Original title: Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer
Authors: Wenqiao Zhang, Zheqi Lv, Hao Zhou, Jia-Wei Liu, Juncheng Li, Mengze Li, Siliang Tang, Yueting Zhuang
The article delves into Multi-source Active Domain Adaptation (MADA), a complex scenario where data come from various sources. Traditional methods for Active Domain Adaptation (ADA) struggle in this setting due to heightened domain gaps. To tackle this, the authors propose a new framework, Detective. Detective integrates a Dynamic Adaptation (DA) model that learns to adapt across multiple domains, approximating a single-source model. It gauges both domain and predictive uncertainty in the target domain, using evidential deep learning, to select the most informative target samples. Additionally, a diversity-aware calculator enhances sample diversity. Experiments demonstrate Detective’s significant outperformance of existing methods across three domain adaptation benchmarks. This work expands ADA by considering multiple source domains, offering a robust solution for sample selection in complex multi-source scenarios, addressing uncertainty and domain gaps effectively.
Original article: https://arxiv.org/abs/2311.12905