Relation network based on multi-granular hypergraphs for person re-identification

摘要

In recent years, person re-identification (re-ID) has become a widespread research topic that focuses on retrieving target pedestrians from a set of images, typically taken by multiple cameras with different fields of view. Since the images of people often suffer from occlusion, misalignment and background clutter, the core challenge is to explore how to extract more discriminative and optimised features. State-of-the-art studies have found that further mining the relationships between local features can provide more sufficient semantic information for the final feature descriptors. Motivated by the idea, in this work, we propose an efficient Multi-Granular Hypergraph Relation Learning (MGHRL) module to explore the dependencies between part features. Specifically, a hypergraph is designed for a particular granularity where the relationships between each local node and other local nodes are modeled by propagating information along the hyperedges. Further, Hierarchical Complementary Identification (HCI) module is introduced to selectively activate diverse salient regions within the multi-scale feature maps to provide a more integrated global feature. By the cooperation of these two modules, the relationship between local features can be established from a global perspective. Extensive experiments on three popular benchmarks including Market1501, CUHK03 and DukeMTMC-reID demonstrate the feasibility and effectiveness of our approach.