Scholarlink

<p>We propose <i>SLNMapping</i>, a novel neural mapping framework for super lightweight reconstruction in large-scale scenes. The core is a new ultra-compact neural map representation composed of a set of feature-independent local signed distance functions (SDFs) with outstanding expressiveness. To support efficient optimization, we introduce a novel parallel local SDF detection algorithm that enables real-time updates of local SDF states. Based on the excellent representation, we develop a three-stage mapping strategy for efficient, accurate, and lightweight large-scale reconstruction from streaming LiDAR frames. First, an incremental mapping module is introduced for accurate online pose estimation and simultaneous construction of a globally consistent neural map. Then, we perform offline global optimization to refine the reconstruction quality for the initial map. Finally, we propose an innovative neural map simplification method tailored for our representation, which aggregates the redundant local SDFs to further reduce the memory usage while preserving geometric fidelity. Extensive experiments demonstrate that our approach delivers superior localization accuracy and achieves state-of-the-art mapping performance with high efficiency and extremely low map memory consumption, especially requiring only about 1/10 the memory on the Oxford Spires dataset compared with existing advanced methods.</p>

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

SLNMapping: Super Lightweight Neural Mapping in Large-Scale Scenes

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Chenhui Shi,
  • Fulin Tang,
  • Hao Wei,
  • Yihong Wu

摘要

We propose SLNMapping, a novel neural mapping framework for super lightweight reconstruction in large-scale scenes. The core is a new ultra-compact neural map representation composed of a set of feature-independent local signed distance functions (SDFs) with outstanding expressiveness. To support efficient optimization, we introduce a novel parallel local SDF detection algorithm that enables real-time updates of local SDF states. Based on the excellent representation, we develop a three-stage mapping strategy for efficient, accurate, and lightweight large-scale reconstruction from streaming LiDAR frames. First, an incremental mapping module is introduced for accurate online pose estimation and simultaneous construction of a globally consistent neural map. Then, we perform offline global optimization to refine the reconstruction quality for the initial map. Finally, we propose an innovative neural map simplification method tailored for our representation, which aggregates the redundant local SDFs to further reduce the memory usage while preserving geometric fidelity. Extensive experiments demonstrate that our approach delivers superior localization accuracy and achieves state-of-the-art mapping performance with high efficiency and extremely low map memory consumption, especially requiring only about 1/10 the memory on the Oxford Spires dataset compared with existing advanced methods.