Multi-scale feature pyramid and multi-branch neural network for person re-identification

The key to person re-identification (Re-ID) is how to extract a representative and robust depth feature of the person, which requires the model to pay attention to both global contour information and local detailed features. To extract more representative features, an effective method is to build a multi-branch deep model by duplicating the backbone structure. This method usually severely increases the computational cost, and continuous convolution and pooling operations cause the loss of detailed information. This paper proposes a lightweight multi-scale feature pyramid structure, which extracts features from network layers of different scales and aggregates them to supplement spatial detail information. Meanwhile, this paper adopts a pair of complementary attention modules, which pay attention to the discriminative areas of person features by focusing on channel aggregation and position perception, respectively. In addition, this paper proposes a multi-level orthogonal regularization method to further enhance the diversity of features. The experimental results show that the mAP of this method on the Market1501 dataset reaches 91.6%. The proposed method outperforms state-of-the-art methods and along with lower complexity.