Faster Vox-Fusion: Memory-efficient and Fast Dense Mapping

Dense mapping plays a significant role for robotics, autonomous driving and VR/AR. Traditional geometric-based dense mapping methods lack the capability to deal with view synthesis while the recently proposed NeRF-based dense mapping methods are time-consuming and short of generalization. Moreover, memory efficiency presents a notable challenge, impacting both the performance and resource utilization. To address these limitations, this paper introduces faster Vox-Fusion, a novel approach that not only significantly accelerates dense mapping (achieving a 25-fold increase in reconstruction speed over Vox-Fusion methods) but also enhances memory efficiency and maintains high levels of detail and generalization in most applications. This is achieved by leveraging the robust tracking capabilities of ORBSLAM3 for accurate camera pose estimation without refinement, and employing a novel hybrid keyframe strategy alongside an optimized dynamic voxel creation process featuring deduplication. Despite a marginal reduction in reconstruction quality (approximately 2-3%), our method requires substantially fewer GPU resources (about 65%), marking a significant advancement in dense mapping technology. This work sets a new benchmark for efficient, high-quality dense mapping, offering substantial benefits for real-world applications.