Partial attention and multi-attribute learning for vehicle re-identification

Abstract

Abstract Intelligent monitoring systems are in increasing need with the rapid growth of traffic nowadays. Vehicle re-identification has vital applications in digital forensics to track suspected vehicles in camera network. It is very challenging to learn discriminative information because of violent changes including the illumination and the viewpoint when a vehicle appears in different cameras, which will lead to the difficulty in distinguishing different vehicles and confusing the same vehicle. To improve the discrimination and the robustness of vehicle re-identification, we propose a partial attention and multi-attribute learning network. Focusing on the local areas which contain abundant discriminative information, we employ partial attention based on vehicle keypoint detection model. Moreover, because the color and the model of a vehicle are relatively stable in different viewpoints, we employ the branch networks to extract multi-attribute features which will improve the robustness. To validate our approach, experiments are carried out on VeRi and VehicleID datasets, and results show that the proposed method achieves higher accuracy compared with other methods.