基于领域自适应的多源多标签行人属性识别

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

当前行人属性识别(PAR)公开数据集中属性标注繁杂且采集场景多样,各数据集中行人属性差异较大,进而导致公开数据库已有的标记信息数据难以直接应用到PAR实际问题中。针对上述问题,提出一种基于领域自适应的多源多标签PAR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one-hot编码加权算法,将多数据集中共有属性的标签对齐;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准确率。通过所提出的特征对齐和标签对齐算法,可有效解决PAR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StrongBaseline在平均准确率上分别提升了1.22、1.62和1.53个百分点,说明该方法在跨数据集PAR中具有一定的优势。

The current public datasets of Pedestrian Attribute Recognition(PAR)have the characteristics of complicated attribute annotations and various collection scenarios,leading to the large variations of the pedestrian attributes in different datasets,so that it is hard to directly utilize the existing labeled information in the public datasets for PAR in practice.To address this issue,a multi-source and multi-label PAR method based on domain adaptation was proposed.Firstly,to transfer the styles of the different datasets into a unified one,the features of the samples were aligned by the domain adaption method.Then,a multi-attribute one-hot coding and weighting algorithm was proposed to align the labels with the common attribute in multiple datasets.Finally,the multi-label semi-supervised loss function was combined to perform joint training across datasets to improve the attribute recognition accuracy.The proposed feature alignment and label alignment algorithms were able to effectively solve the heterogeneity problem of attributes in multiple PAR datasets.Experimental results after aligning three pedestrian attribute datasets PETA,RAPv1 and RAPv2 with PA-100K dataset show that the proposed method improves the average accuracy by 1.22 percentage points,1.62 percentage points and 1.53 percentage points respectively,compared to the method StrongBaseline,demonstrating that this method has a strong advantage in cross dataset PAR.