This paper presents an approach to reconstruct facial expressions using real data sets of people acquired by three-dimensional (3D) scanners. The acquired raw human face surfaces are pre-processed and a statistical shape model of the human face is built using multivariate statistical approaches. Our idea of using tensor model on the multivariate statistical method is to use all the face features found in the training set, with a variety of facial variations simultaneously by separating them into a number of classes. Point-to-point correspondences between the face surfaces are required in order to do the reconstruction processes. The advantage with the tensor-based multivariate statistical method is that it is practical to generate a variety of face shapes applied in different degrees, which would give a continuous and natural transition between the facial expressions. Our experiments focused on dense correspondence to compute the deformation of facial expressions. We have also used some selected landmark points placed on the face surfaces to compute the deformation of facial expressions. The selected landmark points are based on the Facial Action Coding System (FACS) framework and the movements are analysed according to the motion of the facial features. Besides altering human facial expressions, the presented approach could also be used to neutralise facial expression to aid the performance of face recognition.