GauS-SLAM: Dense RGB-D SLAM with Gaussian Surfels

TL;DR

GauS-SLAM tackles geometry distortions and depth-blending ambiguities in Gaussian-based dense SLAM by introducing 2D Gaussian surfels, a surface-aware depth rendering pipeline, and a tightly coupled front-end/back-end with local maps. The 2D surfel representation enables unbiased depth, depth adjustment, and depth normalization, while the local-map design confines tracking to visible surfaces and mitigates occlusion-induced errors. The system demonstrates state-of-the-art tracking accuracy and rendering fidelity on ReplicaReplica and ScanNet++ datasets, with favorable runtime characteristics thanks to incremental mapping and map-merging strategies. Overall, GauS-SLAM advances dense RGB-D SLAM by combining robust Gaussian surface representations with principled depth rendering and a scalable local-map architecture, yielding practical gains in accuracy, view synthesis, and efficiency.

Abstract

We propose GauS-SLAM, a dense RGB-D SLAM system that leverages 2D Gaussian surfels to achieve robust tracking and high-fidelity mapping. Our investigations reveal that Gaussian-based scene representations exhibit geometry distortion under novel viewpoints, which significantly degrades the accuracy of Gaussian-based tracking methods. These geometry inconsistencies arise primarily from the depth modeling of Gaussian primitives and the mutual interference between surfaces during the depth blending. To address these, we propose a 2D Gaussian-based incremental reconstruction strategy coupled with a Surface-aware Depth Rendering mechanism, which significantly enhances geometry accuracy and multi-view consistency. Additionally, the proposed local map design dynamically isolates visible surfaces during tracking, mitigating misalignment caused by occluded regions in global maps while maintaining computational efficiency with increasing Gaussian density. Extensive experiments across multiple datasets demonstrate that GauS-SLAM outperforms comparable methods, delivering superior tracking precision and rendering fidelity. The project page will be made available at https://gaus-slam.github.io.

Figures (11)

• Figure 1: GauS-SLAM is a dense SLAM system using 2D Gaussian surfels2DGS, capable of simultaneously achieving high-precision localization and high-fidelity reconstruction. As shown in the left figure, GauS-SLAM exhibits millimeter-level tracking accuracy on a challenging real-world scenario(b20a261fdf in ScanNet++scannetpp dataset), significantly outperforming the SOTA approach, SplaTAMSplaTAM. The right figure demonstrates GauS-SLAM's SOTA performance on the ReplicaReplica dataset, achieving an absolute trajectory error(ATE-RMSE) of $0.06cm$ and 40.25 dB in rendering quality.
• Figure 2: Two challenges in Gaussian-based tracking methods. (a1) illustrates geometry distortions caused by center depth model of the 3D Gaussian. (a2) shows ill-blended depth arising from depth rendering between different surfaces. (b) demonstrates that, during the alignment, certain interference area exhibit high accumulated opacity making them challenging to be masked out as outliers.
• Figure 3: Overview of GauS-SLAM. This framework consists of a front-end that performs tracking and mapping using a single local map, and a back-end responsible for merging the local map into the global map and submap-based global optimization.
• Figure 4: The comparison of Rendering performance on ReplicaReplica. We present rendered color maps and depth error maps from 2 viewpoints to comparatively evaluate the rendering quality and geometry accuracy of different approaches.
• Figure 5: Comparison of mesh results on ReplicaReplica. Compared to isotropic 3D Gaussians, Gaussian surfels produce smoother mesh reconstructions.