[vc_section][vc_row][vc_column][vc_column_text]Large-scale 3D mapping is an important task for <\/b>robotics and autonomous driving. However, mobile robots and <\/b>autonomous vehicles with limited hardware resources may face <\/b>issues with large memory consumption. It is challenging to achieve <\/b>a balance between mapping quality and memory consumption. <\/b>To address this issue, we propose a new compact implicit neural <\/b>map representation – the Tri-Pyramid that can infer the Truncated <\/b>Neural Distance Field (TNDF) given an arbitrary 3D position. <\/b>Additionally, we introduce a TNDF label rectification method con<\/b>sidering both the direction of ground normals and closest surface <\/b>points to enhance the precision of supervision signals for train<\/b>ing with a set of effective loss functions. Experiments on public <\/b>datasets demonstrated that our method reaches comparable or <\/b>superior performance for dense mapping while significantly reduc<\/b>ing memory consumption compared to previous LiDAR mapping <\/b>approaches. Furthermore, our study confirms the scalability and <\/b>adaptability of our approach from room-scale to city-scale scenes. <\/b>Moreover, we explore the potential of directly leveraging the im<\/b>plicit neural map representation for localization tasks by solving <\/b><\/p>\n
an optimization problem. The experiments showcase the accurate <\/b>localization capabilities of our method in various scenarios.<\/b><\/p>\n