Intrinsic feature extraction for unsupervised domain adaptation

Abstract

PurposeUnsupervised domain adaptation object detection not only mitigates model terrible performance resulting from domain gap, but also has the ability to apply knowledge trained on a definite domain to a distinct domain. However, aligning the whole feature may confuse the object and background information, making it challenging to extract discriminative features. This paper aims to propose an improved approach which is called intrinsic feature extraction domain adaptation (IFEDA) to extract discriminative features effectively.Design/methodology/approachIFEDA consists of the intrinsic feature extraction (IFE) module and object consistency constraint (OCC). The IFE module, designed on the instance level, mainly solves the issue of the difficult extraction of discriminative object features. Specifically, the discriminative region of the objects can be paid more attention to. Meanwhile, the OCC is deployed to determine whether category prediction in the target domain brings into correspondence with it in the source domain.FindingsExperimental results demonstrate the validity of our approach and achieve good outcomes on challenging data sets.Research limitations/implicationsLimitations to this research are that only one target domain is applied, and it may change the ability of model generalization when the problem of insufficient data sets or unseen domain appeared.Originality/valueThis paper solves the issue of critical information defects by tackling the difficulty of extracting discriminative features. And the categories in both domains are compelled to be consistent for better object detection.