Domain Adaptation for Object Detection Based on Domain-Aware Prompting in Remote Sensing Imagery

Domain adaptation (DA) is of critical importance in practical remote sensing object detection, aiming to address the performance degradation caused by significant domain discrepancies between a labeled source domain and an unlabeled target domain. The DA method enables the alignment of cross-domain features, thus improving detection performance in the target domain while avoiding the high cost of manual annotation. However, most existing methods focus on mitigating domain bias by optimizing discriminative visual encoders. These coarse feature alignment strategies often overlook domain-Aware semantic consistency, which leads to feature misalignment, especially for the remote sensing scene with small objects and diverse backgrounds. To address this problem, we propose a novel domain-Aware prompting (DAP) framework to learn domain-invariant semantics by modeling the learnable domain-Aware prompts. Specifically, we first design a reciprocal prompt interaction learning (RPIL) module to generate domain-Aware semantics for each domain by modeling reciprocal interactions between textual and image modalities. Subsequently, we develop a PDPA module to capture shared domain prompt semantics from both local and global levels by constructing domain prompt prototypes. Extensive experiments demonstrate the efficacy of our proposed approach, achieving superior performance on challenging remote sensing datasets. The code is publicly accessible at https://github.com/XZhang878/DAP/