ZODS-RS -- Zero-training Oriented Detection & Segmentation for Remote Sensing

arXiv:2606.10769v1 Announce Type: new Abstract: Remote-sensing and UAV applications need models that generalize across platforms and viewpoints without task-specific training. Yet training-free pipelines often falter on oriented geometry, scale/rotation variation, and crowded ports or airfields, and rarely unify detection and segmentation. We introduce ZODS-RS, a training-free, closed-form pipeline that outputs horizontal boxes (HBB) and instance masks. Built on DINOv3 dense features and SAM-style proposals, ZODS-RS chains: PP (prototype purification via Tyler covariance), R-SEM (rotation-scale equivariant matching with separable kernels and global Hungarian assignment), and UAM (uncertainty-aware pixelwise merging with adaptive priors and optional negative prototypes). A lightweight CWLA fuses multiple DINOv3 layers. On FAIR1M (HBB) we obtain mAP_{0.50:0.95}=mathbf{13.06} and AP_S=mathbf{2.93} (class-averaged over ship/airplane); on xView (HBB) we report mAP=mathbf{16.69}. On our UAV dataset, ZODS-RS achieves mask mIoU=mathbf{31.10} and improves small-object AP by mathbf{+30.70} over Grounded-SAM on a single 5090. This work offers a unified, no-training solution for horizontal-box detection plus instance segmentation in aerial imagery; provides explicit closed-form formulations for PP/R-SEM/UAM tightly coupled with DINOv3; and demonstrates consistent gains on small and crowded targets and under cross-domain shifts while keeping deployment simple.

Source: arXiv cs.CV | 2026-06-10