Exploring Objectness Information via Progressively Decoupled Adaptation for Cross-Domain Detection
Yiming Ge, Hui Liu, Ertong Shang, Junzhao Du, Jie Zhao, Zhaocheng Niu · 2025
Recent works in domain adaptive object detection (DAOD) have demonstrated integrating domain adaptation modules into detectors can improve the models' transferability. Nevertheless, given the presence of multiple objects within an image and uncertainties in their localizations, adapting features to diverse objects faces a potential challenge in distinguishing foreground from background. This may hurt the detector's discriminability. Furthermore, traditional methods solely rely on category information for foreground object identification, thereby overlooking the significant category-agnostic information, namely objectness features, and leading to a negative transfer. In this paper, we propose a novel collaborative architecture named Progressively Decoupled adaptation with Objectness Information (PDOI), which individually adapts proposals and semantics, and decouples them from the training of the detector. PDOI sequentially cascades three parts: a category-agnostic adaptor, a category-conditional adaptor, and a detector. The first adaptor focuses on learning category-agnostic objectness features to enhance cross-domain proposal alignments. Subsequently, the category-conditional adaptor is designed to achieve cross-domain semantic alignments by adaptively learning the semantic attributes within objectness features. Finally, we train the detector using pseudo-proposal/category labels generated by two adaptors. We develop a self-feedback optimization mechanism enabling mutual enhancement between the two adapters and detectors. Our approach consistently outperforms state-of-the-art methods across several DAOD benchmarks.