One Graph Can Generalize: Graph-guided Structural Transfer for Source-free Open-set Domain-adaptive Object Detection

Qiang Zhang, Chenxin Li, Panwang Pan, Yifan Liu, Yuzhihuang, Yunlong Lin, Yingying Wang, Luyao Tang, Haochen You, Zheng Zhu, Jiechao Gao

Keywords: Domain Adaptive Object Detection (DAOD); Source-Free Open-Set Learning; Graph-Based Knowledge Transfer

Abstract: Domain Adaptive Object Detection (DAOD) aims to transfer detection capabilities from a labeled source domain to an unlabeled target domain with different visual characteristics. Despite recent advances, existing DAOD methods face significant limitations: they typically operate under closed-set assumptions, require simultaneous access to source and target data, and struggle to differentiate between domain-shifted known objects and genuinely novel categories. To address these real-world challenges, we reformulate the DAOD problem by explicitly considering three distinct shifts: domain distribution shift (\textit{Shift [i]}), open-set class shift (\textit{Shift [ii]}), and source-free transfer shift (\textit{Shift [iii]}). We propose GraphGen , a unified graph-based framework that simultaneously tackles all three shifts by modeling structural relationships between objects, enabling knowledge transfer without source data access through dynamically updated graphs that capture cross-domain similarities. The framework integrates specialized modules for graph-based feature alignment, novelty discovery, and self-regularization to comprehensively address the challenges of source-free open-set domain adaptation. Experiments on benchmark datasets demonstrate that GraphGen outperforms state-of-the-art methods, with significant improvements in novel class discovery while maintaining strong performance on known categories.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 10910