Research On Fine-grained Face Verification

Current deep learning methods have achieved human-level performance on Labeled Faces in the Wild (LFW) database, but we think it is because that the limited number of pairs on LFW do not capture the real difficulty of large-scale unconstrained face verification problem. Besides the intra-class variations like pose, occlusion and expression, visually similarity of different persons’ faces is an another challenge. It is unavoidable while many researchers ignore it.Therefore, in this paper, we firstly select some visually similar pairs in LFW database by combining the deep learning method and human annotation results. Preserving the matched pairs and replacing the mismatched pairs of LFW with the selected similar pairs, we obtain the Fine-grained LFW (FGLF W)database which can better reflect the real difficulty of face verification. Ex-perimental results show that methods achieving not bad performance on LFW drops more than 11% even 25% on FGLFW. While human obtain only 92.03%on FGLFW. It reflects that visually similar pairs are difficult to current methods and our FGLFW database is a quite challenging database.For dealing with the fine-grained face verification problem, we propose modified methods from two aspects. From the aspect of metric learning, we propose geometry-aware metric learning(GAML) method which can preserve the robustness of the deep feature. From the aspect of deep learning, we pro-pose patch based Convolutional Neural Network and discriminative localization based Convolutional Neural Network. Both network fully utilize the discrim-inative detail information. Experimental results show a good improvement on FGLFW comparing with the baseline model.