自适应感受野网络的行人重识别
作者:
作者单位:

东北大学

作者简介:

通讯作者:

中图分类号:

TP242

基金项目:

国家自然科学基金项目(面上项目,重点项目,重大项目);中央高校基本科研业务专项基金项目;装备预研领域基金;航天系统仿真重点实验室基金


Adaptive receptive network for person re-identification
Author:
Affiliation:

NEU

Fund Project:

The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan);Fundamental Research Funds for the Central Universities;Foundation of Equipment Pre-Research;Foundation of Science and Technology on Space System Simulation Laboratory

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    行人重识别通常删除特征提取网络中的最后一个空间下采样操作,以增加最后输出特征图的分辨率,保留更多的细粒度特征。然而,这种操作会大幅减小神经网络的感受野,而更大的感受野可以为行人重识别提供更多的上下文信息。同时,在实际的视觉皮层中,相同区域的神经元的感受野是不同的,但当前行人重识别网络的设计大多忽视了这一点。为了解决上述问题,本文提出一种新颖的自适应感受野网络。网络的设计受启发于生物的视觉系统,通过在多分支网络上设置不同大小的感受野,结合注意力机制让网络自行选择合适的感受野特征,实现了网络感受野的自适应,并且采用了分组卷积使得自适应感受野模块更加轻量级。同时在各个分支利用空洞卷积增大感受野,补偿删除最后下采样操作所减少的网络感受野。在公开的大规模数据集上进行了实验,本文算法相比于基线方法有显著的提升,当使用ResNet-50作为特征提取网络时,在DukeMTMC-reID、Market-1501数据集上的Rank1和mAP分别达到了89.2%和76.0%、95.2%和87.2%。与现有方法相比,本文算法在精度有明显的提升.

    Abstract:

    Person re-identification typically removes the last spatial down-sampling operation in the backbone to increase the resolution of the final output feature map and preserve more fine-grained features. However, this operation substantially reduces the size of receptive field, and a larger receptive field can provide more contextual information for person re-identification. At the same time, in the actual visual cortex, the receptive field of neurons in the same region are different, but this is largely ignored by the current design of pedestrian recognition networks. To solve the above problems, this article proposes a novel adaptive receptive field network. The design of the network is inspired by the visual system of living organisms. By setting different sized receptive field on the multi-branch network, combined with the attention mechanism to allow the network to select the appropriate receptive field characteristics, the network receptive field adaptive, and the use of packet convolution makes the adaptive receptive field module more lightweight. The receptive field is also increased in each branch using empty convolution to compensate for the reduction of the network receptive field by deleting the last downsampling operation. Experiments were performed on publicly available large-scale datasets, and the algorithm in this article showed a significant improvement over the baseline approach, with Rank1 and mAP on the DukeMTMC-reID, Market-1501 datasets reaching 89.2% and 76.0%, 95.2% and 87.2%, respectively, when using ResNet-50 as backbone. Compared with the existing methods, the algorithm of this article has a significant improvement in accuracy.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2020-05-01
  • 最后修改日期:2020-10-05
  • 录用日期:2020-10-12
  • 在线发布日期: 2020-12-01
  • 出版日期: