Efficient Sampling and Volume Rendering Strategy for Neural Field SLAM

Meng Wang, Yue Qi

Abstract: Neural implicit representations have recently shown encouraging results in various domains, including promising progress in simultaneous localization and mapping (SLAM). Nevertheless, existing methods require substantial computational power (typically running on GPUs like RTX 3090) to maintain fast processing speeds and experience a noticeable slowdown on consumer-grade GPU devices. In this work, we present an efficient sampling and volume rendering strategy central to achieve faster processing speeds without additional memory costs. Two key ingredients enable our goal: (1) a progressive sampling method for tracking to perform efficient and robust camera poses estimated and (2) a feature fusion approach for color volume rendering to reduce calculation cost during the optimization process. Extensive experiments on four datasets (ScanNet, TUM RGB-D, Replica, and Synthetic) demonstrate that our approach can accelerate the optimization process and achieve better or no-par performance compared to the original method on a consumer-grade GPU device.