Multi-source and multi-label pedestrian attribute recognition based on domain adaptation

  

  • Received:2021-06-04 Revised:2021-08-17 Published:2021-09-17
  • Contact: Xia ZhenYu

基于领域自适应的多源多标签行人属性识别

程南江1,2,余贞侠1,陈琳2,乔贺辙3   

  1. 1. 成都信息工程大学
    2. 中国科学院重庆绿色智能技术研究院
    3. 中科院重庆绿色智能研究所
  • 通讯作者: 余贞侠

Abstract: Abstract: The current public datasets of pedestrian attribute recognition (PAR) have the characteristic of complicated attribute annotations and various collection scenarios, leading to the large variations among the pedestrian attributes datasets, so that it is not easy to directly utilize the heterogeneous labeled information in existing public datasets for PAR in practice. To address this issue, a multi-source and multi-label pedestrian attribute recognition method based on domain adaptation is proposed in this study. Firstly, to transfer the style between the different datasets, the features of the samples are al igned by the domain adaptive method. Then, a multi-attribute one-hot coding weighting algorithm was proposed to align the common labels in multiple datasets. Finally, the multi-label semi-supervised loss function was proposed to perform joint training across datasets to improve the attribute recognition accuracy. The proposed feature alignment and label alignment algorithms can effectively solve the heterogeneous problem of attribute annotation in multiple datasets for PAR. Three pedestrian attribute datasets including PETA, RAPv1 and RAPv2, are used to align and test on PA-100K dataset, the experimental results show that the proposed method improves the average accuracy metrics by 1.22 percentage points, 1.62 percentage points and 1.53 percentage points, respectively, over the baseline model, demonstrating that the proposed method has a strong advantage in PAR with the multi-source datasets.

Key words: Pedestrian Attribute Recognition, Multi-label Learning, Domain Adaption, Deep Learning, Convolutional Neural Networks

摘要: 当前行人属性识别公开数据集中属性标注繁杂,采集场景多样,导致各数据集中行人属性差异较大,公开数据库已有的现有标记信息数据不易直接运用到实际问题中。针对上述问题,提出了一种基于领域自适应的多源多标签行人属性识别方法。首先通过领域自适应方法对样本进行特征对齐,完成多个数据集之间的统一风格转换;接着提出多属性一位有效(one-hot)编码加权算法,将多数据集中共有属性的标签进行对齐;最后结合多标签半监督损失函数,进行跨数据集联合训练以提高属性识别精度。通过提出的特征对齐和标签对齐算法,可有效解决行人属性识别多数据集中属性异构性问题。将三个行人属性数据集PETA、RAPv1和RAPv2分别与PA-100K数据集对齐后的实验结果表明,所提出的方法对比基准模型在平均准确率指标分别提升了1.22个百分点、1.62个百分点和1.53个百分点,证明其在跨数据集行人属性识别中具有一定的优势。

关键词: 行人属性识别, 多标签学习, 领域自适应, 深度学习, 卷积神经网络

CLC Number: