A novel and efficient face feature is proposed in this paper. 3D faces from the database are preprocessed and mapped to 2D geometry images. Then the geometry images are decomposed into wavelet responses by multi-scale Gabor transforms. According to analyses and experiments, responses that represent vertical shape information are figured out to be face feature for recognition. Moreover, the feature extracted by multi-scale Haar transforms also obtains satisfying experiment results, which prove that the feature is free from the extraction methods. Extensive experiments conducted on FRGC(Face Recognition Grand Challenge) v2.0 show a better performance compared with existing popular methods. It is also approved that the vertical shape information is promising for dealing with face expressions in 3D face recognition.