文章摘要
陈思文,吴怀宇,陈洋.基于注意力和标签自适应的跨域行人重识别[J].高技术通讯(中文),2022,32(2):143~151
基于注意力和标签自适应的跨域行人重识别
Cross-domain person re-identification based on attention and label adaptation
  
DOI:10.3772/j.issn.1002-0470.2022.02.004
中文关键词: 跨域行人重识别;注意力;标签自适应;知识蒸馏(KD);深度卷积神经网络(DCNN)
英文关键词: cross-domain person re-identification, attention, label adaptation, knowledge distillation (KD), deep convolutional neural network (DCNN)
基金项目:
作者单位
陈思文  
吴怀宇  
陈洋  
摘要点击次数: 569
全文下载次数: 398
中文摘要:
      针对现有的大多数行人重识别算法都依赖于监督训练,而监督训练中人工标注的数据需要昂贵的资源开销从而限制了其在新场景中拓展应用的问题,提出了基于注意力和标签自适应的跨域行人重识别方法。该方法首先对深度卷积神经网络(DCNN)中不同深度的特征层嵌入注意力机制和BNNeck模块,增强模型在不同数据集下对行人的特征表示能力;其次针对没有任何标签的目标数据集,提出了无监督标签自适应方法,将标签信息逐渐扩展至目标数据集中;最后采用知识蒸馏(KD)的方法不断对模型进行微调,使模型逐渐适应新的场景。该方法在Market 1501数据集上的平均精度均值(mAP)为33.1%,在DukeMTMC reID数据集上的mAP为36.1%,与PTGAN、IPGAN等跨域行人重识别算法相比性能有明显提升。
英文摘要:
      In view of the fact that most of the existing person re-identification algorithms rely on supervised training, and the manually annotated data in supervised training require expensive man power and material resources, which limits expanding the application of person re-identification algorithms in new scenarios, a cross-domain person re-identification based on attention and label adaptation is proposed. Firstly, the attention mechanism and BNNeck module are embedded in the feature layers of deep convolution neural network(DCNN) to enhance the feature representation ability of the model to person under different datasets. Secondly, the label adaptation method is proposed to expand the label information to the target dataset. Finally, the method of knowledge distillation (KD) is used to continuously fine-tune the model, so that the model gradually adapts to the new scenarios. The proposed method achieves 33.1% mean average precision (mAP) accuracy on Market-1501 dataset and 36.1% mAP accuracy on DukeMTMC-reID dataset. Compared with the advanced PTGAN, IPGAN and other cross-domain personre-identification algorithms, the performance of the proposed method is obviously improved.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮