PoP‐Mapping: Real‐Time Dense Mapping and Novel View Synthesis With Hybrid Point‐Plane Representation

Dense 3D visual mapping is a fundamental capability for autonomous robotic perception and navigation, where implicit neural representations have recently emerged as a powerful paradigm. However, most existing methods rely on a single homogeneous representation (e.g., plane grids or points) for heterogeneous scenes, struggling to reconcile global structures with high‐frequency details. This article presents point‐plane mapping (PoP‐Mapping), a hybrid neural radiance field (NeRF) system that combines axis‐aligned feature planes for efficient global modeling with adaptive neural points for local refinement, enabling accurate reconstruction from RGB‐D inputs. Moreover, a hierarchical two‐stage (H2S) training strategy is employed to jointly optimize these features, where points are dynamically anchored based on appearance and geometric complexity, allocating neural primitives to regions with high‐frequency details. Furthermore, to overcome the severe degradation and floating artifacts common in unconstrained implicit fields, we incorporate a truncated signed distance function (TSDF)‐aided mesh‐guided sampling strategy. This leverages explicit surface priors to strictly constrain volumetric queries, significantly improving the geometric consistency and stability of novel view synthesis. Experiments on the Replica dataset show that PoP‐Mapping achieves the best average reconstruction and rendering results among the evaluated baselines and reduces floating artifacts in novel‐view synthesis. A deployment on a mobile robot further demonstrates the applicability of PoP‐Mapping to real‐world RGB‐D mapping.