基于多样性图像和实例增强的域适应行人搜索

Domain-Adaptive Person Search With Diverse Images and Instance Augmentation

传统全监督行人搜索方法大多仅适用于同一数据域,在未知数据域上泛化力不足。近来研究人员开始探索域适应行人搜索,旨在提升模型在未知目标域的泛化力,需要解决的主要问题是域对齐和可靠正负实例样本生成。基于此,提出基于多样性图像和实例增强的域适应行人搜索方法,从图像和实例多样性增强的角度出发,更好地实现域对齐和可靠样本生成。该方法包含两个新模块:源域图像增强模块和负实例增强的重识别学习模块。源域图像增强模块仅丰富源域图像数据的多样性,进而提高对目标域的域适应力。在负实例增强的重识别学习模块中,引入多样性负实例挖掘模块从目标域和源域挖掘丰富的负实例样本,提高对重识别特征的可判别性。所提模块仅用于训练阶段,不增加测试阶段推理时间。在行人搜索数据集CUHK-SYSU和PRW上进行实验,验证了所提方法的有效性。其中,在PRW测试集上,所提方法取得了40.8%的平均精度均值(mAP),比现有域适应方法 DAPS提高了6.1百分点。

Most traditional fully supervised person search approaches are only applicable to one data domain and have limited generalization ability on unknown data domains.Researchers have recently started studying domainadaptive person search,aiming to improve the generalization ability of the involved model for unknown target domains,where domain alignment and reliable positive and negative generations are the primary challenges.To this end,herein,a domainadaptive person search approach with diverse images and instance augmentation is proposed,which aims to effectively achieve domain alignment and reliable positive and negative generations.This approach introduces two novel modules:sourcedomain image augmentation and negativeenhanced reid learning modules.The former aims to improve the domain adaption ability of the involved model and the detection precision on target domains by only enhancing sourcedomain data diversity.Meanwhile,the latter introduces a diversenegative mining module to enrich the diversity of negatives and improve the discriminability of learned reid features.The proposed modules were only used during training,which did not increase the involved test inference time.Experiments were performed on two widely employed person search datasets:CUHKSYSU and PRW,demonstrating the effectiveness of the proposed approach and its superiority over traditional people search approaches.For instance,the proposed approach achieves mean average precision(mAP)of 40.8%on the PRW test set,indicating higher performance than that of the existing domainadaptive approach DAPS by 6.1 percentage points.