基于域内域间语义一致性约束的域自适应目标检测方法  

Consistency-aware Domain Adaptive Object Detection via Orthogonal Disentangling and Contrastive Learning

作　　者：钟安雨 王蕊[1,2,3] 张华 邹聪 荆丽桦 ZHONG An-Yu;WANG Rui;ZHANG Hua;ZOU Cong;JING Li-Hua(State Key Laboratory of Information Security,Institute of Information Engineering,Chinese Academy of Sciences,Beijing,100093;Zhejiang Lab,Hangzhou,311100;School of Cyber Security,University of Chinese Academy of Sciences,Beijing,100049)

机构地区：[1]中国科学院信息工程研究所信息安全国家重点实验室,北京100093 [2]之江实验室,浙江杭州311100 [3]中国科学院大学网络空间安全学院,北京100049

出　　处：《计算机学报》2023年第4期827-842,共16页Chinese Journal of Computers

基　　金：国家自然科学基金面上项目(No.62176253);国家自然科学基金企业创新发展联合基金重点支持项目(No.U20B2066);之江实验室开放课题(No.2021KB0AB01)资助.

摘　　要：在目标检测任务中,当训练集和测试集来自不同应用场景时,通常存在检测性能下降问题,这源于不同场景的数据间存在域偏移(domain shift).收集不同场景的数据费时费力,且会增加模型部署成本,降低模型使用效率.针对这一问题,本文从强化特征的语义一致性以获得更好的域无关特征的思路出发,提出基于域内域间语义一致性约束的域自适应目标检测方法.首先,本文考虑了特征解耦过程中的特征域内一致性,提出了一种基于正交分离特征的正交关系一致性约束,该约束能够保留解耦前后特征中的语义信息,以此强化域内特征一致性,从而提升模型识别的准确率.进一步地,本文考虑了在不同域间解耦后特征的域间一致性,引入了基于伪标签的对比学习机制,将来自不同域间的实例级特征进行对齐,以此保证域间特征一致性来提升模型的跨域性能.为验证本文所提出的方法,在本领域常用的数据集Cityscapes-FoggyCityscapes上进行了测试,相对于基线方法本文所提出的方法取得了3.1%的平均准确率(mAP)提升,其中在部分特定子类上提升达到6%;相比较最新方法也有约1%的平均准确率提升.本文还在KITTI-Cityscapes和Sim10K-Cityscapes数据集上测试了所提方法,实验结果表明本方法在其它数据集上也能取得良好的域自适应效果.Traditional object detection methods suffer from performance degradation when the training and test data are from different domains,for example,photos from a sunny day and a cloudy day are two different domains,and an object detection model trained on a sunny day usually performance not well on a cloudy day.This is caused by the domain shift between two domains.Collecting data for every single domain is time-consuming and laborious,which will increase the cost of model deployment and reduce the efficiency of the model being used.Aiming at this problem,the domain adaptive object detection method is proposed.Most domain adaptation methods eliminate domain shifts by finding domain-invariant feature representations in two domains.Although existing domain adaptation methods have achieved great success,there are still differences between the domain-invariant features extracted from the source domain and the target domain,which lead to poor performance when the model uses domain-invariant features from the target domain.Enlightened by the idea of strengthening the semantic consistency of features to obtain better domain-invariant features,this paper proposes a Consistency-aware Domain Adaptive object detection network(ConDA)with orthogonal disentangling and contrastive learning.Specifically,this paper first proposes an orthogonal relation consistency constraint based on the orthogonal disentangled features,which can better improve the intradomain consistency and the transferability of the model.Orthogonal constraints are applied in the feature disentangling process to keep domain-invariant and domain-specific features different.Based on the orthogonal constraints,the relationship consistency loss can be applied by calculating the instance-level feature relationship consistency before and after feature disentangling and then constraining them to be the same.This loss can retain the semantic information during feature disentangling and strengthen the intra-domain consistency of the feature,thus improving the transferabili

关 键 词：域自适应 目标检测 深度学习 特征解耦 对比学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]