Extrapolating single view face models for multi-view recognition

Performance of face recognition systems can be adversely affected by mismatches between training and test poses, especially when there is only one training image available. We address this problem by extending each statistical frontal face model with artificially synthesized models for non-frontal views. The synthesis methods are based on several implementations of Maximum Likelihood Linear Regression (MLLR), as well as standard multi-variate linear regression (LinReg). All synthesis techniques utilize prior information on how face models for the frontal view are related to face models for non-frontal views. The synthesis and extension approach is evaluated by applying it to two face verification systems: PCA based (holistic features) and DCTmod2 based (local features). Experiments on the FERET database suggest that for the PCA based system, the LinReg technique (which is based on a common relation between two sets of points) is more suited than the MLLR based techniques (which in effect are "single point to single point" transforms). For the DCTmod2 based system, the results show that synthesis via a new MLLR implementation obtains better performance than synthesis based on traditional MLLR (due to a lower number of free parameters). The results further show that extending frontal models considerably reduces errors.