Dynamic Video Face Transformation Using Multilinear and Autoregressive Models

Abstract

In this paper we present a prototype system for altering perceived attributes of faces in video sequences, such as the apparent age, sex or emotional state. The system uses multilinear models to decompose the parameters coding for each frame into separate pose and identity parameters. The multilinear model is learnt automatically from the training video data. Statistical models of group identity are then used to alter the identity parameters from one group to another (e.g. from male to female). An autoregressive model is learnt from the pose parameters, and this is applied to alter the dynamics. We have tested our system on a small dataset (for altering apparent gender) with encouraging preliminary results.