Dual Covariance 3DGS SLAM: two covariances per Gaussian, robust tracking at 60 FPS
kwangmoo_yi · x · 2026-09-26
arXiv paper "Dual Covariance Gaussian Splatting SLAM" (Tan & Lam):
- Problem: ICP-based 3DGS SLAM uses each Gaussian's covariance for both rendering and registration — conflicting demands. The mapper flattens covariances against surfaces to minimize photometric error, while robust registration benefits from measurement uncertainty.
- Method: dual-covariance parameterization — one mean, two covariances per Gaussian: a rendering covariance optimized by the mapper, and a tracking covariance derived from an RGB-D sensor noise model. Tracking covariances also anchor image corners where depth geometry is weak.
- Results: robust tracking and reduced odometry drift on TUM RGB-D, ScanNet, Replica, and two outdoor RealSense D435i sequences, at 60 FPS.
Related event: Dual Covariance Gaussian Splatting SLAM Achieves 60 FPS Robust Tracking(2 posts)→
More from Embodied
- Perceptron launches Mk1.5, one embodied AI model for drones, quadrupeds and smart glasses — code_star · 2026-09-26
- Meta's new VR glasses land but still trail Vision Pro, longtime reviewer says — adariostrange · 2026-09-26
- Dual Covariance Gaussian Splatting SLAM decouples rendering and registration — kwangmoo_yi · 2026-09-26
- Meta goes all-in on Muse: voice, glasses, a charm — and a free rootable cloud PC per user — altryne · 2026-09-26
- WetLabs Benchmark tests robots on 9 wet-lab tasks: can machines accelerate science? — ericjang11 · 2026-09-26
- Arm Total Design: the build pipeline from virtual robot to real fleet, 83 partners — lmoroney · 2026-09-26