MG-SLAM: Structure Gaussian Splatting SLAM with Manhattan World Hypothesis

TL;DR

MG-SLAM addresses incomplete Gaussian SLAM reconstructions in indoor environments by leveraging the Manhattan World constraint to guide line- and plane-based structure. It introduces fused line features for robust tracking and a surface-interpolation pipeline that densifies large planar surfaces with new Gaussians, followed by mesh extraction with regularization. Evaluations on Replica-V1, Replica-Apartment, ScanNet, and a physical long-horizon platform show state-of-the-art tracking and mapping quality with strong rendering fidelity, while maintaining real-time performance. The approach advances Gaussian SLAM by integrating structural priors with dense Gaussian representations to achieve more complete, geometrically faithful indoor maps and usable meshes for downstream tasks.

Abstract

Gaussian Splatting SLAMs have made significant advancements in improving the efficiency and fidelity of real-time reconstructions. However, these systems often encounter incomplete reconstructions in complex indoor environments, characterized by substantial holes due to unobserved geometry caused by obstacles or limited view angles. To address this challenge, we present Manhattan Gaussian SLAM, an RGB-D system that leverages the Manhattan World hypothesis to enhance geometric accuracy and completeness. By seamlessly integrating fused line segments derived from structured scenes, our method ensures robust tracking in textureless indoor areas. Moreover, The extracted lines and planar surface assumption allow strategic interpolation of new Gaussians in regions of missing geometry, enabling efficient scene completion. Extensive experiments conducted on both synthetic and real-world scenes demonstrate that these advancements enable our method to achieve state-of-the-art performance, marking a substantial improvement in the capabilities of Gaussian SLAM systems.

Figures (11)

• Figure 1: Visualization of MG-SLAM on scene0000_00 and scene0207_00 of the ScanNet dataset dai2017scannet. Our method leverages robust line segments to achieve superior camera pose estimation and scene reconstruction results. Moreover, by applying structural surface constraints, we enhance and complete the planar surfaces of the scene through the insertion of new Gaussian primitives to fill gaps.
• Figure 2: The two-phase pipeline illustration of our proposed MG-SLAM. The upper section visualizes the parallel processes of the tracking and mapping systems. The lower section presents the post-optimization of scene interpolation and mesh extraction.
• Figure 3: The novel view sythesis of the scene apartment_0 from the Replica Apartment dataset straub2019replica. The top left shows the line segments extracted in 3D space. The bottom left illustrates the overall reconstructed scene.
• Figure 4: Qualitative comparison of our method and the baselines for novel-view synthesis on the ScanNet dataset dai2017scannet. The outcomes show that our method provides more robust and fine-grained reconstructions in real-world complex scenes compared to current NeRF-based and Gaussian-based approaches.
• Figure 5: Qualitative comparison of our method and the baselines on the trajectory collected using our physical platform. The left side displays the line and point features extracted by our tracking system. The right side presents the reconstruction comparisons with Gaussian-based methods, with MG-SLAM showing more reliable reconstruction results in long-horizon and textureless indoor environments.