AutoCompass trains visual localizers from noisy GPS labels

A team from Niantic and Oxford will present a paper at ECCV 2026 arguing that visual localization models can be trained on public maps without accurate heading labels, using only raw and noisy GPS. The approach, called AutoCompass, is by Javier Tirado-Garín and colleagues, and the abstract's headline claim is that "trained from raw GPS labels, models learn to predict accurate headings automatically."

The problem is familiar to anyone building geo-localization at scale. "Neural map matchers estimate an image's 3-DoF pose relative to a 2D map," the abstract states, but the geo-referenced training images those matchers need are almost never labeled cleanly; both position and heading drift. AutoCompass leans into that noise. Instead of demanding a precise ground-truth pose per image, the method defines a tolerance region around each raw GPS reading, and "if available, uses relative poses between training images obtained via SLAM or SfM for a more accurate training signal."

The claim that pushes hardest against convention is the heading one: no compass, no manually annotated orientation. Two researchers we track posted the paper as it circulated ahead of the Friday poster session at ECCV, where it appears as poster 3815 in Poster Session 3.

Per-benchmark accuracy figures and dataset breakdowns sit inside the full paper rather than the public listing.