Glob3R pairs 3D foundation models with global SfM and reports 3.2 cm on TUM RGB-D
ducha_aiki · x · 2026-07-23
Glob3R is a new global structure-from-motion pipeline built with 3D foundation models. The method finetunes a Pi3 matching head in a RoMaV2-style setup, uses SALAD to sort image sequences, and then applies a COLMAP-like global SfM stage.
The attached figures show the full reconstruction flow and results on multiple benchmarks:
- ETH3D pose estimation: the paper reports strong accuracy, with the method reaching 100.0 / 91.3 on the loose-threshold table and 91.58 / 73.27 at the strict threshold for the best comparison shown.
- TUM RGB-D: average trajectory RMSE is 3.2 cm, matching AMB3R and beating several baselines.
- KITTI: the average trajectory RMSE is 13.21 m, better than the listed baselines in the table.
- Ablations show the full method outperforms simplified variants in both reconstruction quality and speed tradeoffs.
Overall, the paper argues that dense warp-based track generation plus global association and bundle adjustment can make SfM more robust on ordered or pseudo-ordered image sequences.
More from Research
- NeurIPS reviews are now rolling out — timrudner · 2026-07-23
- Monthly derivative-model shares now tilt heavily toward China — natolambert · 2026-07-23
- Daily open-model dashboard shows China still far ahead of the US — natolambert · 2026-07-23
- Teaching notebook builds a neural network from scratch with linear algebra and the chain rule — GeostatsGuy · 2026-07-23
- DeepSeek-V4-Pro training on Ascend 910C exposes long-tail kernel bottlenecks — teortaxesTex · 2026-07-23
- Agent memory works better as a statement graph with provenance and append-only history — chrislally · 2026-07-23