NGEL-SLAM: Neural Implicit Representation-based Global Consistent Low-Latency SLAM System

TL;DR

NGEL-SLAM addresses the challenge of achieving global consistency and low latency in neural implicit SLAM by fusing a traditional tracking backbone (ORB-SLAM3) with a fast, octree-based multi-sub-map neural representation. It achieves rapid loop-closure responses via real-time sub-map adjustments followed by targeted fine-tuning, and leverages uncertainty-guided rendering to combine information across sub-maps. The system demonstrates state-of-the-art tracking and mapping performance on Replica, ScanNet, and TUM RGB-D while maintaining low latency, outperforming purely neural SLAM baselines. This hybrid approach offers practical benefits for indoor robotics and AR/VR applications by delivering dense, high-fidelity geometry and textures with globally consistent maps at real-time speeds.

Abstract

Neural implicit representations have emerged as a promising solution for providing dense geometry in Simultaneous Localization and Mapping (SLAM). However, existing methods in this direction fall short in terms of global consistency and low latency. This paper presents NGEL-SLAM to tackle the above challenges. To ensure global consistency, our system leverages a traditional feature-based tracking module that incorporates loop closure. Additionally, we maintain a global consistent map by representing the scene using multiple neural implicit fields, enabling quick adjustment to the loop closure. Moreover, our system allows for fast convergence through the use of octree-based implicit representations. The combination of rapid response to loop closure and fast convergence makes our system a truly low-latency system that achieves global consistency. Our system enables rendering high-fidelity RGB-D images, along with extracting dense and complete surfaces. Experiments on both synthetic and real-world datasets suggest that our system achieves state-of-the-art tracking and mapping accuracy while maintaining low latency.

Figures (8)

• Figure 1: Rendering and tracking results. Compared with NICE-SLAM zhu2022nice, our method renders higher-fidelity images and provides more precise camera tracking results. Additionally, our method performs fast convergence and enables low-latency map updates after loop closure, enabling it to run 10x faster than NICE-SLAM. The ground truth camera trajectory is shown in black, and the estimated trajectory is shown in red.
• Figure 2: System overview. Our proposed system comprises two primary modules: the tracking module and the mapping module. It can be further divided into three processes: tracking, dynamic local mapping, and loop closing. These three processes work together to ensure global consistency and low latency in our system. The tracking process takes an RGB-D stream as input and tracks the camera pose in real-time. If a frame is selected as a keyframe, it is passed to the dynamic local mapping process. In this process, the tracking module performs local BA while the mapping module trains the corresponding local map. When a loop is detected, the loop closing process optimizes the camera poses using global BA and updates the scene representation. All processes are executed in parallel.
• Figure 3: Mapping network. The Mapping network employs a sparse octree structure to store multi-level features and two single MLPs.
• Figure 4: Rendering results on the Replica dataset.
• Figure 5: Rendering results on the ScanNet dataset.