Multi-Granularity Feature Fusion Network for Cross-Domain Person Re-Identification

In recent years, a lot of encouraging work has emerged in the field of domain generalization as one of the core tools for solving the cross-domain Person Re-identification (ReID) task. However, on the one hand, most of the domain generalization methods are devoted to the extraction and aggregation of global or local features, and lack the modeling of structured features of the human body with rich fine-grained information; on the other hand, the existing domain generalization methods emphasize the representation of a single feature, and do not take into full consideration the association and effective fusion of many features. To this end, we propose a Multi-granularity Feature Fusion Network (MFFNet) for crossdomain Re-ID, which utilizes the designed four branching features to construct a complete representation of pedestrians with rich fine-grained and associative information, by drawing on the design idea of graph convolution. Specifically, we effectively construct the intrinsic correlation between local features by introducing the Local Feature Comparison Module (LFCM). In addition, we design a Graph Convolution Module (GCM) to generate pose relational features and topological features with strong discriminative structured information. A series of ablation and comparison experiments on authoritative benchmarks show that the proposed MFFNet achieves competitive performance among similar algorithms.