Knowledge-Driven Multi-Branch Interaction Network for Vehicle Re-Identification

Vehicle re-identification plays a key role in intelligent transportation systems by supporting critical applications such as traffic flow optimization and public safety. With the introduction of triplet loss, the greatly enhanced vehicle discriminative ability brings more possibilities for the practical applications of vehicle re-identification. However, traditional triplet loss may not be fully suited for vehicle re-identification due to the risk of falling into local minima caused by solely selecting hard triplets. To address this problem, a Knowledge-Driven Multi-Branch Interaction Network (KMINet) is proposed which emphasizes robustness, feature diversity and feature fusion. At first, a knowledge-driven triplet mining module is introduced, which dynamically varies the selection of positive and negative samples. This module enables the network to learn the significance of features from a diverse range of triplets. Then, to tackle the challenge of viewpoint variation, a global and local multi-branch structure is designed to leverage both transformer and convolutional blocks to extract global context features and local detail features. This ensures that our model captures comprehensive vehicle characteristics. Finally, we incorporate an attention transfer module within the multi-branch structure to achieve information interaction and reduce feature redundancy across different branches. Our method outperforms state-of-the-art models on the VeRi-776 dataset, achieving 85.71% mAP and 97.73% CMC1. Moreover, the competitive performance is demonstrated on both VehicleID and VERI-Wild datasets, with mAP scores of 92.50% and 88.56%, respectively.