Markov Random Field Structures for Facial Action Unit Intensity Estimation

We present a novel Markov Random Field (MRF) structure-based approach to the problem of facial action unit (AU) intensity estimation. AUs generally appear in common combinations, and exhibit strong relationships between the intensities of a number of AUs. The aim of this work is to harness these links in order to improve the estimation of the intensity values over that possible from regression of individual AUs. Our method exploits Support Vector Regression outputs to model appearance likelihoods of each individual AU, and integrates these with intensity combination priors in MRF structures to improve the overall intensity estimates. We demonstrate the benefits of our approach on the upper face AUs annotated in the DISFA database, with significant improvements seen in both correlation and error rates for the majority of AUs, and on average.