Multimedia Systems Компьютерные сети и коммуникации

Multi-candidate geometry collaboration for single-anchor 6-DoF pose estimation

Gu, Shutong, Sun, Zixuan, Shuai, Hui
Springer Nature · 2026-11-16 · DOI 10.1007/s00530-026-02789-7

Аннотация

Reliable relative pose estimation of previously unseen objects from limited references is important for RGB-D scene analysis and 3D object alignment. Single-anchor 6-DoF pose estimation addresses this task using only one RGB-D anchor observation, without requiring an object-specific CAD model or multi-view scan. However, its accuracy depends heavily on the anchor-side geometry: a single view provides incomplete surface evidence, while reconstructed regions may contain errors that propagate into query registration. Reconstructions conditioned on different visual cues can offer complementary geometric hypotheses, but their discrepancies in shape, scale, and alignment make direct fusion unreliable. To address these challenges, we propose MGC-6D, a multi-candidate geometry collaboration framework that preserves alternative reconstructions and exploits their complementary information. RaySt3R and InstantMesh generate RGB-D-conditioned and RGB-conditioned geometry candidates, respectively. Bidirectional ICP evaluates their compatibility and guides the conditional construction of an observation-constrained, topology-preserving fused candidate. Rather than fixing one geometry at the anchor stage, the framework registers all available candidates to the query and selects the final relative pose using a unified score combining anchor reliability with query-view depth, silhouette, and point-cloud consistency. Experiments on HO3D, Toyota-Light, and REAL275 yield average recall (AR) scores of 50.1, 60.6, and 73.9, respectively. These results indicate the value of retaining and evaluating multiple geometry hypotheses for offline single-anchor pose estimation. The code is available at https://github.com/Lumos000/MGC-6D.

Читать у издателя →