Synthesis of a face image at a desired pose from a given pose

This paper proposes an efficient method to synthesize pose of an image. Pose synthesis is used to predict the face image with minimal error at a desired pose from a given pose. It is frequently required in many applications like production of animated movies, in forensic science and generation of 3D face geometry, etc. It uses principal component analysis (PCA) in conjunction with linear object classes (LOC) method (Vetter and Poggio, 1997). The face image of a pose is modelled as shape and texture vectors and the LOC method is applied on these two vectors of training set separately. The principal components of shape vector give a smaller number of significant dimensions along which the best linear approximation using LOC for the shape vector is calculated. Even though the given vector is approximated using only these dimensions, the error in approximation is significantly less compared to approximating pose image using all dimensions of shape vector. The proposed method is tested on CMU PIE face database and it is found to be significant improvement over the well known linear object classes method.

► We present an approach for synthesis of a face at a desired pose from a given one.
► It can help in face recognition by generating a desired pose available in database.
► It is found that our work improves over the LOC method for large rotations.
► We have also improved the shape prediction by applying PCA along with LOC.
► Experiments on CMU-PIE database show it performs much better than the LOC method.