Automatic facial expressions analysis has been an active area of research in the last decade or so.It has a lot of potential in the field of human-computer interaction, image retreival, human emotion analysis,etc. Human emotion analysis goes well back into the nineteenth century when Darwin first demonstrated that there are specific inborn emotions in every living creature.
In recent times a lot of work is done in the supervised area of emotion detection and many algorithms have detected them with high success rate.These algorithms mostly use a classification based on Facial Action Coding System ( FACS ). Initially this system was for experts to classify facial expressions but some success is made in recent times to classify automatically. All such algorithms are currently supervised in nature and classify a test image into one of the clusters of already classified training images.
The methodology I intend to implement is an Unsupervised Maximum Margin Temporal Clustering ( MMTC )[1] approach to the above problem. Temporal Clustering ( TC ) refers to classification of frames into different clusters in a set of non-overlapping segments. TC is a relatively unexplored problem and the few algorithms which currently exists are based on generative models. In this work, MMTC is a novel learning framework that simultaneously perform temporal segmentation and learns a multi-class SVM for seperating temporal clustering.
Existing TC algorithms have several issues :- like the k-means clusterings are only optimal for spherical clustering. Generative approaches lack mechanism for feature selection,etc..
MMTC is basically an extension of Maximum Margin CLustering (MMC) over TC. MMC has shown promising results by extending the theory of SVMs to unsupervised learning. I will also use MRMMC (Membership Requirement MMC ), a modified MMC as baseline approach to compare the results over the database. Active Appearance Model tracker will be use for facial feature extraction and the database I plan to use is the Facial Expressions In the Wild[2].