VG-Mapping: Variation-Aware 3D Gaussians for Online Semi-static Scene Mapping

TL;DR

This work tackles the challenge of keeping dense maps up-to-date in semi-static environments where objects change between visits. It introduces VG-Mapping, a hybrid representation that fuses a TSDF voxel map with 3D Gaussian splatting and a variation-aware density control to detect and update changed regions online. A dedicated VG-Scene RGB-D dataset (synthetic and real) is presented for benchmarking, and experiments show substantial gains in rendering quality and update efficiency over state-of-the-art baselines, along with improvements in downstream tasks such as open-vocabulary segmentation and 6D pose estimation. The approach enables photorealistic, up-to-date maps that better support robust localization, planning, and interaction in semi-static settings.

Abstract

Maintaining an up-to-date map that accurately reflects recent changes in the environment is crucial, especially for robots that repeatedly traverse the same space. Failing to promptly update the changed regions can degrade map quality, resulting in poor localization, inefficient operations, and even lost robots. 3D Gaussian Splatting (3DGS) has recently seen widespread adoption in online map reconstruction due to its dense, differentiable, and photorealistic properties, yet accurately and efficiently updating the regions of change remains a challenge. In this paper, we propose VG-Mapping, a novel online 3DGS-based mapping system tailored for such semi-static scenes. Our approach introduces a hybrid representation that augments 3DGS with a TSDF-based voxel map to efficiently identify changed regions in a scene, along with a variation-aware density control strategy that inserts or deletes Gaussian primitives in regions undergoing change. Furthermore, to address the absence of public benchmarks for this task, we construct a RGB-D dataset comprising both synthetic and real-world semi-static environments. Experimental results demonstrate that our method substantially improves the rendering quality and map update efficiency in semi-static scenes. The code and dataset are available at https://github.com/heyicheng-never/VG-Mapping.

Figures (5)

• Figure 1: We propose VG-Mapping, an RGB-D online 3DGS mapping system tailored to semi-static scenes. (a) When a robot revisits the same place, dynamic changes across visits may cause inconsistencies between the prior map and the current observation, which we define as semi-static regions. To address this issue, (b) VG-Mapping introduces a variation-aware mapping mechanism that (c) efficiently and accurately updates the changed areas.
• Figure 2: VG-Mapping System Pipeline. In terms of map representation, VG-Mapping combines 3DGS with TSDFs. Leveraging the geometric information provided by the TSDF-based voxel map, we design a variation-aware density control to replace the optimization-driven ADC in vanilla 3DGS.
• Figure 3: Qualitative results on the VG-Scene dataset. The first column shows the scene before changes, where the red masks highlight the objects or regions undergoing semi-static changes in the current view. The second column presents the observations after the changes. Compared to the baselines, our method achieves higher-quality updates in the changed regions.
• Figure 4: Effectiveness analysis of each component. (a) Pruning reduces the influence of outdated Gaussians during optimization, preventing suboptimal results. (b) AVD facilitates the detection of appearance changes without geometric displacement, while (c) GVD resolves ambiguities in regions with similar appearance but different depth.
• Figure 5: Qualitative results of Grounded-SAM on rendered images. The left-hand text denotes the prompt. We visualize the masks, bounding boxes, and confidence produced by Grounded-SAM.