Research On Unsupervised Object Detection Methods For Cross-Domain Scenarios

Object detection is one of the basic tasks in the field of computer vision,which aims to automatically identify and locate the object of interest in the image,and has a wide range of applications in all walks of life.Recently,object detection methods based on deep learning have achieved remarkable achievements,but these methods are often inseparable from tedious and time-consuming manual labeling,and the generalization ability in new scenes is generally poor.The unsupervised domain adaptive object detection method solves this problem by transferring the knowledge information learned by the model in the source domain to the target domain,so that the model can obtain the data recognition ability in the unlabeled target domain.Considering the different number of source domains that can be used,unsupervised domain adaptive object detection method can be further divided into two categories:single-source domain adaptive detection method and multi-source domain adaptive detection method.Despite their significant improvements,these methods still have some shortcomings.Therefore,The paper focuses on the study of unsupervised object detection methods for cross-domain scenarios.To address the feature shift problem caused by distillation learning in single-source domain adaptation,and the feature mutual interference problem caused by aligning domain-specific information in multi-source domain adaptation,corresponding solutions are proposed to improve the detection performance of the model in cross-domain scenarios.The main contributions of this paper are summarized as follows:(1)A single source domain adaptive detection method based on feature mutual exclusion is proposed.In order to address the problem of low accuracy of pseudo-labels in the target domain caused by feature shift due to distillation,feature mutual exclusion method is proposed,including feature distribution mutual exclusion and feature attribute mutual exclusion.On the premise of keeping the same category feature distribution of source domain and target domain aligned,feature distribution mutual exclusion encourages the network to exclude features of different categories in terms of distribution,while feature attribute mutual exclusion encourages the network to exclude features of different categories in terms of their content on different attributes.In addition,in order to make the features extracted from the network contain more attributes related to the target domain detection,we also propose a strong-weak augment consistency method to constrain the consistency of the network’s prediction output,which further improves the effect of the feature mutual exclusion method.The method promotes the network to avoid interference from image appearance,improves its focus on foreground regions and addresses the problem of cross-domain category imbalance.Extensive experiments on various cross-domain scenarios show the effectiveness of the proposed method.(2)A multi-source domain adaptive detection method based on domain invariant feature alignment is proposed.In order to address the interference of unique information in each domain during feature alignment,we propose feature disentanglement alignment method and domain invariant information mining method.The feature disentanglement alignment method first decouples the category features of each source domain into cross-domain invariant features and domain-specific features,and then promotes the alignment of target domain features with cross-domain invariant features.The domain invariant information mining method mines the cross-domain invariant foreground regions based on the fusion of multi-layer features,and then adjusts the loss weights of different regions in adversarial learning to focus on the cross-domain invariant foreground information.Additionally,the distillation learning framework is used to further improve the domain adaptive detection performance.Experiments on multiple public datasets demonstrate the effectiveness of the proposed methods.Overall,this paper proposes effective solutions for both single-source domain adaptation tasks and multi-source domain adaptation tasks,which significantly promote the detection performance of the model in the target domain.The research results are of high value,and lay a foundation for the application of object detection in more complex dynamic scenarios.