基于标签互优化的领域自适应无监督行人重识别网络

Unsupervised domain adaptive person re-identification network based on label mutual optimization

摘要:

为解决无监督领域自适应(unsupervised domain adaptation,UDA) 行人重识别(person re-identification,Re-ID) 方法中数据集间的风格差异和聚类过程中伪标签的噪声问题,更好地利用全局与局部信息,本文提出基于标签互优化(label mutual optimization,LMO)的UDA方法。方法首先在特征提取器中加入特征增强模块,提取出更鲁棒性的特征。然后方法使用多分支网络,分别提取行人的全局特征与局部特征,通过计算交叉一致性分数优化迭代网络,从而降低伪标签噪声带来的影响。最后,方法在聚类算法中应用聚类可靠性评价标准,进一步提升网络的性能,并在多个数据集上验证了本文所提出方法的有效性。

Abstract:

To address the issues of style discrepancies between datasets and pseudo-label noise during clustering in unsupervised domain adaptation (UDA) for person re-identification (Re-ID),and to better leverage global and local information,this paper proposes a label mutual optimization-based UDA method.The approach first incorporates a feature enhancement module into the feature extractor to obtain more robust representations.A multi-branch network is then employed to separately extract global and local pedestrian features,with cross-consistency scores calculated to iteratively optimize the network and mitigate the impact of pseudo-label noise.Finally,cluster reliability evaluation criteria are applied in the clustering algorithm to further enhance network performance.Extensive experiments on multiple datasets validate the effectiveness of the proposed method.