hide
Free keywords:
-
Abstract:
The idea that complex facial or body movements are composed of simpler components (usually referred to as ‘movement primitives’ or ‘action units’) is common in motor control (Chiovetto et al. 2010) as well as in the study of facial expressions (Ekman & Friesen, 1978). However, such components have rarely been extracted from real facial movement data. METHODS: We estimated spatio-temporal components that capture the major part of the variance of dynamic facial expressions, using a motion retargeting model for 3D facial animation (Curio et al, 2010) and applying dimension reduction methods (NMF and anechoic demixing). The estimated components were used to generate artificial stimuli, assessing the minimal required number of components in a perceptual Turing test, and their contributions to expression classification and expressiveness ratings. RESULTS: For an anechoic mixing model two components were sufficient for perfect reconstruction of the original expression. Often one component is sufficient for classification, while ratings tend to depend gradually on two, or more components.