位置感知注意力及其在行人重识别中的应用-《计算机技术与发展》

文章信息/Info

Title:: A Novel Position-aware Attention Module and Its Use inPerson Re-identification

作者:: 陈江萍¹ ; 张索非² ; 宋越³ ; 吴晓富¹ ; 林嘉¹; 1. 南京邮电大学通信与信息工程学院,江苏南京 210003;
2. 南京邮电大学物联网学院,江苏南京 210003;
3. 95958 部队,上海 200120

Author(s):: CHEN Jiang-ping1 ; ZHANG Suo-fei2 ; SONG Yue3 ; WU Xiao-fu1 ; LIN Jia1; 1. School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;
2. School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;
3. 95958 Troops,Shanghai 200120,China

Keywords:: position encoding; non-local block; position-aware attention module; feature pyramid branch; person re-identification

摘要:: 行人重识别领域的众多工作都表明,采用多分支神经网络搭配注意力模块是一种实现高性能特征嵌入的有效方式。传统方案主要关注于多分支网络结构的设计,而在注意力机制的设计上存在明显不足,如当前注意力机制缺乏对特征位置信息的有效挖掘和利用。为此,该文在多尺度特征金字塔分支( Feature Pyramid Branch,FPB) 网络的框架下,分析了不同注意力模块的引入对系统性能的影响;在此基础上,讨论了两种在注意力机制中融入位置信息的方法,提出了一种新的位置感知注意力模块,该模块具有即插即用的优点,便于融入各种主干网络。在多个流行行人重识别标准数据集上的实验表明,融入位置感知注意力模块的 FPB 网络相比于原 FPB 网络,仅需增加 0. 29 M 参数就可以显著提升最终的模型识别准确率:rank-1 在 Market1501 上提高 0. 7 百分点,在 DukeMTMC 上提高 1. 5 百分点,在 CUHK03-Labeled 上提高 2. 4百分点,在 CUHK03-Detected 上提高 3. 8 百分点。

Abstract:: Recent work in the field of person re - identification shows that using multi - branch neural networks is an effective way toachieve high performance feature-embedding. Traditional schemes mainly focus on the design of various efficient multi-branch networkstructures,but there are obvious deficiencies in the design of attention mechanism. For example,the current attention mechanism lacks theeffective mining and utilization of feature position information. Therefore,we investigate the effects of using different attention modulesin a multi-scale Feature Pyramid Branch ( FPB) network. Then,two methods are discussed for introducing position information intoattention modules, and a novel position - aware attention module is proposed, which can be used into various backbone networks.Experiments on popular person re-identification datasets show that compared with the original FPB network,the proposed FPB networkwith position-aware attention could achieve significantly better performance with 0. 29 M parameters increase in model,the gain in therank-1 accuracy is about 0. 7% on Market1501,1. 5% on DukeMTMC,2. 4% on CUHK03-Labeled and 3. 8% on CUHK03-detected.