GAP-GDRNet Geometry-aware Monocular 6D Pose Estimation for Spacecraft Using Synthetic Geometric Supervision

Jul 16, 2026

Upload by Maeve

Monocular spacecraft 6D pose estimation remains challenging due to weak texture, thin structures, illumination variation, and occlusion. The work introduces GAP-GDRNet, a geometry-aware RGB framework derived from GDRNet and evaluated on a synthetic single-target spacecraft benchmark. It strengthens geometry-guided regression via AFR before dense geometric prediction for global structure attention and local weak-texture enhancement, and via PGSA within Patch-PnP for patch-level geometric aggregation. Dense supervision is produced with Blender rendering and provides masks, model-coordinate maps, intrinsics, and 6D pose labels, achieving strong pose accuracy and real-time speed.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages26

Reading Time1 h 6 min

AccessFree to download

Introduction

Related Work

Monocular RGB-Based 6D Object Pose Estimation

arXiv :2607 .02360v4 [ cs .CV] 10 Jul 2026 GAP-GDRNet: Geometry-aware monocular 6D pose estimation for spacecraft using synthetic geometric supervision Zongwu Xiea , Yonglong Zhanga , Yifan Yanga , Yang Liua,∗, Guanghu Xiea a State Key Laboratory of Robotics and Systems, Harbin Institute of Technology, Harbin, 150001, Heilongjiang, China Abstract Monocular spacecraft 6D pose estimation remains difficult under weak texture, thin structures, illumination variation, and occlusion. This article presents GAP-GDRNet, a geometry-aware RGB framework built on GDRNet for a single-target synthetic spacecraft benchmark. The method strengthens the geometry-guided regression pipeline at two points. First, AFR is placed before dense geometric prediction to combine global structural attention with local weak-texture enhancement. Second, PGSA is inserted into Patch-PnP to relate downsampled geometric regions before final pose regression. Dense supervision is obtained from a Blender-based rendering and annotation process that provides masks, model-coordinate maps, camera intrinsics, and 6D pose labels. On the self-built spacecraft dataset, GAPGDRNet achieves a rotation error of 1.96◦ , a translation error of 0 .0165 m, and 95 . 16% ADD@0 .02 m, outperforming the reproduced GDR-Net baseline by 3 .88 percentage points while running at 35 .97 FPS. Tests on T-LESS and LM-O further show consistent gains over the reproduced baseline on textureless and occluded non-spacecraft objects. Keywords: Spacecraft pose estimation, monocular vision, 6D pose estimation, geometry-guided regression, synthetic data, on-orbit servicing ∗ Corresponding author Email address: [liuyanghit@hit.edu.cn](liuyanghit@hit.edu.cn) (Yang Liu) 1. Introduction Accurate 6D pose sensing of non-cooperative spacecraft is required in rendezvous, proximity operation, on-orbit servicing, and space situational awareness. A monocular camera is attractive in these tasks because it provides compact measurements with limited hardware complexity, but the image must still support recovery of the target rotation and translation. Spacecraft differ from many terrestrial pose-estimation targets: large smooth panels, thin structural elements, repeated or symmetric parts, hard shadows, and high-contrast backgrounds often appear together. As a result, the visual evidence available from a single RGB image can be weak, unevenly distributed, or partially occluded. Learning-based RGB pose estimators have made dense object-specific geometry usable in monocular settings. Geometry-guided direct regression is particularly relevant here because it retains dense intermediate predictions while avoiding an external PnP solver at test time. As illustrated in Fig. 1, the RGB image is first converted into geometric representations such as 2D– 3D correspondences and surface regions, and the final pose is then recovered by a learned Patch-PnP module. GDR-Net follows this route by predicting geometric representations from the RoI image and feeding them to Patch-PnP for final pose regression [1] . For spacecraft imagery, two limitations remain. Dense coordinate and mask prediction must draw on both the coarse spacecraft layout and small boundary cues, yet illumination changes and weak texture can suppress the latter. Patch-PnP also aggregates geometric evidence after spatial encoding, where distant but related regions, such as the main body and extended panels, may not interact sufficiently through local convolutions alone. GAP-GDRNet keeps the GDR-Net input-output setting and changes the internal representation path. Before dense prediction, attention-based feature refinement (AFR) combines a global grouped coordinate attention branch with a median-enhanced local feature branch. The former emphasizes longrange structure; the latter preserves boundary, contour, and weak-texture responses. Inside Patch-PnP, patch-level geometric self-attention (PGSA) relates downsampled geometric tokens before

FAQs

Why is monocular spacecraft 6D pose estimation difficult?

It is difficult under weak texture, thin structural elements, illumination changes, and occlusion, which makes reliable rotation and translation evidence sparse or uneven in a single RGB image.

What key improvements does GAP-GDRNet make over GDRNet?

GAP-GDRNet adds AFR before dense geometric prediction to enhance global structure and weak-texture/local boundary cues, and inserts PGSA into Patch-PnP to improve patch-level geometric relationships before final pose regression.

How is dense supervision for training obtained?

Supervision is generated using a Blender-based rendering and annotation pipeline, producing masks, model-coordinate maps, camera intrinsics, and 6D pose labels for a synthetic spacecraft dataset.

Jul 16, 2026

Upload by Maeve