Masked Graph Attention Network for Person Re-identification

The mainstream methods for person re-identification (ReID) mainly focus on the correspondence between individual sample images and labels, while ignoring rich global mutual information resides in the whole sample set. We propose a method called Masked Graph Attention Network (MGAT) to address this problem. MGAT operates on the complete graph constructed with the extracted features, where nodes are able to directionally attend over other nodes' features under the guidance of label information in the form of mask matrix. By using MGAT module, the previously neglected global mutual information is exploited to generate an optimized feature space with more discriminant power. Meanwhile, we propose to feedback the optimization information learned by MGAT module to the feature-embedding network to enhance the mapping capability, thus avoiding the difficulty to handle large-scale graphs in testing phase. To evaluate our method, we conduct experiments on three commonly used ReID datasets. The results show that our method outperforms most mainstream methods, and is highly comparable to the state-of-the-art method.