一种基于中心和关系感知的跨域目标检测网络

Center-guided Relation-aware Network for Cross-domain Object Detection

域偏移已成为跨域目标检测领域一个棘手的问题.当把在源域训练好的检测器用于目标域时,由于源域和目标域的分布差异,检测器往往会有显著的性能下降.为了处理上述问题,本文提出了一个基于类别中心和关系感知的跨域目标检测模型,该模型通过图卷积的方式同时对域差异信息,类别信息和关系信息进行建模.本文所提出的模型有以下几个优点:1)据我们所知,这是跨域目标检测方向里第一个同时对上述3种信息建模的网络;2)提出的模型设计了包括域对齐,类别中心对齐,关系感知模块在内的3种机制,可以有效学习域不变的特征以减小域偏移.在4个标准数据集上的结果表明本文模型的结果可以与当前最好的模型相当甚至超过当前最好的模型.

Domain shift has become a thorny issue in cross-domain object detection,which has led to a significant performance drop when adapting an object detector trained with source domain to target domain.To address the above issue,we propose an end-to-end Center-guided Relation-aware Network(CRN)for cross-domain object detection by modeling domain discrepancy,semantic category information and relation information jointly via Graph Convolutional Networks(GCNs).The proposed CRN model enjoys several merits.First,to the best of our knowledge,this is the first work to model the three kinds of information jointly in a deep model for cross-domain object detection.Second,the proposed model has designed three effective mechanisms including domain-level alignment,category center alignment,and relation-aware module,which can help learn domain invariant yet discriminative features to reduce domain discrepancy through the information transferring via GCNs.Extensive experimental results on standard benchmarks demonstrate that our algorithm performs favorably against state-of-the-art cross-domain object detection methods.