MSFFNet: Multistage Feature Fusion With Metadata Injection for Vehicle Reidentification

Vehicle re-identification is a crucial task in intelligent transportation systems and urban management, aiming to retrieve images of the same vehicle from different camera viewpoints. Prevailing solutions typically rely on computationally expensive modules or costly annotations, hindering their practical deployment. To overcome these limitations, we propose the novel Multi-Stage Feature Fusion Network (MSFFNet), a lightweight and robust feature alignment framework. First, we design the Multi-Stage Fusion Head (MSFH) that generates robust feature representations by fusing multi-stage features based on the inherent hierarchical feature pyramid of CNNs. The resulting fusion features capture both the abstract semantics of high-level features and the fine-grained local details of low-level features. To alleviate the ineffective fusion caused by the varying distance distributions between features at different stages, we introduce the Graph Convolution-based Fusion Block (GCFB) to study the relationships between these features. Second, to mitigate the influence of irrelevant information during fusion, we introduce the Metadata Auxiliary Module (MAM) that injects easily available metadata, such as camera IDs and viewpoints, into fused features. This allows the model to focus on discriminative vehicle information. Our model produces compact feature representations, occupies less storage, and achieves faster retrieval speeds, with a computing time of 1.74 milliseconds per image. Extensive experiments demonstrate that our method achieves state-of-the-art performance on three widely used vehicle ReID benchmarks, establishing an optimal balance between speed and accuracy.