In this paper, we propose a discriminative keypoint selection-based 3D face recognition method that is superior to prevalent techniques in terms of both computational complexity and performance. We use the average face model (AFM) for face registration to efficiently locate the axis of symmetry in the rotated face mesh and recover a full frontal face from a 3D face model commonly corrupted due to pose variances. Instead of using the keypoint detection method, we use the feature selection algorithm to find the most discriminant keypoints for face identification and reduce computational time for not only feature extraction but also keypoint matching. The results of the experiments conducted on the Bosphorus database and the UMB-DB show that our algorithm can improve rank-1 identification accuracy, thus confirming its robustness against pose variances, expressions, and occlusions.