Target-Only Relational Learning for Object Detection under Domain and Category Shifts
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Domain Adaptive Object Detection (DAOD) transfers a detector from a labeled source domain to an unlabeled target domain. Deployment becomes substantially harder when target appearance changes, unknown categories occur, and source images cannot be retained. We study these conditions jointly and propose GraphGen, a source-free framework that builds sparse relation graphs over target-region features. Graph propagation regularizes target representations, uncertainty-aware pseudo-labeling separates known and unknown proposals, and graph consistency stabilizes teacher–student adaptation. On Cityscapes to Foggy Cityscapes and Pascal VOC to Clipart, GraphGen improves known-class detection and novel-object recall over the methods included in our common protocol, with gains of up to 2.22 mAP and 1.19 AR points. The results support target-side relational reasoning, while the remaining dependence on pseudo-label quality motivates broader multi-seed and cross-domain evaluation.
- Primary area:
- transfer learning, meta learning, and lifelong learning