An Improved Dynamic Alignment Method for Vehicle Re-Identification

Abstract

In this paper, we propose an improved vehicle re-identification method based on the combination between the AlignedRelD and the Stochastic Weight Averaging (SWA). AlignedRelD extracts a global feature and local features of a vehicle’s image and performs joint learning. Local automatic alignment is achieved by computing the shortest path between the two sets of local features, so that global feature learning can benefit from local feature learning. By running an optimizer with a high constant learning rate, the SWA averages the weight of the model to ensure that a better weight combination can be found. Our method achieves rank-1 accuracy of 94.4% on VeRi-776 and 95.1% on VehiclelD(small), outperforming state-of-the-art methods by a large margin. In order to better solve the task of vehicle re-identification in residential area, we have made the Oeasy-Parking dataset and experimented with our methods, and achieved good results.