Learning multiple consistent representation for vehicle re-identification

This paper addresses the challenges in vehicle re-identification, including high intraclass variance, low interclass variance, lack of fine-grained annotation, and misalignment of feature space. Our research is grounded in deep learning and weakly supervised learning theories, examining these challenges from multiple perspectives. A novel vehicle re-identification framework, the Learning Multiple Consistent Embeddings Network (MCEN), is proposed to address the difficulty of accurately locating and extracting key areas of vehicles, which is a result of limited fine-grained annotations. First, a multi-branch, multi-attention network is constructed, and a part diversity regularization term is introduced to encourage different attention branches to concentrate on deep features with distinct semantic meanings. Next, a part consistency regularization term is introduced to align features extracted by each branch within their respective Euclidean spaces, mitigating inter-class inconsistency and ensuring the alignment of local features. The network is trained in an end-to-end manner. The experimental results show that MCEN outperforms existing methods. In the VeRi-776 dataset, MCEN achieves an 8.3% improvement in mAP and a 3.1% increase in Top-1 accuracy over the PRN model. Unlike traditional feature-based methods and local feature detection approaches, MCEN regularization strategies improve robustness against occlusion and variations in camera angles. Furthermore, in the VehicleID dataset, MCEN shows superior performance in all three subsets (small, medium, and large), surpassing the state-of-the-art methods. These results validate the effectiveness of our proposed model in addressing real-world challenges in vehicle re-identification.