EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

GPRS-Net: Global Statistical Prior Guided Relation-Aware Symmetric Point Cloud Completion

Zelin Li, Shuanfeng Zhao, Zhengheng Li

IEEE Access · 2026 · 14 · 95559-95575

DOI: 10.1109/access.2026.3690668 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Zelin Li und 2 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Qinchuangyuan “Scientist+Engineer” Team Building Project of Shaanxi Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Point cloud completion aims to recover geometrically complete and topologically consistent 3D shapes from partial observations. However, existing methods still struggle with thin structures, sharp features, and strongly anisotropic regions, where local discontinuities, point scattering, and blurred boundaries are likely to occur, especially under weak symmetry and complex topology. This issue largely stems from the lack of reliable global shape constraints during symmetry guided coarse skeleton generation, while isotropic neighborhood aggregation tends to introduce cross part interference and feature confusion in anisotropic areas. To address this problem, we propose a Global Statistical Prior guided Relation Aware Symmetric Network (GPRS Net) for point cloud completion. The proposed method first introduces a directional projection statistics module to extract global statistical priors from the depth distributions of multidirectional projections of the input point cloud. These priors are fused with keypoint features to guide per keypoint transformation and translation regression, producing a structurally consistent symmetric coarse skeleton. A relation aware edge aggregation encoder is then designed to jointly model scale differences and directional deviations on KNN graphs, enabling adaptive edge feature recalibration and aggregation for stronger local geometric representation. Finally, a two stage progressive refinement backbone is adopted to progressively recover fine details while preserving overall structural consistency. Experiments on PCN and MVP demonstrate consistent improvements, as well as better structural stability and detail continuity in challenging cases.

Förderung

  • Qinchuangyuan “Scientist+Engineer” Team Building Project of Shaanxi Province

Zitieren

APA

Li, Z., Zhao, S., & Li, Z. (2026). GPRS-Net: Global Statistical Prior Guided Relation-Aware Symmetric Point Cloud Completion. IEEE Access, 14, 95559-95575. https://doi.org/10.1109/access.2026.3690668

Vancouver

Li Z, Zhao S, Li Z. GPRS-Net: Global Statistical Prior Guided Relation-Aware Symmetric Point Cloud Completion. IEEE Access. 2026;14:95559-95575. doi:10.1109/access.2026.3690668

BibTeX
@article{Li2026,
  author = {Li, Zelin and Zhao, Shuanfeng and Li, Zhengheng},
  title = {GPRS-Net: Global Statistical Prior Guided Relation-Aware Symmetric Point Cloud Completion},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {95559--95575},
  doi = {10.1109/access.2026.3690668}
}
Kostenlos starten →