AutoCompass trains neural map matchers from raw GPS weak labels, outperforming strong-label baselines

zhenjun_zhao · x · 2026-09-05

An ECCV 2026 paper from the Naver/Niantic-affiliated team behind GLACE. AutoCompass is a supervision scheme for training neural map matchers from inaccurate absolute pose labels: it shows heading labels are unnecessary (models learn accurate headings from raw GPS), defines a tolerance region around raw GPS to improve positional accuracy, and optionally leverages relative poses from SLAM/SfM as stronger supervision. Across driving and egocentric benchmarks it consistently outperforms counterparts trained with strong reliance on absolute pose labels.

Original post →

More from Research

Research channel →