Abstract: In order to deal with automatic facial expression recognition task, a 3D facial expression recognition algorithm was proposed based on automatically detected fiducial points from 3D mesh models, range images and the corresponding 2D feature images. First, the fiducial points were automatically detected on the 3D mesh models, range images and the corresponding 2D feature images. And some fiducial points got from non-mesh models were mapped back to 3D mesh models to get full fiducial points. Then, Euclidian distances between these fiducial points were extracted as feature to feed the SVM classifier. Compared with two state-of-the-art algorithms, the classification results show that the fully automatic algorithm can achieve highly competitive classification rate. The average recognition rate was 87.1%, especially the recognition rate for surprise and happiness expression were 92.3% and 91.7% respectively.