Niantic's AutoCompass trains accurate visual localization from noisy GPS labels
ducha_aiki · x · 2026-09-08
Niantic Spatial researchers present AutoCompass (ECCV 2026), a supervision approach for training neural map matchers that estimate an image's 3-DoF pose against public 2D maps despite noisy labels.
Key findings:
- Heading labels are unnecessary: models trained from raw GPS learn accurate heading prediction automatically.
- Tolerance regions: defining a tolerance region around raw GPS positions improves accuracy.
- Relative poses help: SLAM/SfM-derived relative poses between training images give a stronger supervision signal when available.
Across driving and egocentric benchmarks, AutoCompass consistently outperforms methods trained with the usual reliance on accurate absolute pose labels.
More from Research
- Afinetheorem: a different lab reportedly has an N-S-related proof, "we'll see this week" — Afinetheorem · 2026-09-08
- Alpöge and Buckmaster release smooth-forced blowup results; Tao says machinery may extend to Navier-Stokes — burny_tech · 2026-09-08
- NeurIPS desk-rejected 178 papers as AI-generated — while its chairs' own papers flagged at 24-69% — tughanbulut · 2026-09-08
- VERA paper undercuts its own rhetoric: structured Jacobian-based IDM beats raw 14B video model for robot control — GeorgiaChal · 2026-09-08
- Researcher challenges scaling orthodoxy: benchmark trends aren't universal laws of perception and reasoning — GeorgiaChal · 2026-09-08
- If AI Helps Crack the Navier–Stokes Millennium Problem, It Would Signal a Scientific Revolution — kimmonismus · 2026-09-08