This work aims to create a model able to discern the parameters of shape and action units from 3D human face meshes. The adopted dataset was acquired by using Kinect and consist of 360 3D representation of human faces. More precisely, 20 different users performed 6 specific facial expressions (happy, sad, scared, angry, disgusted, surprised) by using 3 emphasis degree (low, medium, high). The collected dataset was labelled and then modelled in a three-dimensional tensor. Then, a multilinear dimensionality reduction technique (Higher-order singular value decomposition - HOSVD) was applied to separately extract the face deformation features related to the shape units and the action units. These specific features are finally exploited to independently rebuild the user human face by using much fewer data with respect to the starting dataset, specifically the 83% less, maintaining approximately 90% of variance.
