This paper presents a hierarchical deformable model for robust human face detection, especially with occlusions and under low resolution. By parsing, we mean inferring the parse tree (a configuration of the proposed hierarchical model) for each face instance. In modeling, a three-layer hierarchical model is built consisting of six nodes. For each node, an active basis model is trained, and their spatial relations such as relative locations and scales are modeled using Gaussian distributions. In computing, we run the learned active basis models on testing images to obtain bottom-up hypotheses, followed by explicitly testing the compatible relations among those hypotheses to do verification and construct the parse tree in a top-down manner. In experiments, we test our approach on CMU+MIT face test set with improved performance obtained.