Graph3R unifies view graph initialization with feed-forward reconstruction, recall +8.99pts
zhenjun_zhao · x · 2026-10-19
Jingwen Yu, Ping Tan, Hong Zhang, et al. introduce Graph3R, unifying view graph initialization with feed-forward reconstruction via a shared geometric representation, tackling the problem where uncurated image sets spanning disjoint scenes corrupt poses and geometry.
Pipeline: π³ multi-view features → a lightweight verification head predicts pairwise covisibility → a view graph groups views by scene → shared features are masked per group before decoding poses and dense geometry, reconstructing disjoint scenes separately in a single backbone pass.
Results: across three benchmarks, average max recall at 100% precision improves by up to 8.99 points over the baseline, with better mixed-scene reconstruction; multi-session mapping gains cross-session localization; ETH3D pose accuracy of 76.19/88.09 (10°/1m, 20°/5m) beats RobustPi3 and peers; ETH3D multi-view depth consistency (avg MVS 0.2230) leads, at faster runtime.
More from Research
- Interpretability debate: residual stream superposition may be real but unimportant — burny_tech · 2026-10-19
- Residual stream superposition can be real yet unimportant if neurons are monofunctional — burny_tech · 2026-10-19
- AI agent rebuilds black hole images from raw EHT data, matching 91%-98% of official results — fofrAI · 2026-10-19
- ORL trains diffusion LMs on their own trajectories, boosting a 4B model's throughput by 6.9x past speculative decoding — burny_tech · 2026-10-19
- New consciousness preprint claims unexpected neural findings, explained via Claude — eherrerosj · 2026-10-19
- NVIDIA and Stanford Unveil World from Motion: Generative 4D Gaussian Reconstruction from a Single Video — ZGojcic · 2026-10-19