LOD-InfiniTAM: Adaptive Depth Sampling for Accurate RGB-D SLAM

Dense RGB-D Simultaneous Localization and Mapping (SLAM) is the mainstream approach for augmented reality and autonomous robotics, fields that demand high accuracy and real-time performance in markerless environments. However, the measurement noise of depth sensors increases quadratically with distance, introducing significant interference into SLAM systems. In this paper, we propose an adaptive depth sampling method based on a Level-of-Detail (LoD) algorithm that selectively filters observations from distant or textureless regions. Integrated into the InfiniTAM framework, the method reduces both noise-induced errors and redundant computation. Evaluations on TUM RGB-D dataset demonstrate that the method improves tracking accuracy by 22%, reduces artifacts in volumetric reconstruction, and maintains real-time performance at over 240 FpS.