This paper presents an effective 3D face keypoint detection, description and matching framework based on three principle curvature measures. These measures give a unified definition of principle curvatures for both smooth and discrete surfaces. They can be reasonably computed based on the normal cycle theory and the geometric measure theory. The strong theoretical basis of these measures provides us a solid discrete estimation method on real 3D face scans represented as triangle meshes. Based on these estimated measures, the proposed method can automatically detect a set of sparse and discriminating 3D facial feature points. The local facial shape around each 3D feature point is comprehensively described by histograms of these principal curvature measures. To guarantee the pose invariance of these descriptors, three principle curvature vectors of these principle curvature measures are employed to assign the canonical directions. Similarity comparison between faces is accomplished by matching all these curvature-based local shape descriptors using the sparse representation-based reconstruction method. The proposed method was evaluated on three public databases, i.e. FRGC v2.0, Bosphorus, and Gavab. Experimental results demonstrated that the three principle curvature measures contain strong complementarity for 3D facial shape description, and their fusion can largely improve the recognition performance. Our approach achieves rank-one recognition rates of 99.6, 95.7, and 97.9% on the neutral subset, expression subset, and the whole FRGC v2.0 databases, respectively. This indicates that our method is robust to moderate facial expression variations. Moreover, it also achieves very competitive performance on the pose subset (over 98.6% except Yaw 90°) and the occlusion subset (98.4%) of the Bosphorus database. Even in the case of extreme pose variations like profiles, it also significantly outperforms the state-of-the-art approaches with a recognition rate of 57.1%. The experiments carried out on the Gavab databases further demonstrate the robustness of our method to varies head pose variations.