A fusing method based on cluster-centric feature similarities is proposed. Firstly, locally linear embedding is used to learn low dimension embedding from original data space, and STHAC is used to divide projects as clustering in LLE feature space. Then, feature similarities are got from local appearance-based cluster, relevant similarity matching of exemplar points and cluster subspaces are done in a Bayesian maximum-a posteriori classification framework. Finally, face recognition is finished by importance probability of exemplars. The effectiveness of proposed method has been verified by experiments on video face databases CMU Mobo, Honda/UCSD and ChokePoint. Experimental results show that proposed method has higher recognition precision than several traditional methods.