Discriminant Feature Learning with Self-attention for Person Re-identification

Person re-identification (re-ID) across cameras is a crucial task, especially when cameras’ fields of views are non-overlapping. Feature extraction is challenging due to changing illumination conditions, complex background clutters, various camera viewing angles, and occlusions in this case. Moreover, the space mis-alignment of human corresponding regions caused by detectors is a big issue for feature matching across views. In this paper, we propose a strategy of merging attention models with the resnet-50 network for robust feature learning. The efficient self-attention model is used directly on the feature map to solve the space mis-alignment and local feature dependency problems. Furthermore, the loss function which jointly considers the cross-entropy loss and the triplet loss in training enables the network to capture both invariant features within the same individual and distinctive features between different people. Extensive experiments show that our proposed mechanism outperforms the state-of-the-art approaches on the large-scale datasets Market-1501 and DukeMTMC-reID.