Partial Attention-Based Direction-Aware Vehicle Re-identification

With the rapid development of urban transportation, vehicle re-identification has become a focal point in traffic management and vehicle tracking problems. In order to address the problem of small inter-class similarity among vehicles, previous studies utilize vehicle parsing models to extract local features. Therefore, we introduce the Squeeze-and-Excitation attention mechanism to extract important discriminative information from these local features. Furthermore, we propose a local co-occurrence attention mechanism to represent the proportion of common parts feature matching. To address the issue of large intra-class differences caused by vehicle direction change, we propose a lightweight and effective direction weighted fusion strategy. Experiments on two large datasets show that the proposed algorithm performs competitively.