LingGS: Decoupling Feed-Forward Geometry Prediction from Camera Tracking for High-Fidelity Gaussian SLAM

arXiv cs.GR details LingGS, an online Gaussian SLAM pipeline that separates feed-forward geometry prediction from camera tracking. Instead of relying on learned pose prediction, it uses a frozen streaming 3D foundation model as a virtual RGB-D sensor, then hands camera motion estimation to an explicit geometric solver.

That split matters because tracking quality is often the bottleneck in monocular RGB Gaussian splatting SLAM, especially when the system has a limited optimization budget. By reprojection-based pose solving plus a streaming Gaussian mapper, LingGS aims to keep reconstruction dense while improving pose accuracy over learned tracking approaches.

The paper reports that geometric tracking improves camera pose accuracy across multiple streaming geometry backbones, suggesting the idea is not tied to one specific model family. In practice, that makes the approach more interesting for teams building real-time reconstruction, telepresence, capture tools, or scene understanding systems where pose drift quickly ruins output.

On indoor and outdoor datasets, the system is said to significantly improve reconstruction quality over prior methods while running at about 7 FPS on a single RTX 4090. The authors also claim 2-3x the throughput of the evaluated prior SLAM systems, which is the kind of gain that can change whether a research prototype is merely impressive or actually usable.