Exemplar Hidden Markov Models for classification of facial expressions in videos

Facial expressions are dynamic events comprised of meaningful temporal segments. A common approach to facial expression recognition in video is to first convert variable-length expression sequences into a vector representation by computing summary statistics of image-level features or of spatio-temporal features. These representations are then passed to a discriminative classifier such as a support vector machines (SVM). However, these approaches don't fully exploit the temporal dynamics of facial expressions. Hidden Markov Models (HMMs), provide a method for modeling variable-length expression time-series. Although HMMs have been explored in the past for expression classification, they are rarely used since classification performance is often lower than discriminative approaches, which may be attributed to the challenges of estimating generative models. This paper explores an approach for combining the modeling strength of HMMs with the discriminative power of SVMs via a model-based similarity framework. Each example is first instantiated into an Exemplar-HMM model. A probabilistic kernel is then used to compute a kernel matrix, to be used along with an SVM classifier. This paper proposes that dynamical models such as HMMs are advantageous for the facial expression problem space, when employed in a discriminative, exemplar-based classification framework. The approach yields state-of-the-art results on both posed (CK+ and OULU-CASIA) and spontaneous (FEEDTUM and AM-FED) expression datasets highlighting the performance advantages of the approach.