Multi-view Feature Complementary for Multi-query Vehicle Re-identification
Chaobin Zhang, Youqing Wu, Hanqin Shi, Zhengzheng Tu · 2023
Existing vehicle re-identification methods rely mainly on single query, which retrieves images of a given vehicle from other cameras based on the features extracted from a single image. However, the limited information contained in a single vehicle image significantly restricts the performance of vehicle re-identification in complex surveillance networks. In this paper, we propose a more practical and accessible task called multi-query vehicle re-identification, which utilizes multiple vehicle images from different viewpoint to overcome the viewpoint limitation of single query. Based on this task, we propose a novel multi-query feature complementary network (MFCNet), which fuses the information of vehicle images from multiple viewpoint during training to obtain the global features of vehicles. We then use these global features to optimize the local features extracted from a single viewpoint, enhancing the heterogeneity of vehicle features from different viewpoint. During testing, the complementary feature information from different vehicle perspectives is adaptively combined to reduce feature redundancy.