Transformer框架下面向车辆重识别的特征对齐与判别性增强

Feature Alignment and Discriminative Enhancement for Vehicle Re-Identification under the Transformer Framework

车辆重识别在近几年取得了显著研究进展,但由于不同摄像头下汽车外观特征具有歧义性,这对识别性能的提升带来了极大挑战。为此,在Transformer框架下,提出面向车辆重识别的特征语义对齐与判别性增强方法。该方法首先使用预训练后的车辆姿态估计模型实现对车辆关键点的提取,然后利用关键点携带的语义信息,根据图像块的坐标,设计一种特征聚集方法,将Transformer中具有相同语义的token划归到同一组内,实现特征的语义对齐,提升特征鲁棒性与判别性。此外,考虑到不同语义特征之间具有一定的内在关系,进一步构建图卷积网络来进一步优化特征质量。所提出的方法在两个公开的大型车辆数据集上均表现出了先进的效果,证明了方法的有效性以及优越性。

Although significant research progress has been made in vehicle re-identification in recent years,due to the ambiguity of vehicle appearance characteristics under different cameras,this has brought great challenges to the improvement of recognition performance.For this reason,this paper proposes a feature semantic alignment and discriminative enhancement method for vehicle re-identification under the framework of transformer.This method first uses the pre-trained vehicle attitude estimation model to extract the key points of the vehicle.Then,using the semantic information carried by the key points,according to the coordinates of the image block,a feature aggregation method is designed,and the tokens with the same semantics in the transformer are grouped into the same group,which realizes the semantic alignment of the features and improves the feature robustness.In addition,considering the inherent relationship between different semantic features,a graph convolutional network is further constructed to further optimize the feature quality.The method in this paper has shown advanced effects on three publicly available large-scale vehicle data sets,which proves the effectiveness and superiority of the method in this paper.