arxiv.org web signal

AutoCompass learns visual localization from weak GPS labels

By Published Updated

TL;DR

  • AutoCompass trains neural map matchers using only raw GPS labels and shows heading labels are unnecessary because models learn heading automatically.
  • The method adds a tolerance region around each noisy GPS point and, when available, uses relative poses from SLAM or SfM as a stronger signal.
  • The authors report gains over absolute-pose-supervised baselines on both driving and egocentric localization benchmarks.

Heading labels for visual localization models can be thrown away. That is the claim of AutoCompass, posted to arxiv, which trains neural map matchers using only raw GPS coordinates.

"Trained from raw GPS labels, models learn to predict accurate headings, automatically," the paper states. The recipe is threefold: drop heading supervision entirely, define a tolerance region around each noisy GPS point rather than trusting it as a hard label, and, where SLAM or SfM traces exist, feed in the relative poses between training images as a stronger signal.

The paper reports: "Across driving and egocentric benchmarks, AutoCompass consistently outperforms counterparts trained with the usual strong reliance on absolute pose labels." Two researchers in our tracker shared the link. The abstract publishes no per-benchmark numbers.

Shared on Bluesky by 2 AI experts

Originally reported by arxiv.org

Read the original article →

Original headline: AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels