Synthetic Data with Global-Local Level for Adversarial-based Domain Adaptation

Abstract

ABSTRACT Domain adaption has attracted quite a lot of interest in object detection. It is hoped that domain adaption can mitigate the discrepancy between the source domain with label-rich and the target domain with no label or label-scarce. In this paper, we address the domain transfer issue from three aspects: (1) We design a synthetic image generator with automatically labelled based on 3d scenes. (2) A structured domain adaptation framework was designed and two domain adapters on global-local levels, respectively, were constructed to mitigate the cross-domain gap. (3) By the virtue of synthetic data, different strategies are adopted to deal with sample transferability in the global and local levels, respectively, to eliminate the performance bottleneck. Recently, adversarial-based methods were used to make alignment with distributions for source and target objects that have been proven effective when adapting object classifiers. We reinforce the two domain adapters to learn a domain classifier with synthetic data and a consistency regularisation is applied to boost the performance. We conduct extensive experiments to demonstrate that the proposed model achieves large performance margins, which show significant advances over the other state-of-the-art models, performing a promotion of 10% −15% mAP on various domain adaptation scenarios.