EliGSiR: continual RGB-D Gaussian Splatting mapping under bounded compute hits 21.52 dB on TUM RGB-D
zhenjun_zhao · x · 2026-09-19
arXiv paper EliGSiR (Ellensohn, Rueckert, Rauch) tackles continual RGB-D mapping where new observations arrive online while reconstructed regions must be preserved.
Core mechanisms:
- Map-Guided View Scheduling filters redundant incoming views and revisits retained ones based on map state.
- Load-Adaptive Fidelity adjusts supervision resolution to current mapping load instead of a fixed schedule.
- Targeted Geometry Growth decouples depth supervision from Gaussian creation, adding capacity only where repeated observations indicate missing or misplaced structure.
Evaluated on Replica, TUM RGB-D, ScanNet++ and real sensor sequences, reaching 21.52 dB on TUM fr3/longofficehousehold with ground-truth poses.
More from Embodied
- ROS Weekly: ROSCon Global next week in Toronto, InOrbit open-sources Robot Ops — kscottz · 2026-09-19
- MOSS robot demo replays real Jev decisions in simulation, switching tasks from cans to bottles — kamathsblog · 2026-09-19
- Niantic Spatial launches Gaussian Splat Relighting beta for embodied AI simulation — Scobleizer · 2026-09-19
- AMB3R-SLAM: Kilometer-Scale Monocular SLAM on One Consumer GPU, 70% Lower ATE — zhenjun_zhao · 2026-09-19
- SLAMSqueezeBench benchmarks nine SLAM systems under realistic edge compute and memory constraints — zhenjun_zhao · 2026-09-19
- ABC-130K: largest open bimanual teleop dataset with 3,500 hours, headed to CoRL 2026 — pabbeel · 2026-09-19