| 引用本文: | 王路遥1,2,3；王凤随1,2,3；陈元妹1,2,3.多分支融合变分细化蒸馏的跨模态行人重识别(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2024，41（4）：77-85 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 本文已被：浏览 404次 下载 1262次 |
码上扫一扫！ |
|
|
| 多分支融合变分细化蒸馏的跨模态行人重识别 |
|
王路遥1,2,3；王凤随1,2,3；陈元妹1,2,3
|
|
1. 安徽工程大学 电气工程学院,安徽 芜湖 241000
2. 检测技术与节能装置安徽省重点实验室,安徽 芜湖 241000
3. 高端装备先进感知与智能控制教育部重点实验室,安徽 芜湖 241000
|
|
| 摘要: |
| 目的 针对目前跨模态行人重识别研究中对行人细腻区域关注不足以及网络易受噪声影响的问题,提出一
种多分支融合变分细化蒸馏学习方法。 方法 首先,网络通过多分支聚合不同粒度的全局特征,督促深层网络学习
两种模态的全局信息和细节信息,丰富行人的特征描述符;然后,结合变分细化蒸馏策略,对特征信息进行再压缩,
保留与任务相关的深层信息,同时丢弃无用的干扰物;最后,将网络捕获的不同特征用多种损失函数联合监督,以
提高网络对行人表征的敏感度。 结果 所提方法在 SYSU-MM01 数据集的全搜索模式下,R-1 和
66. 93%和
mAP 分别达到
65. 25%;在 RegDB 数据集的可见光到红外设置下,R-1 和 mAP 分别达到 78. 26%、77. 83%。 结论 通过
消融实验、对比实验和可视化实验,充分验证了所提方法的有效性。 |
| 关键词: 行人重识别 跨模态 多分支聚合 变分细化蒸馏 多损失 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Cross-modal Person Re-identification Based on Multi-branch Fusion Variational Refinement Distillation |
|
WANG Luyao1 2 3； WANG Fengsui1 2 3； CHEN Yuanmei1 2 3
|
|
1. School of Electrical Engineering Anhui Polytechnic University Anhui Wuhu 241000 China
2. Anhui Key Laboratory of Detection Technology and Energy Saving Devices Anhui Wuhu 241000 China
3. Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment Ministry of Education Anhui
Wuhu 241000 China
|
| Abstract: |
| Objective Aiming at the problem of insufficient attention to the delicate area of pedestrians and the
vulnerability of the network to noise in the current cross-modal person re-identification research this paper proposed a
multi-branch fusion variational refinement distillation learning method. Methods Firstly the network aggregated global
features of different granularity through multiple branches urging the deep network to learn the global information and
details of the two modes to enrich the feature descriptors of pedestrians. Then combined with the variational refinement
distillation strategy the feature information was recompressed the deep information related to the task was retained and
the useless interferences were discarded. Finally the different features captured by the network were jointly supervised by
multiple loss functions to improve the sensitivity of the network to pedestrian representation. Results R-1 and mAP reached 66. 93% and 65. 25% respectively with the proposed method in the full search mode of the SYSU-MM01 dataset the R-1 and mAP reached 78. 26% and 77. 83% respectively in the visible to infrared setting of the RegDB
dataset. Conclusion Through ablation experiments comparative experiments and visualization experiments the
effectiveness of the proposed method is fully verified. |
| Key words: person re-identification cross-modality multi-branch aggregation variational refinement distillation multiple
losses |
