We propose NPMV-SLAM, an RGB-Depth neural implicit dense SLAM system based on multi-resolution voxels and feature point clouds. The method demonstrates exceptional performance in tracking accuracy and reconstruction quality, aiming to address the shortcomings of traditional visual SLAM in texture detail modeling and geometric consistency, as well as the limitations of existing neural implicit methods in real-time performance and scene scalability. (1) We innovatively propose a position-enhanced encoding mechanism that fuses multi-resolution hash voxel grids with feature point clouds. This design fully leverages the high sensitivity of point clouds to high-frequency geometric details and the global structural continuity provided by voxels, achieving complementary advantages during network training and inference, thereby comprehensively enhancing the system’s reconstruction generalization capability. (2) Furthermore, we design an adaptive sampling strategy guided by point cloud density priors. This strategy fundamentally alleviates the core issue of insufficient scene scalability through data-driven online point cloud reconstruction. By filtering out invalid, non-surface sampling points, it concentrates computational resources on object surface regions, significantly reducing computational redundancy in empty areas, and achieves efficient point cloud spatial indexing with the aid of a vector database similarity search algorithm. While maintaining operational efficiency, our method significantly improves both detailed reconstruction capability and global reconstruction completeness. Experiments conducted on multiple indoor scenes from the Replica and TUM datasets show that our approach achieves notable improvements in tracking accuracy, rendering quality, and mapping accuracy, successfully balancing precision and efficiency.
Huang et al. (2026) studied this question.