We propose a novel approach to expression recognition from non-frontal view facial images. Under the unified Bayes theoretical framework, the recognition problem can be formulated as an optimization problem of minimizing an upper bound of Bayes error.