融合样本关系优化和重排序的换衣行人重识别

PDF下载HTML

Title:

Clothes-changing person re-identification by sample relationship optimization and re-ranking

作者:

丁钰¹, 毕晓君^2,3

1. 哈尔滨工程大学 信息与通信工程学院, 黑龙江 哈尔滨 150001;
2. 民族语言智能分析与安全治理教育部重点实验室, 北京 100081;
3. 中央民族大学 信息工程学院, 北京 100081

Author(s):

DING Yu¹, BI Xiaojun^2,3

1. College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;
2. Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE, Beijing 100081, China;
3. Department of Information Engineering, Minzu University of China, Beijing 100081, China

关键词:

深度学习; 换衣行人重识别; 局部特征提取; 样本关系优化; Transformer; 短路连接; 圆损失; 重排序

Keywords:

deep learning; clothes-changing person re-identification; local feature extraction; sample relationships optimization; Transformer; shortcut connections; circle loss; re-ranking

分类号:

TP391

DOI:

10.11992/tis.202404005

摘要:

针对换衣行人重识别模型存在局部特征提取能力有限、样本关系优化不足的问题，提出一种融合样本关系优化和重排序的换衣行人重识别模型。首先，设计具有短路连接结构的Transformer模型，将网络的初始特征与深层特征进行融合，来优化每一个样本的特征表示；其次，引入圆损失对优化难度不同的样本赋予不同的权重，更好地优化不同样本之间的关系；最后，设计k′-互近邻重排序策略，对样本间相似性排名进行重新排序，来进一步提高重识别的准确率。在公开的换衣数据集上进行对比实验，结果表明本文提出的模型相比其他先进模型取得了更好的重识别效果。

Abstract:

Current clothes-changing person re-identification models often have limited local feature extraction capabilities and insufficient sample relationship optimization. To this end, this paper proposes a novel clothes-changing person re-identification model by sample relationship optimization and re-ranking. Firstly, we design a Transformer model with shortcut connections to fuse initial and deep features, thereby optimizing the feature representation of each sample. Meanwhile, we use circle loss to assign different weights to sample pairs with varying optimization difficulties, which can better optimize the relationships among different samples. Finally, we designed a k′-reciprocal re-ranking strategy, which can re-rank the similarity rankings and further enhance the re-identification accuracy. Extensive experiments conducted on publicly available datasets LTCC and PRCC demonstrate that comparing with other advanced models, the proposed model achieves better re-identification effect.