CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection
#Source-Free Domain Adaptive Object Detection #CGSA #Object-Centric Learning #DETR #Slot-Aware Adaptation #Cross-Domain #ICLR 2026
📌 Key Takeaways
- Researchers developed CGSA, a novel framework for source-free domain adaptive object detection
- CGSA integrates Object-Centric Learning into SF-DAOD using slot-aware adaptation with DETR-based detection
- The approach outperforms previous methods across multiple cross-domain datasets
- The research highlights potential for object-centric design in privacy-sensitive adaptation scenarios
📖 Full Retelling
Researchers Boyang Dai, Zeng Fan, Zihao Qi, Meng Lou, and Yizhou Yu introduced CGSA, a novel framework for source-free domain adaptive object detection, in a paper submitted to arXiv on February 26, 2026, aiming to overcome limitations in current approaches by incorporating object-level structural cues that have been previously overlooked. Source-Free Domain Adaptive Object Detection (SF-DAOD) focuses on adapting detectors trained on labeled source domains to work with unlabeled target domains without retaining any source data. The researchers identified that most existing methods concentrate on tuning pseudo-label thresholds or refining teacher-student frameworks while neglecting important object-level structural cues within cross-domain data. CGSA represents the first framework to integrate Object-Centric Learning into SF-DAOD by combining slot-aware adaptation with DETR-based detection technology. The approach integrates a Hierarchical Slot Awareness module into detectors to progressively break down images into slot representations that serve as visual priors. These slots are then directed toward class semantics through a Class-Guided Slot Contrast module, which maintains semantic consistency and promotes domain-invariant adaptation. Extensive experiments conducted across multiple cross-domain datasets demonstrated that CGSA outperforms previous SF-DAOD methods. The research highlights the potential of object-centric design in privacy-sensitive adaptation scenarios, as CGSA can effectively adapt to new domains without requiring access to the original source data. The paper has been accepted by the ICLR 2026 conference, indicating the significance and quality of this contribution to the field of computer vision and artificial intelligence.
🏷️ Themes
Computer Vision, Machine Learning, Privacy-Preserving AI
📚 Related People & Topics
Secretary of State for the Environment, Transport and the Regions
Defunct UK Cabinet ministerial position
The Secretary of State for the Environment, Transport and the Regions was a United Kingdom Cabinet position created in 1997, with responsibility for the Department of the Environment, Transport and the Regions (DETR). The position and department were created for John Prescott by merging the position...
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
--> Computer Science > Computer Vision and Pattern Recognition arXiv:2602.22621 [Submitted on 26 Feb 2026] Title: CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection Authors: Boyang Dai , Zeng Fan , Zihao Qi , Meng Lou , Yizhou Yu View a PDF of the paper titled CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection, by Boyang Dai and 4 other authors View PDF HTML Abstract: Source-Free Domain Adaptive Object Detection (SF-DAOD) aims to adapt a detector trained on a labeled source domain to an unlabeled target domain without retaining any source data. Despite recent progress, most popular approaches focus on tuning pseudo-label thresholds or refining the teacher-student framework, while overlooking object-level structural cues within cross-domain data. In this work, we present CGSA, the first framework that brings Object-Centric Learning into SF-DAOD by integrating slot-aware adaptation into the DETR-based detector. Specifically, our approach integrates a Hierarchical Slot Awareness module into the detector to progressively disentangle images into slot representations that act as visual priors. These slots are then guided toward class semantics via a Class-Guided Slot Contrast module, maintaining semantic consistency and prompting domain-invariant adaptation. Extensive experiments on multiple cross-domain datasets demonstrate that our approach outperforms previous SF-DAOD methods, with theoretical derivations and experimental analysis further demonstrating the effectiveness of the proposed components and the framework, thereby indicating the promise of object-centric design in privacy-sensitive adaptation scenarios. Code is released at this https URL . Comments: The paper has been accepted by the conference ICLR 2026 Subjects: Computer Vision and Pattern Recognition (cs.CV) ; Artificial Intelligence (cs.AI) Cite as: arXiv:2602.22621 [cs.CV] (or arXiv:2602.22621v1 [cs.CV] for this version) https://doi.org/10.48550/arXiv.2602.226...
Read full article at source