We propose a supervised manifold learning algorithm to discover appearance manifold of facial expressions for capturing dynamic emotion transitions. In particular, by incorporating prior class information into Locality Preserving Projections (LPP) \cite{He-Niyogi:nips03}, we propose a Supervised Locality Preserving Projections (SLPP) algorithm, which not only preserves local structure, but also encodes class information in the embedding. Extensive experiments on two databases demonstrate superior performance of SLPP to the original LPP, LDA, and PCA in learning subspace of facial expressions. For constructing a feature space representation, we introduce Boosted Local Binary Patterns (BoostLBP) for compactness without sacrificing discriminative power. BoostLBP based SLPP achieves generalization performance of 94.6\% (92.0\% if neutral expression included) for basic emotion recognition. When learning expression manifold from image sequences, we propose a novel alignment algorithm to align manifolds of different subjects on one generalized manifold. The learned generalized appearance manifold provides a unified framework for dynamic emotion analysis.

Return to Conference Programme