Multi-Granularity Feature Fusion for Person Re-Identification
Yue Yang, Lizuo Jin · Advances in transdisciplinary engineering · 2022
Due to the limited ability of feature expression learned by one branch, designing a Multi-granularity network for feature extraction has become one of the important directions in the field of person re-identification. This paper designs a Multi-granularity feature fusion network (MFN) to enhance person feature extraction. The network is composed of global branches and local branches, and the former use the convolution pyramid to extract multiple scales features, through the channel attention module (Split Attention, SA) fusion of global branches and local branches, so that the global branch to obtain a strong ability to express persons features; the local branches is a feature map extracted from the backbone network based on the idea of Part-based convolutional baseline. Split the feature map horizontally into 4 branches, and the ID loss is calculated separately for each local branch. The classification loss in the total loss is consist of the ID loss of the 4 local branches. The two interact in parallel to improve the recognition ability. The experimental results on the Market1501 and DukeMTMC-ReID datasets, Rank-1/mAP reached 94.9%/86.4% and 87.5%/76.8% respectively.