Dual Covariance 3DGS SLAM Decouples Rendering from Registration, Tracks at 60 FPS
zhenjun_zhao · x · 2026-09-23
A new arXiv paper (2609.25746) proposes Dual Covariance Gaussian Splatting SLAM.
- Problem: In ICP-based 3DGS SLAM, each Gaussian's covariance serves both rendering and registration with conflicting demands—the mapper flattens it against surfaces to minimize photometric error, while robust registration benefits from measurement uncertainty.
- Method: A dual-covariance parameterization where each Gaussian keeps one mean but two covariances: a rendering covariance optimized by the mapper and a tracking covariance derived from an RGB-D sensor noise model. Tracking covariances also serve as Gaussian anchors for image corners, constraining directions where depth geometry is weak.
- Results: Evaluated on TUM RGB-D, ScanNet, Replica and two outdoor RealSense D435i sequences (wheeled and handheld), achieving robust tracking, reduced odometry drift, at 60 FPS.
More from Embodied
- Musk predicts 1 billion humanoids in 10 years, 100 billion in 20, at 10x every 3-4 years — davidpattersonx · 2026-09-23
- Hugging Face ships @huggingface/lerobot JS package to read robot datasets in-browser — mishig25 · 2026-09-23
- Flex-π: 6B world action model predicts 3D pointmaps and DINO features, not just color video — chris_j_paxton · 2026-09-23
- MatcherCompass: A Deployment-Aware Benchmark for Choosing Image Matchers on Real Robots — zhenjun_zhao · 2026-09-23
- Rabbit launches OS3, a cloud agent that drives Windows, Mac and Linux remotely — emmanuelvivier · 2026-09-23
- Drawing robot research with Gemini models featured in IEEE RAM Arts and Robotics special issue — m_wulfmeier · 2026-09-23