HETA++: A Novel Global SfM Pipeline
zhenjun_zhao · x · 2026-07-20
HETA++ introduces a novel pipeline for global SfM: starting with view track graph construction, rotation averaging, and feature track building, followed by relative translation refinement, global filtering, and hybrid explicit initialization, and finally completing the reconstruction via selective angular refinement, selective Bundle Adjustment, and full BA.
The paper presents camera pose AUC and reconstruction time comparisons on the 1DSfM dataset, showing that HETA++ achieves a better accuracy/speed trade-off across multiple scenes. Tables also display comparisons with methods like COLMAP, CReTA, LiGT, HETA, and GLOMAP.
More from Research
- Project CETI gets a Jeopardy! shout-out with a SETI-style whale clue — begusgasper · 2026-07-22
- Expert Warns Against Binary Switches for AIxBio Research, Advocates Tiered Access — davidmanheim · 2026-07-22
- Research finds memory compression makes AI agents drop safety rules and hit 59% violations — gerardsans · 2026-07-22
- DriftWorld claims a world model that runs at 30+ FPS and trains on 1–2 GPUs — du_yilun · 2026-07-22
- A physics reward can improve video generation without creating a real physics engine — Dapper-Drawer4546 · 2026-07-22
- Why a 1GW Chinese AI data center may be plausible after all — teortaxesTex · 2026-07-22