PROJECT TITLE :

VARID Viewpoint-Aware Re-IDentification of Vehicle Based on Triplet Loss

ABSTRACT:

Research on vehicle re-identification, also known as "Re-ID," has garnered a significant amount of attention in recent years due to the growing prevalence of intelligent traffic control and monitoring. The problem of re-identifying vehicles is significantly more difficult and unpredictable than other cross-view searching tasks such as re-identifying people. This is due to the fact that changes in viewpoint can have a significant impact on how vehicles appear. Existing studies primarily concentrate their attention on the process of extracting global features based on visual appearance in order to represent the identity of the target vehicle. However, the effect of viewpoint variation is rarely taken into consideration. In this paper, we take into account the view information in order to improve vehicle Re-ID. Additionally, we propose latent view labels through clustering and incorporate view information into deep metric learning in order to address the problem. Additionally, in order to further improve the intra-class compactness of feature space, we develop a center constraint that is more stringent. In addition, we use an orthogonal regularization in order to improve the ability to differentiate between the various types of vehicles. VARID outperforms the state-of-the-art by a significant margin, achieving a mAP of 79.3% on VeRi-776 and 88.5% on VehicleID respectively. The proposed method has been shown to perform noticeably better than methods that are considered to be the state of the art, according to additional and more comprehensive experimental analyses and evaluations on four benchmarks.