Semantic Guided Feature Disentanglement and Reconstruction for Domain Adaptive Object Detection

Domain Adaptive Object Detection (DAOD) aims to mitigate the domain gap between labeled source domains and unlabeled target domains. Recent advances in visionlanguage pre-trained models (VLMs) have opened new avenues for enhancing DAOD generalization through crossmodal alignment. However, existing VLM-based methods often perform either unimodal feature adaptation or direct visual-textual fusion, which entangles domain-invariant and domain-specific information, thereby hindering effective feature transfer and robustness under domain shifts. To address these challenges, we propose a Semantic-Guided Feature Disentanglement and Reconstruction (SG-FDR) framework that introduces semantic supervision to explicitly decompose visual features into domain-invariant and domain-specific components, preserving cross-domain consistency while capturing target-specific variations. Specifically, the Semantic-Guided Disentanglement Module employs domain-adaptive prompts to guide semantic-level feature separation, whereas the Adaptive Feature Reconstruction Module selectively distills transferable cues from domain-specific features and integrates them with invariant features to enhance adaptability. Extensive experiments on multiple DAOD benchmarks demonstrate that our method consistently outperforms state-of-the-art approaches, achieving robust performance under complex domain shifts.