the underlying expression dynamics. We demonstrate that such information can be utilised to animate a synthetic head avatar. Submitted