Abstract In recent years, deep learning-based object detection has achieved significant advances, enabling its widespread deployment across diverse real-world applications. Conventional approaches typically assume consistent data distributions between the source and target domains, a premise that often fails to hold in practical scenarios, leading to substantial performance degradation in detection systems. Consequently, the domain shift problem has emerged as a critical research focus in the computer vision community, as evidenced by the proliferation of innovative methods presented annually in top-tier conferences and journals. Despite these advancements, comprehensive surveys dedicated specifically to domain-adaptive object detection (DAOD) remain scarce.To address this gap, this paper provides a detailed survey of DAOD algorithms. We first introduce foundational concepts, including deep domain adaptation and object detection, then systematically decompose DAOD into two subproblems to elucidate its developmental trajectory from a fundamental perspective. We further present the latest advances in DAOD algorithms, categorizing them into feature alignment, adversarial training, reconstruction-based methods, knowledge distillation, and other emerging paradigms. For each category, we analyze the research landscape and compare performance across benchmark datasets. Finally, through a thorough review and synthesis of existing approaches, we outline promising future research directions for DAOD.
Tingting et al. (2025) studied this question.