Clustering Facial Expressions with Transductive Learning

By initiating the process without guidance from labels, transductive learning techniques allow the discovery of the internal, natural structures existing in datasets. In parallel, various paradigms have been proposed to explore the human facial dynamics while showing an expression. Highly used paradigms are the Facial Action Coding System (FACS) and respectively, the intensity-valence encoding. In this paper, we are using standard clustering techniques over state of the art facial landmarking to investigate the natural preference of a paradigm. The surprising conclusion is that, although clustering is constructed upon a database build to illustrate the FACS system, face expression images do cluster along arousal and valence.