Human Face Analysis with Nonlinear Manifold Learning

Since human face movements distribute on a nonlinear manifold, there are inherent alignment residuals brought by the global linearity hypothesis in the traditional Principal Component Analysis (PCA) based Active Appearance Models (AAM). In this paper, a famous manifold learning method, Local Linear Embedding (LLE) is improved to model human face shape space for reducing the inherent alignment residuals. The experimental results show that the method, LLE-AAM, obtains lower alignment residuals to the tiny alterations of human face and still make successful alignment when PCA-AAM failed to some large alterations. According to the statistical analysis, LLE-AAM could reduce the residual to a certain extent.