F²SLAM turns feed-forward geometry into persistent factors, cutting uncalibrated ATE to 0.002 m on Replica
zhenjun_zhao · x · 2026-10-19
Researchers including Javier Civera released F²SLAM (arXiv 2610.05207) with open-source code.
- Key idea: instead of treating feed-forward 3D predictions as an external state aligned post hoc, F²SLAM converts them directly into optimization-native, target-weighted measurements on a persistent dense factor graph, letting broad multi-view evidence refine the SLAM state.
- A high-frequency stream maintains local tracking and graph connectivity; a low-frequency stream selectively refreshes measurements after a state-consistency check; both constrain poses, inverse depths, and optional intrinsics via one dense bundle adjustment.
- Results: best average uncalibrated ATE on TUM RGB-D (0.027 m) and Replica (0.002 m); well ahead of uncalibrated foundation-model baselines on KITTI; on 7-Scenes vs SLAM-Former, completion 0.030→0.015 m and Chamfer 0.027→0.019 m.
- Runs at 7.4 FPS on one A100 with lower peak GPU memory than SLAM-Former (15.4 vs 27.7 GB).
More from Research
- CASTER formalizes zero-shot robot manipulation from synthetic videos as constrained optimization — mangahomanga · 2026-10-20
- What makes human video generation good for robot manipulation? A constrained-optimization analysis — mangahomanga · 2026-10-20
- New MDP paper constructs learning algorithms for the bias-to-Blackwell optimality hierarchy — _onionesque · 2026-10-20
- DeepMind's Gemini agent discovers and Lean-verifies new acceleration rules for 11 BB(6) Turing machines — omarsar0 · 2026-10-20
- Atomwell debuts: 9B foundation model shows clean scaling laws on 80M experimental measurements — chaitjo · 2026-10-20
- SFT vs RLVR: how fine-tuning paradigms shape LLM confidence calibration in high-stakes use — burkov · 2026-10-20