LAUSR

Like pedestrian re-identification, vehicle re-identification (re-id) is an important part of building smart cities, and its purpose is to identify the same vehicle in vehicle images captured by multiple cameras. Vehicle re-id is more challenging than pedestrian re-id because many vehicles have similar colors and shapes, and their visual differences are usually very subtle. Existing vehicle re-id methods often rely on additional, expensive annotations to distinguish different vehicles. In contrast, we propose a two-branch network based on global attention mechanisms (MultiAttention-Net), which distinguishes subtle differences through adaptive learning. We introduce a global attention mechanism to highlight the differences between similar vehicles; however, compared with global appearance features, local features are more discriminant. Therefore, we propose combining global and local features to train the network to further improve the performance of vehicle re-id. During testing, only global features are used to measure the similarity between vehicle images. The experimental results show that the proposed MultiAttention-Net re-id method performs well on the challenging VeRi and VehicleID datasets.