Self Attention based multi branch Network for Person Re-Identification

Recent progress in the field of person re-identification have shown promising improvement by designing neural networks to learn most discriminative features representations. Some efforts utilize similar parts from different locations to learn better representation with the help of soft attention, while others search for part based learning methods to enhance consecutive regions relationships in the learned features. However, only few attempts have been made to learn non-local similar parts directly for the person re-identification problem. In this paper, we propose a novel self attention based multi branch(classifier) network to directly model long-range dependencies in the learned features. Multi classifiers assist the model to learn discriminative features while self attention module encourages the learning to be independent of the feature map locations. Spectral normalization is applied in the whole network to improve the training dynamics and for the better convergence of the model. Experimental results on two benchmark datasets have shown the robustness of the proposed work.