An efficient framework utilizing both local features and geometrical distribution for detecting facial components is presented. First, candidate facial components are efficiently collected by cascaded boosting of Haar-like features. The candidates may include false positives and multiple detections. Then, geometrical distribution of facial components is imposed on the candidates to select the optimal configuration. For simplicity, we suppose full dependence between the components and model it with multivariate Gaussian. The effectiveness of the framework is evaluated with experiments.