3D face has been introduced in the literature to deal with the unsolved issues of 2D face recognition, namely lighting and pose variations. In this paper, we study and compare the distinctiveness of features extracted from both the registered 2D face images and 3D face models. Sparse Representation Classifier (SRC) is exploited to calculate all similarity measures which are compared with the ones by a baseline of Nearest Neighbor (NN). As individual features of 2D and 3D are far from distinctive for discriminating human faces, we further present an adaptive score level fusion strategy for multimodal 2D-3D face recognition. The novel fusion strategy consists of an offline and an online weight learning process, both of which automatically select the most relevant weights of all the scores for each probe face in each modality. The weights calculated offline are based on the EER value of each type of features, while the online ones are dynamically obtained according to matching scores. Both types of weights are then fused to generate a final weight. Tested on the complete FRGC v2.0 dataset, the best rank-one recognition rate using only 3D or 2D features is 79.72% and 77.89%, respectively; while the new proposed adaptive fusion strategy achieves 95.48% with a 97.03% verification rate at 0.001 FAR, highlighting the benefit of exploring both 3D and 2D clues as well as the effectiveness of our adaptive fusion strategy.