Spatio-temporal keypoints for video-based face recognition