Scholarlink

<p>Neural implicit SLAM combines SLAM-based image tracking with NeRF scene reconstruction, achieving strong results but facing two challenges: (1) difficulty distinguishing object geometry and (2) degraded texture rendering in complex scenes. Geometry reconstruction is hindered by incomplete depth maps and random pixel sampling, while texture rendering suffers from volume rendering uncertainty and pose drift, causing blurred textures. We propose a neural implicit SLAM framework leveraging gradient and depth priors to enhance both shape and texture. To address geometry, we introduce a pixel sampling strategy emphasizing object edges, reinforcing shape constraints. For textures, we design an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\alpha \)</EquationSource> </InlineEquation>-composition rendering method guided by depth, together with a novel keyframe selection for accurate poses. Experiments on Replica and ScanNet demonstrate that our approach improves shape recognition of objects with similar textures and enhances texture rendering accuracy, advancing NeRF-based SLAM in challenging scenes.</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

Object shape differentiation and texture rendering for neural implicit SLAM

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Jiaming Lu,
  • Ruyu Liu,
  • Jianhua Zhang,
  • Xiaofeng Liu,
  • Xu Cheng

摘要

Neural implicit SLAM combines SLAM-based image tracking with NeRF scene reconstruction, achieving strong results but facing two challenges: (1) difficulty distinguishing object geometry and (2) degraded texture rendering in complex scenes. Geometry reconstruction is hindered by incomplete depth maps and random pixel sampling, while texture rendering suffers from volume rendering uncertainty and pose drift, causing blurred textures. We propose a neural implicit SLAM framework leveraging gradient and depth priors to enhance both shape and texture. To address geometry, we introduce a pixel sampling strategy emphasizing object edges, reinforcing shape constraints. For textures, we design an \(\alpha \) -composition rendering method guided by depth, together with a novel keyframe selection for accurate poses. Experiments on Replica and ScanNet demonstrate that our approach improves shape recognition of objects with similar textures and enhances texture rendering accuracy, advancing NeRF-based SLAM in challenging scenes.