GeoAdapt: Fine-Grained Keypoint Localization via Deformable Feature Refinement for Ground-Based Optical Remote Sensing

Abstract

Ground-based optical remote sensing of aerial targets at kilometer-scale standoff distances requires accurate keypoint localization for six-degree-of-freedom (6-DoF) pose recovery under variable illumination, motion blur, and atmospheric degradation. Many lightweight detectors use fixed-kernel convolutions, whose spatially invariant sampling may limit adaptation to heterogeneous target geometries and spatially varying image degradation. We introduce GeoAdapt, a compact keypoint detection framework that inserts deformable convolution v2 (DCNv2) modules between the feature pyramid network (FPN) neck and the detection head. GeoAdapt also replaces the standard object keypoint similarity (OKS) loss with a combination of Wing Loss and Bone Loss. The complete model contains 5.95 M parameters, 47.9% fewer than the You Only Look Once version 8 small pose model (YOLOv8s-pose). On a synthetic ground-based optical remote sensing benchmark, GeoAdapt achieved a percentage of correct keypoints (PCK) at a threshold of 0.05 times the bounding-box diagonal (PCK@0.05D) of 89.3% and a rotation error of 11.6°, improving PCK by 11.7 percentage points over YOLOv8s-pose. Zero-shot evaluation on manually annotated real ScanEagle and Matrice 200 imagery showed consistent advantages over YOLOv8s-pose and YOLO11s-pose in all six test scenarios. A factorial ablation indicated a positive interaction between DCNv2 and the Wing+Bone loss.