Image-Pair Correlation Learning for Vehicle Re-Identification

Vehicle re-identification plays an important role in intelligent transportation systems. It aims to identify vehicles with the same identity between images captured by different cameras. How to reasonably estimate the similarity between features plays an important role in vehicle re-identification. Traditional vehicle re-identification methods suffer from high intra-class difference and low inter-class difference due to view difference, which poses a significant challenge for accurate vehicle re-identification. Many Siamese network-based methods for vehicle re-identification can learn intra- and inter-class distances, but they tend to overlook similarity metrics between classifiers and similarity learning of element-level features, which could further enhance similarity learning between images. To address this issue, we propose an image-pair correlation learning network for vehicle re-identification. Imposing constraints on the distances between features in different ways to reduce the intra-class distance and increase the inter-class distance. We design a classifier similarity estimation module and a similarity metric module of features at element-level to learn the similarity of images from different views. Extensive experiments on AI City Challenge 2020 Track2 dataset and VeRi-776 dataset demonstrate the effectiveness of our methods.