异质中心角距离约束的多粒度跨模态行人重识别
Multi-granularity cross-modality person re-identification with hetero-center angular constraints
邹业欣 1蒋敏1
作者信息
- 1. 江南大学人工智能与计算机学院,江苏无锡 214122
- 折叠
摘要
针对红外光和可见光图像之间巨大差异导致的跨模态行人重识别正确匹配图像异常困难的问题,提出一种多粒度特征学习的跨模态行人重识别网络(CM-MGN).有效结合全局特征和不同粒度的局部特征,学习更具判别性的行人特征.为有效减小模型的计算复杂度和解决传统三元组损失中异常样本选取的问题,提出基于角距离的异质中心三元组损失(HCAT).在RegDB和SYSU-MM01数据集上的实验结果表明,该方法的Rank-1精度分别达到了 92.33％和62.83％,较其它方法取得了更优性能.
Abstract
Aiming at the problem that it is extremely difficult to correctly match images for cross-modality person re-identification(Re-ID)caused by the huge difference between infrared and visible images,a multi-granularity feature learning cross-modality person Re-ID network(CM-MGN)was proposed.Global features and local features with different granularities were combined effectively,more discriminative pedestrian features were learned.To effectively reduce the computational complexity of the model and solve the problem of selecting outlier samples with the traditional triplet loss,the heterogeneous center triplet loss based on angular distance(HCAT)was proposed.Experiments on RegDB and SYSU-MM01 datasets show that the Rank-1 accuracy of this method is 92.33％and 62.83％,respectively,which is better than that of other methods.
关键词
跨模态/行人重识别/多粒度/局部特征/异质中心/角距离/深度学习Key words
cross-modality/re-identification/multi-granularity/loacl features/hetero-center/angular distance/deep learning引用本文复制引用
基金项目
国家自然科学基金项目(61362010)
国家自然科学基金项目(61201429)
出版年
2024
NETL
NSTL
万方数据