HiP3R: hierarchical pose graphs cut long-horizon 3D reconstruction ATE from 8.3m to 4.0m
zhenjun_zhao · x · 2026-10-24
A NTU team (Chen Change Loy's group) released HiP3R, using hierarchical pose graph optimization to let feed-forward models handle thousands-of-frames long videos.
- Rationale: chunk-level Sim(3) graphs efficiently fix scale drift and close large loops but can't correct intra-chunk errors, while frame-level optimization is flexible yet globally expensive and can break metric scale. HiP3R first optimizes a chunk-level Sim(3) graph for global layout and scale, then fixes that layout and estimates one shared SE(3) camera per frame, merging overlapping window predictions into a single trajectory
- Loop closures from distant revisits are generated via visual place recognition, joint-window inference, and two-endpoint registration
- Results: 4.017m mean ATE across benchmarks (LingBot-MAP backend) vs. 8.288m for the runner-up; with VGGT-Omega backend, mean ATE on Oxford/KITTI/VBR drops from 43.06m to 11.25m
More from Research
- Cell paper: in vivo gene writing with engineered retrotransposons — anshulkundaje · 2026-10-24
- FADA adapts humanoid robots to new conditions in ~2 minutes, no rewards needed — GuanyaShi · 2026-10-24
- Energy-Based Transformer research teased, with claims diffusion's end is near — burny_tech · 2026-10-24
- "Looped Model Atlas" visualizes LLM architecture papers as an interactive magic circle — burny_tech · 2026-10-24
- Rocq formalization shows Peano Arithmetic proves "K230 halts iff NOT Con(ZFC)" — burny_tech · 2026-10-24
- AI now designs custom phages on demand to kill superbugs, replacing sewage hunting — RobertWKemp · 2026-10-24