基于影响函数的跨领域小样本的行人重识别方法

Influence function based method on cross-domain few-shot person Re-ID problem

在行人重识别领域,如何快速在一个新城市进行行人重识别系统的部署为行人重识别领域带来了巨大的挑战。一般情况下,在新城市中没有足够的标注数据来训练一个强大的行人重识别模型,只能依赖少量由目击者提供的罪犯照片。首次对该实际应用场景进行正式研究,将其定义为跨领域小样本行人重识别问题,并仔细讨论分析其与现有行人重识别场景的异同。随后,提出一种基于影响函数的样本权重法来指导模型的训练,并在三个公开数据集Market、Duke和CUHK上对方法进行实验对比。实验结果表明,该方法可以有效地处理不同数据集之间的偏置,性能也超过了已有方法。

In person re-identification(Re-ID)community,how to deploy a Re-ID system quickly in a new city to help the police catch a criminal is a challenging task.In general,the new cities don’t have enough labelled data to train an excellent model,only relying on a tiny number of criminals’pictures provided by witnesses.This paper firstly formulated the scenario as a cross-domain few-shot problem and discussed the difference between conventional supervised Re-ID and unsupervised Re-ID.Then it introduced a re-weighting instance method based on influence function to guide the training procedure of the Re-ID model.Finally,it evaluated the proposed method on public datasets including Market,Duke,and CUHK.Extensive experimental results show that the method can effectively address the domain bias of different datasets and the absence of labelled data on the target dataset,achieving the state of the art.