Scholarlink

We present a robust multiview point cloud registration framework that addresses two critical limitations of conventional methods: computational inefficiency in dense pose graph construction and vulnerability to outlier contamination in global optimization. Our approach introduces a sparse graph construction paradigm with three key innovations: A tri-scale outlier scoring metric combining density anomalies across spatial resolutions, A persistent saliency measure through hierarchical geometric feature analysis, and An adaptive edge selection criterion balancing overlap completeness and registration reliability. For pose synchronization, we develop a history-aware IRLS variant with residual-driven weight adaptation, dynamically blending initial confidence and iterative refinements through temporal residual fusion. Extensive evaluations on 3DMatch, ScanNet, and ETH datasets demonstrate state-of-the-art performance, achieving 96.4%/84.2% registration recall on 3DMatch/3DLoMatch (+0.5pp/+1.2pp improvement over the current best method SGHR's 95.9%/83.0%) with approximately 70% fewer pairwise registrations. Ablation studies confirm the complementary benefits of our outlier-saliency scoring and adaptive weight iteration mechanisms.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Multiview Point Cloud Registration with Reliable Sparse Graph and Adaptive Reweighting

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zewei Pan,
  • Linsen Li

摘要

We present a robust multiview point cloud registration framework that addresses two critical limitations of conventional methods: computational inefficiency in dense pose graph construction and vulnerability to outlier contamination in global optimization. Our approach introduces a sparse graph construction paradigm with three key innovations: A tri-scale outlier scoring metric combining density anomalies across spatial resolutions, A persistent saliency measure through hierarchical geometric feature analysis, and An adaptive edge selection criterion balancing overlap completeness and registration reliability. For pose synchronization, we develop a history-aware IRLS variant with residual-driven weight adaptation, dynamically blending initial confidence and iterative refinements through temporal residual fusion. Extensive evaluations on 3DMatch, ScanNet, and ETH datasets demonstrate state-of-the-art performance, achieving 96.4%/84.2% registration recall on 3DMatch/3DLoMatch (+0.5pp/+1.2pp improvement over the current best method SGHR's 95.9%/83.0%) with approximately 70% fewer pairwise registrations. Ablation studies confirm the complementary benefits of our outlier-saliency scoring and adaptive weight iteration mechanisms.