GS-Share: Enabling High-fidelity Map Sharing with Incremental Gaussian Splatting

TL;DR

GS‐Share is a photorealistic map‐sharing system with a compact representation that achieves higher fidelity, particularly for extrapolated views, with improvements of 11%, 22%, and 74% in PSNR, LPIPS, and Depth L1, respectively.

Abstract

Constructing and sharing 3D maps is essential for many applications, including autonomous driving and augmented reality. Recently, 3D Gaussian splatting has emerged as a promising approach for accurate 3D reconstruction. However, a practical map-sharing system that features high-fidelity, continuous updates, and network efficiency remains elusive. To address these challenges, we introduce GS-Share, a photorealistic map-sharing system with a compact representation. The core of GS-Share includes anchor-based global map construction, virtual-image-based map enhancement, and incremental map update. We evaluate GS-Share against state-of-the-art methods, demonstrating that our system achieves higher fidelity, particularly for extrapolated views, with improvements of 11%, 22%, and 74% in PSNR, LPIPS, and Depth L1, respectively. Furthermore, GS-Share is significantly more compact, reducing map transmission overhead by 36%.

Figures (10)

• Figure 1: Illustration of view interpolation and extrapolation. Interpolated views closely resemble the training views, whereas extrapolated views differ significantly from them.
• Figure 2: Overview of GS-Share. GS-Share takes RGB-D images as input and aligns them to a unified coordinate system using COLMAP colmap, followed by a global map construction process. Since the global map must be transmitted to users, its size is critical, motivating a compact representation. To further enhance map quality, GS-Share generates an auxiliary virtual map and renders virtual images from it. After post-processing, these virtual images serve as pseudo ground truth to augment the training data. Once the global map is reconstructed, users register with the server to access it. GS-Share transmits the full map during the initial stage and the map increments in subsequent updates. By leveraging the received increments, users reconstruct the latest Gaussian map with minimal transmission overhead.
• Figure 3: The process of virtual-image-based map enhancement. It uses $L_{pred}$ as the loss for training the predictor and $L^{v}$ for training the global map. During the training of the global map, the parameters of the predictor are frozen.
• Figure 4: The process of training and transmitting map data. (a) and (b) represent the procedures without and with incremental map update, respectively. AE refers to arithmetic coding arithmetic, while AD denotes the corresponding decoding procedure.
• Figure 5: Performance comparison under different compression ratios on extrapolated views.