Alleviating the Equilibrium Challenge with Sample Virtual Labeling for Adversarial Domain Adaptation

Wenxu Shi, Bochuan Zheng

Abstract: Numerous domain adaptive object detection (DAOD) methods leverage domain adversarial training to align the features to mitigate domain gap, where a feature extractor is trained to fool a domain classifier in order to have aligned feature distributions. The discrimination capability of the domain classifier is easy to fall into the local optimum due to the equilibrium challenge, thus cannot effectively further drive the training of feature extractor. In this work, we propose an efficient optimization strategy called \underline{V}irtual-label \underline{F}ooled \underline{D}omain \underline{D}iscrimination (VFDD), which revitalizes the domain classifier during training using \emph{virtual} sample labels. Such virtual sample label makes the separable distributions less separable, and thus leads to a more easily confused domain classifier, which in turn further drives feature alignment. Particularly, we introduce a novel concept of \emph{virtual} label for the unaligned samples and propose the \emph{Virtual}-$\mathcal{H}$-divergence to overcome the problem of falling into local optimum due to the equilibrium challenge. The proposed VFDD is orthogonal to most existing DAOD methods and can be used as a plug-and-play module to facilitate existing DAOD models. Theoretical insights and experimental analyses demonstrate that VFDD improves many popular baselines and also outperforms the recent unsupervised domain adaptive object detection models.

Primary Subject Area: [Experience] Multimedia Applications

Secondary Subject Area: [Experience] Multimedia Applications

Relevance To Conference: Domain adaptation significantly enhances the performance of multimedia applications, particularly in object detection tasks. Current domain adaptation methods often rely on domain adversarial training to align features and mitigate domain gaps. However, the effectiveness of these methods is limited by challenges such as falling into local optima during training. Our proposed approach, Virtual-label Fooled Domain Discrimination (VFDD), introduces an efficient optimization strategy that revitalizes the domain classifier using virtual sample labels. By making the separable distributions less distinct, VFDD ensures a more easily confused domain classifier, driving further feature alignment. This innovative technique overcomes the equilibrium challenge and significantly improves upon existing DAOD methods. VFDD can be seamlessly integrated into existing models, enhancing their performance and outperforming recent unsupervised domain adaptive object detection models.

Submission Number: 948