Context-Sensitive Conditional Ordinal Random Fields for Facial Action Intensity Estimation

Abstract

We address the problem of modeling intensity levels of facial actions in video sequences.

Machine Learning Methods for Social Signal Processing

“…The cs-CORF provides means of accounting for all six context questions from the W5+ context model. In (Rudovic et al 2013b), the authors demonstrate the influence of context on intensity estimation of facial expressions by modeling the context questions who (the observed person), how (the AU intensity-related changes in facial expressions), and when (the timing of the AU intensities). The context questions who and how are modeled by means of the newly introduced context and context-free covariate effects, while the context question when is modeled in terms of temporal correlation between the ordinal outputs, i.e., the AU intensity levels.…”

“…x are the feature measurements, and the latent variable z is non-linearly related to the ordinal labels y via the ordinal probit function, used to define the node features in the cs-CORF model. For more details, see (Rudovic et al 2013b). …”

“…(Rudovic et al 2013b) simultaneously accounts for the context questions who, how and when. x are the feature measurements, and the latent variable z is non-linearly related to the ordinal labels y via the ordinal probit function, used to define the node features in the cs-CORF model.…”

“…Recently, (Rudovic et al 2013b) proposed the Context-sensitive Conditional Ordinal Random Filed (cs-CORF) model for dynamic estimation of intensity of AUs, and facial expressions of pain. This model is a generalization of the CORF models (Kim & Pavlovic 2010, Rudovic et al 2012b, proposed for expression classification and temporal segmentation.…”

“…Based on the modeling approach, these can be divided into static methods (Mahoor et al 2009, Mavadati et al 2013, Savrana et al 2012, Kaltwang et al 2012, Jeni et al 2013) and dynamic methods (Rudovic et al 2013b). The static methods can further be divided into classification-based methods (e.g., (Mahoor et al 2009, Mavadati et al 2013) and regression-based (e.g, (Savrana et al 2012, Kaltwang et al 2012, Jeni et al 2013).…”