Compared to the fruitful research outputs in 2D face recognition, the research in hyperspectral face recognition is quite limited in literature. When most available works process 2D slices of hyperspectral data separately, a 3D Gabor wavelet based approach is proposed in this paper to extract features in spatial and spectrum domain simultaneously. As a result, the information contained in the 3D data can be fully exploited. Experimental results show that the proposed approach substantially outperforms the methods available in literature such as spectrum feature, PCA and 2D-PCA on the HK-PolyU Hyperspectral Face Database under the same testing protocol. When only one sample per subject is available for training, our method also achieves very robust performance.