GauSem-SLAM: Gaussian Semantic Submaps with Loop Closure for Globally Consistent SLAM
Bowen Zhang, Yufan Liu, Lebin Liang, Dong Li, Mingrui Li, Xuanxuan Zhang
AI summary
Problem
Current semantic SLAM methods struggle with effective loop closure and globally consistent semantic mapping, often treating semantics as external priors or applying costly rigid adjustments to dense point clouds that degrade 3DGS map quality.
Approach
The method constructs semantic 3DGS submaps with a spatial-semantic allocation strategy, embeds semantic information into Gaussian primitives, and uses a Semantic-Aware Loop Closure module with Semantic-Guided Registration (SGR) to correct intra- and inter-submap loops, followed by two-stage global map refinement.
Key results
- Integrates semantic submap construction, tracking, mapping, and loop closure into a unified pipeline.
- Introduces Semantic-Guided Registration (SGR) for efficient inter-submap alignment using DINOv2 features.
- Outperforms baselines in tracking (0.24 cm ATE RMSE) and mapping (39.94 PSNR, 95.81% mIoU) across three public datasets.
- Prevents geometric fragmentation at submap boundaries via spatial-semantic allocation and semantic sampling.
Why it matters
Enables real-time, globally consistent semantic scene understanding for robotics and embodied AI applications requiring accurate pose estimation and photorealistic rendering.
Abstract
3DGS has shown outstanding performance in multi-view geometry, driving its adoption in visual SLAM. However, real-time semantic 3DGS mapping faces challenges. Current methods typically treat semantics as external priors, making it hard to integrate them into SLAM tracking or loop closure correction. Moreover, traditional semantic SLAM corrects accumulated drift by applying rigid adjustments to dense point clouds, which is costly for 3DGS maps and limits loop closure performance. We propose GauSem-SLAM, which uses a Gaussian semantic submap representation with a progressive allocation strategy, integrating semantics into tracking, mapping, loop detection, and submap management. We fully exploit semantic information by designing a robust loop detection module that combines DINOv2 semantic features with semantic landmarks. Furthermore, we introduce Semantic- Guided Registration (SGR), a method for computing inter- submap loop constraints. Through intra-submap and inter- submap loop correction, followed by a two-stage global map refinement, our system achieves globally consistent pose esti- mation and mapping. Experiments on three public datasets demonstrate that our method outperforms prior methods in both tracking and mapping.