A Multiscale Attention Mechanism Based Vehicle Re-Identification

Vehicle re-identification is effected by the vehicle appearance feature, especially for the illumination variation, line of sight occlusion, and image angle change. To solve this problem, a vehicle feature extraction approach based on multi-scale fusion attention mechanism is proposed. First, ResNet-50 is used as backbone network to strengthen the extraction of more discriminative local feature. Second, a plug-and-play spatial attention module is inserted into the backbone network, allowing the backbone network to pay more attention to the local information, which makes the final extracted vehicle features more discriminative. Experimental results on the VeRi-776 dataset indicate that the proposed approach achieved the highest Rank-1 index of 94.28% with mAP of 78.52%, which outperforms the other compared spatial attention mechanism based methods.