Recent emergent face-related applications, such as face recognition and facial expression recognition, usually rely on accurate facial feature point localization. However, the variations in facial appearance, especially due to expressions, often make accurate localization of facial features very difficult. This paper proposes a graphical-model based approach for facial feature localization on expressional face images. By using the model, for localization while the influence between its local appearance and relative position is balanced. The experimental results show that our algorithm gives more accurate results than ASM and the AdaBoost-based facial feature detectors on Cohn-Kanade face database.