OA-Pro: Learning Object-Aware Detection Head with Prompt Tuning for Domain Adaptive Object Detection

Domain adaptive object detection (DAOD) aims to generalize detectors trained on a labelled source domain to an unannotated target domain. Existing methods reduce the domain bias within the detection backbone and overlook the domain bias present in the detection head, e.g., domain-level discrepancy in different domains, and object-level inter-domain differences between objects. Inspired by highly generalized Visual-Language Models (VLMs), applying a VLM as robust detection backbone followed by a detection head that can be adjusted adaptively for different domains and objects, i.e., an object-aware detection head, is reasonable to learn a detector discriminative in both domain and object-level. To achieve this, we propose a novel DAOD framework named Object-Aware detection head with Prompt tuning (OA-Pro) to increase the discriminative ability in domain and object-level. Specifically, OA-Pro learns domainadaptive and object-adaptive prompt to exploit object-aware domain knowledge. To learn domain-invariant and domainspecific knowledge, the domain-adaptive prompt comprises the domain-invariant tokens, domain-specific tokens, domain-related textual description, and class label. Meanwhile, the objectadaptive prompt consists of the category-wise prompt and objectwise projectors, extracting object-aware domain knowledge in a coarse-to-fine paradigm. Comprehensive experiments over multiple DAOD benchmarks demonstrate that the proposed OA-Pro can produce an effective detection head for boosting DAOD.